The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
S E P T E M B E R /O C T O B E R 1997 CONTENTS 39 Chain-Weighted GDP: Has Our Perspective Changed? Michael R. Pakko Our perspective on the U.S. economy's recent performance has been challenged recently by changes in the methodology used to adjust the National Income and Product Accounts for inflation. Michael R. Pakko surveys the changes embodied in the revised data, examining the question of whether or not the revisions alter our view of the overall pattern of economic fluctuations known collectively as the business cycle. Volume 79, Number 5 3 Economic Models of Employee Motivation Joseph A. Ritter and Lowell J. Taylor Workers present employers with a range of tricky problems. They can be crooked, subversive, surly, or indolent, even if they are paid on time. Joseph A. Ritter and Lowell J. Taylor explore economists’ main theories of how compensation is used to address employee motivation and how these models help to explain puzzling features of labor markets. Although these theories are often regarded as competitors, the authors treat them as complementary tools in understanding how employers deal with the complex problem of motivating workers. 23 51 The Demand for Divisia Money in a Core Monetary Union Katrin Wesche Proponents of an aggregation theoretic approach to money demand argue that simple-sum measures do not capture the theoretical notion of money, especially for broad monetary aggregates. European monetary aggregation, which uses indices for monetary services, seems attractive because these indices can account for the imperfect substitutability between different currencies. In this article, Katrin Wesche applies the aggregation theoretic framework to money holdings of European residents and compares the resulting index to simple-sum M3. She concludes that the Divisia index of European monetary services may provide additional insight into money demand during the period of transition to monetary union. Technical Analysis in the Foreign Exchange Market: A Layman’s Guide Christopher J. Neely Economists have traditionally been skeptical of the value of technical analysis, the use of past price behavior to guide trading decisions in asset markets. Instead, they have relied on the logic of the efficient markets hypothesis. Christopher J. Neely briefly explains the fundamentals of technical analysis and the efficient markets hypothesis as applied to the foreign exchange market, evaluates the profitability of simple trading rules, and reviews recent ideas that might justify extrapolative technical analysis. The Business Cycle and 61 Working Paper Series SEPTEMBER/OCTOBER 1997 Joseph A. Ritter is a research officer at the Federal Reserve Bank of St. Louis. Lowell J. Taylor is an associate professor at the Heinz School of Public Policy and Management, Carnegie Mellon University. Eran Segev and Joshua D. Feldman provided research assistance. Economic Models of Employee Motivation use the terms “wage” and “compensation” interchangeably throughout the article) high enough to deter undesirable behavior by making a job too good to lose are said to pay efficiency wages. It is fairly easy to see whether a firm is using some sort of piece rate plan. There is quite a bit of controversy, however, about whether firms that do not use piece rates adopt efficiency-wage or performancebonding plans. We follow our overview with a discussion of the nature of the evidence supporting the different models. Joseph A. Ritter Lowell J. Taylor T o most people it is a common sense proposition that hiring workers is a trickier problem than buying ballpoint pens. It is often difficult to find the right worker to hire, and workers who have already been hired can quit, steal, be hung over, refuse to cooperate with other workers, or simply not work very hard. In some workplaces some of these problems are relatively easy to solve, either by direct supervision or by directly linking pay to production. In general, however, things like ability, effort, and honesty are difficult to verify and consequently present special problems for personnel managers and economic theorists. The ways firms solve the problems of selecting, motivating, and retaining employees are potentially interesting to a wide cross-section of economists because they can affect how labor markets function and, therefore, how the entire economy operates. This article presents an overview of economists’ main hypotheses about the compensation strategies businesses use to address these kinds of problems. Broadly speaking, these solutions fall into three categories (with considerable diversity within each): piece rates, performance bonding, and efficiency wages. Piece rates link pay directly to workers’ output. Performance bonding uses a combination of up-front payments from workers and conditional repayments to guarantee workers’ performance. Firms that pay wages (we SIMPLE SUPPLY-ANDDEMAND MODELS OF LABOR MARKETS On one level, economists can analyze labor markets using the same supply-anddemand model they might apply to, say, wheat. Supply increases as the price (wage) received by the supplier increases. Demand increases as the price paid decreases. Equilibrium occurs where supply equals demand. For many purposes it is important to recognize that workers are not perfectly interchangeable; most nurses are not economists. This complication is easily handled by treating the markets for nurses and economists separately, each with its own supply and demand curves. Similarly, workers within the same profession are not typically interchangeable. An important dimension along which different kinds of workers can be distinguished is the collection of applicable knowledge and skills that economists call human capital. Levels of human capital vary not only across individuals, but also over time for a given individual. As an employee accumulates human capital, or as existing human capital deteriorates, the employee’s compensation can be expected to change. A worker’s willingness to accept a particular job will be affected by agreeable and disagreeable facets of the job. Workers require a higher wage to accept a hazardous F E D E R A L R E S E R V E B A N K O F S T. L O U I S 3 SEPTEMBER/OCTOBER 1997 job than a safe one. They may accept lower wages to work in a nice place, have flexible hours, or perform work that requires little effort. Differences in wages that come from these kinds of reasons are called compensating differentials. The theory of labor demand is especially important for this article. The core of that theory is based on the observation that hiring an additional employee (or employee hour) will increase the profits of the firm as long as the employee’s compensation is less than the value of the additional output the firm can produce after hiring the employee. The latter quantity is called the value of marginal product (VMP) and is calculated by multiplying the additional employee’s marginal product by the price of the firm’s product. This relationship defines the firm’s labor demand curve. Since the marginal product is likely to decrease as the firm hires more labor (holding other inputs fixed), the firm’s labor demand curve is downward-sloping: A firm that must pay higher wages will demand less labor. If there are no impediments, a labor market will reach equilibrium where supply equals demand. The theory of supply and demand does a good job of explaining the broad outlines of labor markets, but a closer look reveals some cracks. This article concentrates on the fact that (unlike wheat, for example) the same worker behaves differently in different economic circumstances; the same worker might, for example, work hard at $30 per hour but loaf at $7 per hour. The simple supply-and-demand framework cannot encompass this possibility, so different kinds of models are needed. worker motivation, labor economists have focused largely on three core problems: sorting potential employees, achieving optimal performance on the job, and regulating turnover. Sorting Job Applicants In the textbook supply-and-demand approach to labor markets, sorting applicants is assumed to be a simple problem. That theory presumes that an employer knows how productive an applicant will be if he or she takes the job. An accounting firm (or anybody else) knows that an accounting major is likely to be a more effective accountant than a high-school dropout. But that kind of insight is only the tip of the iceberg and would not help to land an applicant a job in the accounting firm’s personnel office. The difference between good and bad employees often depends on qualities that are difficult to discern (willingness to work hard, for example). If the firm designs the right incentives, however, it can encourage desirable applicants, even though the firm cannot easily identify them when they apply. Performance on the Job Workers’ behavior on the job can disrupt the firm’s attempts to make money in many ways. A surly worker might drive away a customer. An employee who shows up late might make it difficult for other workers to do their own jobs. A worker might be careless or simply not work very hard. Workers might steal from their employer. The list is virtually endless. Beyond the obvious, several aspects of these situations are important. First, none of the examples just mentioned is necessarily tied to any observable characteristic of a job applicant. Businesses use an arsenal of screening devices to try to avoid problems, but their effectiveness is manifestly limited. To get optimal performance from employees, the firm cannot rely on applicant screening alone but must also design effective incentives for existing employees. SPECIAL PROBLEMS IN LABOR MARKETS A central task of economic theory is to boil a problem down to its essentials so that it can be thoroughly understood and carefully analyzed. In principle, after the core of the problem is understood, economists turn their attention to the nuances that separate their models (artificial economies) from reality. In the area of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 4 SEPTEMBER/OCTOBER 1997 Second, certain on-the-job problems are particularly critical because employees work together in most firms. A worker who shows up 10 minutes late, for example, is not a problem if his job is to sit in front of a computer and write articles, but if he works on an assembly line, he may force several hundred workers to start work 10 minutes late. Third, the size and complexity of most workplaces make it impossible to detect all negative behavior. This fact triggers a powerful principle: To deter any behavior that is unlikely to be detected, punishment must be disproportionate. For example, suppose Joe likes hanging around the water cooler enough that he would be willing to pay the firm a dollar to be allowed this liberty for an extra hour each day, but his manager notices excess water cooler attendance only one time in a hundred.1 To deter this behavior, then, the firm must impose a penalty greater than a dollar if Joe is caught hanging around the water cooler. This is, in our view, the central reason the basic supply-and-demand model is unlikely to be completely satisfactory in labor markets. The most severe punishment a firm can impose is firing, but in the basic supply-and-demand model, firing imposes very little cost on the worker, because the model assumes that markets are anonymous and function quickly and efficiently. Thus a terminated worker has no difficulty in finding a comparable job.2 productive for some time. To the extent that firms cannot shift these costs to the new worker (through probationary wages, for example), firm-specific human capital is costly to the firm. Quit rates are not entirely outside the firm’s control, however. Compensation policies can provide incentives for workers to stay on the job. Turnover If the agency problem is related to the worker’s productivity, there is an obvious approach to solving it: Establish a direct connection between the worker’s output and his compensation. Many workers are compensated in ways that resemble piece rates: garment workers who are paid on the basis of output, sales workers paid on commission, auto mechanics in large dealerships whose pay is partly on a per-repair basis, and agricultural workers whose pay depends on the amount of fruit picked or rows of grape vines pruned. One pervasive problem in firms that tie pay closely to some objective measure of output is that they often get exactly The Agency Problem All of the problems mentioned in this section are corollaries of the maxim, “If you want something done right, you have to do it yourself.” The problem confronting the owner of a business is how to design incentives that will induce the workers to do it right or, more precisely, to behave in a way that maximizes the firm’s profits. This is an example of what economists call an agency problem: A principal (in this case the firm’s owner or manager) designs incentives for an agent or agents (the workers), who take actions that affect the principal’s well-being. The agency problem stems from the fact that there is a different connection between the agent’s actions and well-being than between the agent’s actions and the principal’s well-being.3 For example, it is in the firm’s interest for a worker to work hard (the action preferred by management), but the worker may prefer to spend the morning at the water cooler. 1 Perhaps because the manager is usually golfing. See footnote 3 below. 2 Recent work on the dynamics of labor markets uses matching models in which the firm and worker bargain over gains generated by a good match. If the worker’s share is small, firing costs the worker little. Even when the worker’s share is larger, so that termination is a significant penalty, its credibility as a disciplinary device is limited because firing is costly to the firm too. 3 The owner(s) of a large firm face another agency problem: how to get the manager of a firm to act in the interest of the owner(s). This problem has also been extensively studied under the heading of executive compensation. See Jensen and Murphy (1990). PIECE RATES Turnover can be very costly to the firm for two reasons. First, isolating and hiring a new worker can cost thousands of dollars for some jobs. Firms that outsource part of this activity to “head-hunters” (presumably because they think it is cheaper than doing it themselves) typically pay a commission that is a substantial fraction of the new worker’s annual pay. Second, new workers almost always need to accumulate some knowledge specific to the new job (firmspecific human capital). This process may require explicit training, or it may just mean that the new worker will not be fully F E D E R A L R E S E R V E B A N K O F S T. L O U I S 5 SEPTEMBER/OCTOBER 1997 4 The literature on incentive pay in economics studies the limited extent to which efficient employment relationships can be achieved in the face of this kind of slippage. A nice summary of this and other issues in incentive pay is Gibbons (1996). what they pay for: behavior that changes the measure of output rather than output itself (Baker, Gibbons, and Murphy, 1994). Fraud and accounting tricks often allow employees to manipulate the output measurement without changing output. Or, perhaps worse, easily observable quantity may rise at the expense of less apparent quality. The dilemma is summarized by Gibbons (1996): “When measured performance omits important dimensions of total contribution [to the firm], firms understand that they will ‘get what they pay for,’ and so may choose weak incentives in preference to strong but frequently dysfunctional incentives.” In other words, firms facing these types of distortions may choose to use incentive systems that are less direct and less precise than piece rates. The biggest impediment to the implementation of piece rates is that the output of individual workers is not easily measured in many jobs; reasonable, objective measures of performance do not exist. One reason is that it is usually difficult, if not impossible, to separate a particular worker’s performance from the overall performance of a group or firm. Inadequate output measurement makes piece rates far less effective. Firms’ motives for using “weak” incentives can be even deeper than obviously defective or easily manipulated measurement systems. Holmstrom and Milgrom (1994) argue that when workers perform several tasks, incentives must be finely balanced to ensure that all the tasks get adequate attention. But if, for example, one task is easy to measure and another, equally important, task is hard to measure (cooperation, for example), it will be impossible to use “strong” incentives—piece rates—to motivate performance on the first task without also inducing the worker to neglect the second task. Sometimes, therefore, firms may forego the opportunity to use piece rates (or use only weak ones), even when they would ostensibly be feasible and effective. Holmstrom and Milgrom conclude that “the use of low powered incentives within the firm, while sometimes lamented as one of the major disadvantages of internal organization, is also an important vehicle for inspiring cooperation and coordination.” Although a supervisor may be able to judge whether the worker is doing a good job over some period of time (we choose fuzzy words deliberately) and set pay accordingly, for two reasons this approach is not really a piece rate. First, evaluation by supervisors breaks the tight relationship between performance and pay that true piece rates can achieve in a simple environment.4 Second, it introduces a time dimension to the relationship between work and compensation that changes it in fundamental ways from the simple immediate reward system of piece rates. The remaining approaches discussed here stress this time dimension. PERFORMANCE BONDING In the face of workers’ inclinations to do various things contrary to the best interests of the firm, it is useful to divide compensation in two pieces. One piece is the level of compensation that the worker requires before agreeing to work for the firm at all. This piece includes any compensating differentials the firm must pay. For the next three paragraphs (only) we will call this component the base wage. The second piece of compensation convinces the worker to perform optimally—to work hard, stay sober, be unlikely to quit, and so forth. For the moment we will call this the bonus. If piece rates were feasible, this bonus could be zero. It might also be zero if it is easy to monitor the worker’s performance in relevant ways. As we argued above, these cases are far from universal. The base wage does not help motivate the worker, because it simply measures the alternative value of his time. It does not motivate him to do things he is disinclined to do (work hard, for example). Compensating differentials reflect the market’s valuation of things such as high effort, but if the employer cannot perfectly monitor the employee’s behavior, a compensating differential will not ensure that high effort F E D E R A L R E S E R V E B A N K O F S T. L O U I S 6 SEPTEMBER/OCTOBER 1997 is forthcoming. Clearly, the bonus, also, will not motivate workers if it is not conditional on performance in some way. So the firm must have a scheme whereby the worker is periodically evaluated and receives the bonus only if the evaluation suggests that his performance exceeds some threshold. Suppose that the evaluation is reasonably honest and accurate (closely related to the worker’s actual performance). If the bonus is big enough, it will provide adequate incentive for the worker to perform as the firm wants. How big it needs to be will depend on how likely it is that the firm’s evaluation will detect suboptimal performance. There is a flaw in this plan, however: The worker’s compensation (base wage plus bonus) may exceed the value of his marginal product if the bonus is too large. In this case, firms could simply decide it is not profitable to hire workers whose compensation exceeds the value of their marginal product and make no further effort to solve the agency problem. But here is a better idea: The firm could require the worker to give it some money at the beginning of the evaluation period and promise to pay it back with interest at the end, conditional on adequate performance. Now the firm is free to hire workers up to the point at which the value of the marginal product of labor equals the base wage because the workers are paying their own bonus. In economics jargon, they are posting a bond to guarantee their own performance. The firm still must compensate workers to do things they do not want to do (pay a compensating differential, in other words), but the bond guarantees that the firm will get what it pays for (if the bond is large enough to offset whatever temptations cause the firm’s agency problem). At first this idea appears to be a case of economic theory run amok. Jobs that require an explicit bond, as just described, are extremely rare, and this seems to be conclusive evidence against this theoretical approach. Indeed, Carmichael (1989) is blunt about this fact: “I know of no labor markets anywhere in the world or in history where this practice has been widespread.” But to write the idea off would be to underestimate the ingenuity of economic theorists. Work–Life Incentives Edward Lazear (1979, 1981, 1995) has argued that actual compensation plans implicitly use the bonding idea and, moreover, that recognizing this fact can help to explain some features of labor markets that otherwise appear quite odd. Lazear’s basic insight is that if firms and workers have full access to capital markets—that is, if they are able to save and borrow effectively—neither side should care whether workers’ compensation exactly equals the value of marginal product (VMP) on any given day. Instead, both care about the present value of wages and VMP over the working life of the employee. This observation suggests a new strategy that makes the performance bond an implicit part of compensation, rather than an explicit up-front payment. Lazear (1995) calls this approach “work–life incentives.” The same idea goes by various names, including “life-cycle incentives” and “upward-sloping age-earnings profiles” or “tenure-earnings profiles.” Lazear poses a simple agency problem: Suppose workers can work at either a high or a low effort level, and that they are indifferent among the options of working hard for wage W H, working at a low effort level for wage W L, or not having the job.5 W H and W L are the workers’ high- and low-effort reservation wages. (In reality, of course, firms must decide on an acceptable effort level, but adding that decision would not substantially change any part of this article.) Their difference, e, measures the monetary value to the worker of the extra effort. Employees who work hard are more productive than those who supply low effort, so VMP H > VMPL. Suppose that workers’ productivity at each effort level does not change during their lifetimes. A firm that could be sure its workers were working hard would pay W H and hire additional workers until VMP H fell to WH. A firm that knew its workers to be shirkers F E D E R A L R E S E R V E B A N K O F S T. L O U I S 7 5 This effort-supply problem is frequently used because, despite its extreme simplicity, it captures ideas like Lazear’s that are valid for many different and more complex situations. SEPTEMBER/OCTOBER 1997 difference between them at any time s between hiring and retirement is Figure 1 Work–Life Incentives Wage Profile Wage WT and is shown in Figure 2. This quantity is the value of a job that pays wages Wt from t= s until t=T. In Figure 2, the difference first rises as the initial negative Wt – WH terms get dropped off the beginning of the sum, and the positive ones get less discounting because they are not so far in the future (the term [1/(1+r)]t–s gets bigger as t – s gets smaller). Eventually, however, the terms getting dropped off the start of the sum are positive, and there are fewer and fewer terms to sum, so the difference falls. By retirement, the difference falls to WT – WH. At any point during his working life, a worker who chooses to work at low effort gets a utility gain e, but gambles that he will be caught (with probability d for detection) and lose a valuable job.6 This will be a good bet; that is, a risk-neutral worker will shirk, if 7 W W H WO T 6 Of course, the firm has some control over d. It should be understood here as a stand-in for how difficult in general it is to monitor an employee’s performance. 7 A risk-neutral worker is indifferent between accepting and rejecting a fair bet. A riskaverse worker would require a bigger gain from shirking to accept a given risk to his job. Because the worker does not lose Ws if he shirks and is caught in s (he gets paid up until the day he is fired), the wage profile must still be sloped a little bit even when d =1. 8 The worker always has an incentive to shirk in T because there is no stream of future payments left to lose. The firm could use a pension paid after T to give the job value in T (and before) as long as it could take the pension away up to the very last minute, if necessary. t would pay W L and hire until VMPL = W L. Some firms will choose the latter strategy, but if high effort is worth more to a particular firm than to workers (VMPH – VMPL > e), the firm will want to choose a compensation mechanism that persuades the worker to work at the high effort level. These are the firms with agency problems. For the reasons discussed above, paying workers WH throughout their careers will not by itself convince them to work hard, even though their pay includes a compensating differential for high effort. Even a threat of termination would do no good, because their next best option is just as desirable as a high-effort job at wage WH; that is what we mean by a reservation wage. In other words, the job itself has no value to the worker. A firm following this strategy gets low output for high wages, a losing proposition. Lazear observes that there is a simple way to make the job valuable to the worker. Consider the lifetime wage profile labeled W in Figure 1, which has been tilted so that the present value of wages paid on W between hiring at date t = 0 and retirement after date t = T equals the present value of a constant wage WH, that is, (1) In Figure 2, the worker will work hard up to time s**. By adjusting the slope of the W wage path (but leaving its present value unchanged), the firm can make s** equal T, thus giving workers incentives for adequate performance most of the time.8 Deferring compensation, as work–life incentives do, also discourages quits among current employees. An employee does not receive full compensation for past work until the end of his career; as a result, the job continues to have value and there is always an incentive to hang on a little longer. For a similar reason, work–life incentives also help to screen out applicants who, for one reason or another, would be more likely to quit: A worker who takes a job for just a year or two at a firm that uses work–life incentives is underpaid, since wages are initially below W H. where r is the interest rate. What happens to the difference between the present value of W and that of WH as time passes? The F E D E R A L R E S E R V E B A N K O F S T. L O U I S 8 SEPTEMBER/OCTOBER 1997 It is easy to see that work–life incentives can solve a broad range of agency problems. In fact, in principle this approach can be applied simultaneously to every agency problem mentioned in the introduction except encouraging better job applicants. The reason Lazear’s approach is so versatile is that it works entirely by making the worker’s job valuable enough that he will not risk dismissal; it does not matter how you interpret e, as long as the expected loss from undesirable behavior [the righthand side of (1)] is larger.9 Tournaments and efficiency wages function in the same way. Why can’t Lazear’s approach help improve a firm’s applicant pool? The job does not have value until the worker posts bond, which happens after hiring. As a careful look at Figure 2 reveals, the agency problem is not completely solved even when the profile is adjusted so that s** = T, because the value of the job is created by the accumulation of deferred pay, starting at zero. Initially, therefore, the value of the job is less than e/d, so some mechanism other than work–life incentives must be used to motivate workers during this interval.10 The firm could require the worker to post an explicit bond at the beginning. But that would remove a major attraction of Lazear’s theory, that it does not require outright (net) payments from workers to firms, which are rare. There is one last problem to wrap up. Since wages are at their highest late in life in Lazear’s model, workers have an incentive to hang on past T. The firm does not want this to happen because these high wages do not correspond to high current productivity; they are deferred compensation for past productivity. But this is not a flaw, it is a feature. Lazear (1979, 1995) observed that this “problem” could serve as an explanation of widespread mandatory retirement policies—policies that force employees to retire at a certain age, regardless of their productivity. Mandatory retirement policies are now illegal for most workers in the United States, but Lazear (1995) shows how defined-benefit pensions (pensions that promise a set Figure 2 Difference Between Present Values of Wage Profiles e/d WT – W H 0 monthly benefit, based on years of service and rate of pay) can also be structured to bring about timely retirement. Decreasing life expectancy (as the worker ages) causes the present value of any given benefit level to decline as retirement age increases. Since the firm sets the rate at which benefits increase with years of service, it can therefore determine the age at which the present value is maximized. If the worker chooses to work past the age preferred by the firm, the present value of his pension starts to decrease, even if the monthly benefit level is still increasing. The worker is thus given a strong financial incentive to retire at the age preferred by the firm. Another empirical implication of Lazear’s model, perhaps obvious enough to escape notice, is that earnings profiles slope upward throughout a worker’s career, even for workers who do not change jobs. This matches what labor economists have found in data on individual earnings histories. The upward-sloping earnings histories in the data do not seem to be fully explained by increasing productivity (human capital) as workers accumulate experience (Medoff and Abraham, 1980). Lazear’s analysis provides a supplementary reason for earnings to increase with experience.11 Tournaments Malcomson (1984) developed the idea that the internal hierarchy of a firm can be used as an effective incentive system.12 In F E D E R A L R E S E R V E B A N K O F S T. L O U I S 9 s** T s* t 9 There are probably limits on how large the right-hand side of (1) can be made in practice. We discuss these at the end of this section. 10 Akerlof and Katz (1989) pursue the implications of this observation. The problem disappears if we assume, as Lazear (1979, 1981) does, that shirking is detected with certainty at each instant in a continuous-time model. In that case the parameter that corresponds to d would be infinite. 11 Details of the nexus between seniority and wages are surveyed by Hutchens (1989). 12 Lazear and Rosen (1981) initiated the study of tournaments in labor economics. Studying the incentive effects of tournaments of the form, “The employee of the year will get a trip to Hawaii,” they showed that in some circumstances such tournaments could replicate the outcome of a piecerate system, but with the advantage of needing information only about the relative rankings of workers, instead of their absolute productivity levels. SEPTEMBER/OCTOBER 1997 13 The problem of motivating the individual(s) at the top of the hierarchy remains. This is, again, the problem of executive compensation mentioned in footnote 3. this type of model, a worker enters the firm at some level in a pyramid of possible jobs. The jobs at higher levels in the pyramid are rarer and pay more than those at the entry level. Periodically, the firm will promote a fraction of the employees from each level according to their ranking in some evaluation process, so that jobs at higher levels in effect become prizes in an ongoing tournament. The firm may supplement the prizes with terminations for employees who are not promoted. The chance of moving up and the competition needed to do so provide strong incentives for good performance.13 In a way similar to the work–life incentives described in the previous section, the size and number of prizes and penalties are set up so that the expected present value of compensation during a worker’s career equals the present value of his reservation wage. The incentives then operate in almost exactly the same way as work–life incentives: When the worker enters the hierarchy, he is initially paid less than the value of his marginal product (thus accumulating a bond). Wage increases are not certain in this model because luck (the quality of co-workers, for example), as well as effort, can influence success, but his expected lifetime earnings profile slopes up. High expected future income comes from a chance at promotions rather than increasing pay in the current job (as in Lazear’s model). Our comments in the previous section about quits apply here too. Tournaments have some problems similar to the “you get what you pay for” problem that plagues piece rates. Because promotions are based on relative evaluations, workers may collude to reduce output (though, as in other cartels, this strategy is prone to defections) or spend time sabotaging each other’s chances for promotion rather than working. From an economist’s point of view, the idea of hierarchies as incentive systems shares an attractive feature with work–life incentives. The logic of work–life incentives simultaneously solves an agency problem and provides an explanation for mandatory retirement, a phenomenon that had puzzled economists. Similarly, tournament models provide a workable solution to an agency problem, and they help explain why hierarchies exist at all, why firms often prefer to promote existing employees rather than to hire new ones, and why the variance of earnings within an organization is greater for employees with more seniority. Problems with Performance Bonding Few economists would dispute that mechanisms like those described in this section exist, and that managers of firms are aware of and try to exploit the incentives that the mechanisms provide. Controversy arises over whether compensation schemes based on the bonding principle can be pushed far enough to solve completely the motivation problems that firms face. This controversy is important because bonding models allow firms to solve their agency problems at no cost and without altering the basic principles of supply and demand in the labor market. Although these models break the tight link between wages and value of marginal product, firms still end up equating the two, but they are averaged over a worker’s lifetime or across workers who enter a tournament. Therefore, ex ante decisions are not affected by the use of performance bonding. Bonding produces, in economists’ jargon, first-best solutions. If first-best solutions exist—that is, if bonding approaches can fully solve the agency problem—they will presumably be firms’ preferred approach. Barriers to their use open the door to second-best solutions like efficiency wages, which are considered in the next section. The most important criticisms of performance bonding fall into four categories: imperfect financial markets, legal barriers, cheating (moral hazard) problems, and problems that come from hidden information. Explicit bond posting is rare in labor markets. In fact, it is unusual to see firms taking anything other than the job itself from workers who are fired. In other words, to the extent that bonding arrangements are used by firms, the value of the bond is somehow embedded in the job itself. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 10 SEPTEMBER/OCTOBER 1997 Understanding why explicit performance bonds are hardly ever used obviously helps explain why firms might choose roundabout practices like work–life incentives and tournaments. The near impossibility of using explicit performance bonds is also important because there are limits to the implicit bonding schemes Lazear and others have proposed. For example, it is easy to construct examples in which adequate work–life incentives (steep enough wage profiles) require negative wages early in a worker’s career. In other words, it is not always possible to tilt the wage path enough to get high effort and avoid explicit payments from workers to firms. Some other mechanism, such as efficiency wages, may therefore be necessary to change workers’ behavior sufficiently. So why are explicit performance bonds so rare? One reason may be that workers who are just starting a job have difficulty coming up with enough money to post an explicit bond. This conjecture challenges the assumption that workers can lend and borrow freely. Instead, they are liquidity constrained; they can save (lend), but their borrowing ability is limited. Dickens et al. (1989) discuss a second reason explicit bonds may not be a useful option: There are limits on the types of contracts that governments will enforce. In particular, under American and English common law, courts refuse to enforce contract provisions they interpret as penalties (as distinct from damages). When the probability of detecting workers’ misbehavior is low, performance bonds must be large because the disincentive to workers comes from the expected loss (the size of the bond times the probability of losing it), not the actual loss. Courts will typically not enforce contracts in which workers forfeit bonds that are disproportionately large. The courts do not, however, view firing as a penalty in this sense. Therefore implicit bonding arrangements are not limited by this legal standard. Implicit bonding also does not require explicit enumeration of the types and quantities of undesirable behavior that will result in penalties. Explicit contracts would be limited to a relatively small set of legally verifiable actions.14 The remaining problems with performance bonding apply to implicit bonds as well. In addition to the common-law legal principle just mentioned, many countries have laws that interfere with the use of performance bonds. In the United States, for example: (1) mandatory retirement is illegal for most workers; (2) minimum wage laws interfere with firms’ ability to pay very low wages to workers at the start of their careers; and (3) employers are required to vest workers in defined-benefit pension plans after five years. (This makes the job less valuable because it separates claim to a pension from continuation of the job.) One problem with performance bonds that stands out in most people’s minds is cheating by the firm, an example of moral hazard. If the worker’s performance were objectively verifiable, the employer could probably use piece rates or something like them. In most jobs, though, performance is judged, somewhat subjectively, by management. This gives the firm a clear incentive to misrepresent the worker’s performance in order to keep the bond. This is a compelling argument, but there are some considerations that mitigate it. First, since other workers usually have their own subjective evaluation of a worker who is fired, firms that regularly exploit this opportunity may develop a bad reputation. If either existing workers or new applicants recognize that there is a substantial chance that they will lose their bond even if they perform well, the bond no longer provides the desired incentive. In addition, workers would require compensation in some form, probably higher wages, for the expected loss of the bond. Second, promotion tournaments avoid the problem to a certain extent in the following way: If firms use a fixed number of prizes that will definitely be awarded according to the relative rankings of existing workers, the firms have no incentive to cheat. If they must fill the slots anyway, they are happy to fill them with the best workers. Of course, firms have F E D E R A L R E S E R V E B A N K O F S T. L O U I S 11 14 Hart (1995) contains a very persuasive discussion about the practical (and, thus, legal) limits of legal contracting. SEPTEMBER/OCTOBER 1997 15 The information is private in the technical sense that only the firm knows it, and it cannot be costlessly verified if the firm chooses to reveal it. The latter condition gives firms an opportunity for strategic misrepresentation. an incentive to avoid filling the slots at all (that is, awarding prizes) unless they serve some further function in the organization, but this ploy is easily observed by workers, so it would quickly destroy the incentive effects of the tournament. Ritter and Taylor (1997) argue that seniority-based layoffs have a similar advantage. Lower layoff probabilities for more experienced workers result in an upward-sloping experience-expected earnings profile, like that achieved by tournaments, even if the profile of actual wages is flat. The firm does not care which workers it lays off, since each is paid a wage equal to the value of his marginal product. It thus has no incentive to cheat. The final category of criticism is based on two principles: (1) that workers will insist on competitive rates of return on the bonds they post and (2) that the firm has better information about the rates of return that workers will actually receive than do the workers. If a business shuts down, workers who have posted bonds through low wages early in their careers lose the entire value of their bond. Similarly, if a firm hits a rough spot and responds by eliminating higher-level positions to make itself more competitive, prizes are removed from its promotion tournament, lowering the expected payoff to the bonds that workers posted by accepting low wages in entry-level positions. The first principle implies that workers will expect to be compensated for events like these. The second says that firms have private information about how likely these events are.15 Ritter and Taylor (1994) show that in these circumstances, risky firms (where workers insist on a higher rate of return on their bonds) have an incentive to pretend to be safe firms so that they can pay lower rates of return on the bonds. Workers, unable to distinguish between the two, require a rate of return above what they would demand from known safe firms. This makes performance bonding costly and, therefore, undesirable for safe firms, which separate themselves from risky firms by paying efficiency wages. EFFICIENCY WAGES Bonding mechanisms like work–life incentives and tournaments can provide an effective resolution to the agency problem because they make jobs valuable to workers. Workers have an investment for which the return is tied to continuation of the job. They are therefore less likely to quit or to take actions that would result in their dismissal. How should firms proceed if any of the economic or institutional reasons discussed above limit their use of performance bonds? The most obvious solution is to make jobs valuable in a direct manner— by paying more. The firm’s strategy here entails the use of a “carrot” and a “stick.” As in the work–life incentives model, the stick is the threat of dismissal. The carrot is the promise of a high-paying job. To see how this works, we return to the simple effort-supply problem that motivated our discussion of work–life incentives. We assumed for simplicity that workers could either work hard (high effort) or shirk (loaf). Workers have reservation wages W H and W L for high and low effort levels, which are related by W H = W L + e. We call e the difference in effort levels, but it is really the amount of money that makes the worker indifferent between high and low effort. Each day, a worker must decide whether to work hard or loaf. The consequences of this decision mirror those in the work–life incentives model: If he loafs, he gets immediate gratification worth e, but the probability is d that he will get caught, be fired, and lose a series of wages that exceed his reservation wage. Thus he will loaf at time s if (2) It is not a coincidence that (2) looks the same as (1); they express the same gamble for the worker. There are, however, two differences that are not immediately apparent from the inequality alone. In the work–life incentives model, the worker F E D E R A L R E S E R V E B A N K O F S T. L O U I S 12 SEPTEMBER/OCTOBER 1997 posts a bond by accepting Wt < W H early in the his career, so (1) does not hold at that stage. A premise of the efficiency-wage literature is that, for one reason or another, bonds cannot fully solve the agency problem. The crudest efficiency-wage models assume they cannot be used at all. In (2), therefore, _ W H all the time, but in (1), Wt < W H Wt > at the beginning of the worker’s career. An important consequence of not allowing a bond to be posted is that it is impossible both to solve the agency problem and to match the present value of a worker’s lifetime pay with the present value of his reservation wage; to solve the agency problem, the firm must pay an efficiency-wage premium. The efficiency wage is the lowest wage that will induce high effort, that is, the wage that would make (2) into an equality. Because the wage premium reduces profits, paying efficiency wages would be a second-best solution for the firm if some form of performance bonds could be used. Because it must pay a wage premium, an efficiency-wage firm demands less labor and produces less output than an otherwise identical firm that has no agency problem (or can solve its problem with performance bonds).16 Our formulation in (1) and (2) highlights that the problem might be solved by some combination of performance bonds and efficiency wages, depending on how far the firm can push performancebonding strategies. The second subtle difference between (2) and (1) is in the interpretation of the reservation wage, and it arises because of the possibility of involuntary unemployment. We postpone discussing this until the next section. The sum in (2) is the present value of efficiency-wage premiums—the value of the job relative to the reservation wage. To deter shirking, the firm must set the wage high enough to make the present value at least as great as e/d. Wages any higher than that would cut unnecessarily into profits. Thus the value of the job must always equal e/d.17 Figure 3 shows the lifetime wage profiles that come out of the efficiency-wage and work–life incentives Figure 3 Work–Life Incentives and EfficiencyWage Profiles Wage Work–Life Incentives Efficiency Wage 0 High-Effort Reservation Wage models using the same W H, e, and d.18 How do the solutions shown in Figure 3 change as the situation changes (across firms, for example)? First, if monitoring is more difficult (d is smaller) or more effort is required (e is higher), the efficiency wage will rise; larger carrots must be dangled to achieve optimal performance. In performance bonding models, bigger bonds are necessary—workers must give the firm larger carrots to be dangled in front of them. In the work–life incentives model, this means that the wage profile must be steeper, since the bond is accumulated during the phase in which the worker is underpaid. (For the same reason, s* increases.) A fall in d works on the cost side of the worker’s mental calculus. He recognizes that the chances of “getting caught” have fallen, and therefore a bigger penalty is required to induce him to forego a gain of e. An increase in e is simply an increase in the benefit of loafing and requires a more valuable job, so again the wage profile is steeper. Suppose there is always a chance that the job will end for reasons unrelated to performance. The worker’s wife could get an attractive job in a different city or the firm could shrink. We have not built this wrinkle into our simple versions of the models, but it is easy to apply the logic of the previous paragraph to see how this consideration affects the solutions. It all works through the value of the job. If a job separa- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 13 T t 16 Typically, efficiency-wage models assume that the firm still operates on its neoclassical labor demand curve; that is, it hires labor until the VMP equals the wage. Its equilibrium VMP is thus higher than the equilibrium VMP of an otherwise identical firm with no agency problem. 17 This means that effort-regulation efficiency-wage models share the problem of inducing high effort in T (when there are no future wage premiums) mentioned in footnote 8. Similar strategies would solve the problem. The problem does not arise in many of the other types of efficiency wage models described below. 18 The slope of the work–life profile in Figure 3 is set so that (1) holds with equality at T (s** = T ). The efficiency wage path, like the work–life profile, assumes that the incentive problem is solved somehow in T. That being the case, the value of the job is always e/d, which gives an efficiency wage of r e 1+ r d in every period. SEPTEMBER/OCTOBER 1997 that firms always prefer to pay for high effort. The competitive equilibrium wage is W HC, where supply equals demand. This is also the high-effort reservation wage; any worker paid less than W HC would immediately move into a comparable position with another firm. No worker would care about losing his job or whether his next job was working hard for W HC or not so hard for some loweffort reservation wage (not shown). Now introduce the agency problem. For the reasons given in previous sections, all workers would shirk at wage W HC, and firms would not be getting their money’s worth. In efficiency-wage models, firms make jobs more valuable to deter shirking. They do this individually by raising wages and reducing their own labor demand. Although they do not collude, their actions move them collectively up the high-effort demand curve to wage W E, where they employ only NE workers. The reservation wage is no longer W HC, because jobs are no longer available at that wage. Instead, the reservation wage in (2) is the wage that, combined with a high effort level, would make the worker as well off as remaining unemployed with a chance of getting a higher wage, W E, sometime in the future. This reservation wage may be above or below W HC, depending on the desirability of unemployment (which depends on things like the level of unemployment insurance benefits). In contrast to the competitive equilibrium, there is now involuntary unemployment N – NE in the sense that the workers who are unemployed would be willing to work at the prevailing wage. In the simple supply-and-demand model, firms never offer wages that differ much from the market-clearing wage W HC. If wages were above that level, workers who could not find jobs would offer to work at less than the going wage, bidding down the wage. In the efficiency wage equilibrium, workers without jobs cannot successfully underbid their employed neighbors. Recall that efficiency wages are chosen so that (2) becomes an equality. Suppose an unemployed worker approaches a firm’s manager, offering to work for less than the efficiency wage. The Figure 4 Efficiency–Wage Equilibrium Wage Supply W W E High-Effort Demand HC NE N Employment tion is more likely, there is a larger chance that the worker will never see some of the high wages promised in the future. This factor reduces the value of the job, so the firm must either pay a higher efficiency wage or require a larger bond. Using this reasoning, a firm that finds itself paying efficiency wages might also find it profitable to offer relatively stable employment, since stable employment would reduce the efficiency wage. Such a firm would sometimes operate off its VMP curve. Efficiency Wages and Unemployment We have described efficiency wages from the standpoint of a single firm. When Shapiro and Stiglitz (1984) first introduced a close relative of the efficiency-wage model presented above, their primary focus was on the implications of this model of compensation for unemployment rates. This section presents the core of their argument. Suppose there are lots of identical employers, each facing the same agency problem—encouraging high effort. There are also N workers who each supply one unit of labor, inelastically. If there were no agency problem, labor could be bought and sold like wheat. The applicable supplyand-demand graph would look like Figure 4. Suppose, as we have throughout this article, that the marginal product of effort is so high F E D E R A L R E S E R V E B A N K O F S T. L O U I S 14 SEPTEMBER/OCTOBER 1997 manager would like to pay the lower wage. But the manager understands that workers who are paid less than the efficiency wage will find it optimal to shirk, since lowering the wage makes the right-hand side of (2) less than the left-hand side. The unemployed worker’s offer is therefore declined. Since all firms behave in this same manner, unemployment persists in equilibrium. (When firms differ, the result can be dual labor markets, which we discuss shortly.) It is not hard to see how performance bonds would circumvent this problem and eliminate the involuntary unemployment. Suppose the unemployed worker approaches a firm, offering to work at less than the efficiency wage and offering to post a bond to be forfeited if he is detected shirking. A clever manager would understand that a big enough bond would deter shirking. The manager would accept this offer. This last point leads to two additional observations. First, firms that pay efficiency wages will, whenever possible, want to also use partial performance bonding. Worker bonds complement efficiency wages in coaxing high effort from workers, thus reducing the efficiency-wage premium. Second, efficiency wages and resulting unemployment persist only to the extent that firms cannot resolve the agency problem by using performance bonds. If bonding schemes were costless to implement, wages would be bid down to the competitive level and unemployment would disappear. suggests that there are some highly paid, stable jobs in which employees do work that is complicated and hard to measure. Many other jobs are characterized by simple work, poor pay, no job security, and little prospect of promotion. In short, as Doeringer and Piore (1971) argue, the American labor market seems to have a dual labor market with a “primary sector” of good jobs and a “secondary sector” of less desirable jobs. Dual labor market theorists like Doeringer and Piore argue that even hard-working, well-qualified workers in the secondary sector often cannot find employment in the primary sector. In a dual labor market, good workers can be stuck in bad jobs. In the basic supply-and-demand model, workers with equal ability and training who are doing equally difficult or distasteful work are paid the same. In this model, there may well be poorly paid jobs, but these jobs tend to have low-skill workers doing easy work. The supply-and-demand model predicts that equally productive workers will have similar lifetime earnings. The central idea of dual labor market theory—that good workers can be stuck in bad jobs—just doesn’t make sense in the competitive model. Pure performancebonding models also envision a perfectly competitive environment, so this observation applies there, too. Bulow and Summers (1986) argue that efficiency-wage models like the one we present here can provide an explanation for dual labor markets. Imagine a labor market in which all workers are identical but their jobs differ. In some jobs, low effort is acceptable or worker performance is easy to evaluate, so firms can effectively pay piece rates. Workers in this “secondary sector” receive a competitively determined wage. “Primary sector” jobs, in contrast, have agency problems that firms can resolve only by paying efficiency wages. All workers would like to have one of the valuable primary-sector jobs, but many well-qualified workers will end up in secondary-sector jobs. Critics of dual labor market theories argue that labor markets efficiently sort workers into appropriate jobs, given their Efficiency Wages and Dual Labor Markets The distinctive feature of efficiencywage jobs is that they are valuable from the start; they are jobs that people want but can’t easily get. The “carrot” that elicits high effort in an efficiency-wage job is the credible promise of high wages extending into the future. Efficiency-wage jobs also tend to offer stable employment. In addition, firms that pay efficiency wages might complement the efficiency-wage policy with performance bonds, so these jobs would have job ladders and pensions. A casual look at jobs in the economy F E D E R A L R E S E R V E B A N K O F S T. L O U I S 15 SEPTEMBER/OCTOBER 1997 ability, training, and inclinations. They argue that labor markets do not really produce primary and secondary sectors. Instead, markets sort workers according to characteristics that are not observable to labor economists (like willingness to work hard or cooperate with coworkers), creating the illusion of dual labor markets. Critics of efficiency-wage models also point out that if there really is a secondary sector, efficiency-wage models would not imply unemployment. Instead, people who could not get high-wage jobs would accept low-wage ones. In fact, this outcome depends on how the job search is modeled. If workers cannot search efficiently for primary-sector jobs while they are employed, the equilibrium level of unemployment will make workers indifferent between searching for a high-wage job while unemployed and accepting a lowwage job. This might still be interpreted as involuntary unemployment. reducing labor turnover is an important objective for managers. How does this problem affect compensation policy? An employee just starting out with a firm typically won’t know very much about nonwage features of the job. How difficult will the work be? Is it interesting? Are the working conditions pleasant? Will he like his boss and colleagues? Once he has spent time on the job, a typical worker will learn about these aspects of the job, and what he learns will affect his inclination to stay with the firm or seek employment elsewhere (while still employed). Indeed, the decision to quit or stay hinges on the value of the job (which in turn depends on both wage and non-wage features of the job) compared with the value of the alternative. One option for the firm is a low-wage, high-turnover strategy. The firm can simply set the wage at the lowest level necessary to fill vacancies immediately, fully understanding that many workers will quit as they discover undesirable nonwage aspects of the job. For firms with high turnover costs, though, a better strategy will be to reduce turnover by paying a wage higher than necessary to fill open jobs. As in the effort-regulation model, employers pay workers more than their reservation wages in order to alter their behavior. In the labor-turnover model, higher wages reduce recruiting and training costs and generate a more experienced labor force. Salop (1979) establishes that when all firms use this strategy, involuntary unemployment can persist in the economy. Also, if firms’ turnover costs differ, the market generates wage dispersion in which workers of equal ability receive different wages. Attracting Good Workers. Adverseselection models (Weiss, 1980, 1990) are based on another real-world problem that firms frequently encounter. A manager hiring a new worker wants to know how smart, conscientious, congenial, and motivated—in short, how productive—the worker is. The manager understands that workers have differing levels of productivity but can make only an informed guess Other Efficiency-Wage Models In the efficiency-wage model we outlined above, firms get higher productivity (less shirking) by paying workers more than their reservation wage. As we have seen, the market consequence of this employment-relations strategy can be dramatic. Most striking is the result that firms will not cut wages in response to involuntary unemployment, because cutting wages would reduce productivity. The effort-regulation problem we described is only one of a number of agency problems that have been addressed with efficiencywage models. The following hypotheses about how efficiency wages might arise differ from the widely used effort-supply model in using only carrots and no sticks; the firm does not use dismissals. Controlling Turnover. For many firms, orienting and training new employees can be an expensive, time-consuming activity. It can take months or even years for workers to become fully adjusted and productive in some work environments. Since firms face a big loss when employees join a firm only to quit a short time later, F E D E R A L R E S E R V E B A N K O F S T. L O U I S 16 SEPTEMBER/OCTOBER 1997 about the applicant’s productivity. Often firms learn about workers’ productivity only after the workers have been on the job for some time. In an extreme case, in which a firm can discern nothing about the future productivity of workers, the firm would have to resort to hiring at random from the pool of applicants. Now suppose that, in general, the most productive workers also have the best opportunities (as self-employed workers or employees in other firms), so that more productive workers have higher reservation wages. Then if a firm offers the lowest wage necessary to fill open positions, it will be choosing from among applicants with generally low productivity. As the firm increases the wage it offers, the pool of applicants expands to include better applicants, and the average productivity of the pool increases. The firm’s optimal strategy entails trading off higher wages against increased average productivity. Wage Norms. The models we have discussed so far are based on the general premise that workers act in their own narrowly defined interest. Akerlof (1982) set out a “sociological” perspective on worker behavior in which the employment relationship is viewed as a “gift exchange.” A firm that pays workers only the lowest wage necessary to get them to show up for work finds that workers reciprocate with minimal effort. A firm that gives workers a “gift” of higher wages (without requiring higher effort) finds that workers reciprocate with a “gift” of higher effort norms (which are enforced, in part, by peers). The model has characteristics similar to those of the basic effort-regulation model, but with behavioral foundations more similar to those hypothesized by sociologists than to the opportunistic utility maximization favored by economists. Annable (1988) advances a subtle argument about the formation and rigidity of wage norms, starting from the premise that “it is a tenet of personnel management that violations of established wage relationships will lead to worker dissatisfaction.” The wage relationships are both intertemporal and interpersonal and are established either spontaneously through “equity, custom, and tradition” or by explicit coordination activity among workers. The norms thus established translate into a relationship between wages and effort (broadly defined) that the firm will find difficult to influence. The firm must therefore take this relationship as given if it chooses the profit-maximizing wage, just as in the simple effort-regulation model. Annable argues that once a set of norms has been established, they will tend to be rigid because they are a public good for workers; the benefits of the coordination activity needed to change norms are shared by all workers, not just those bearing the cost of coordination. Avoiding Unionization. Union organizing entails different costs and benefits for workers and firms. The idea behind union-threat models is that by voluntarily giving workers one of the biggest benefits of unionization—higher wages—the firm can change the workers’ cost-benefit calculus. Workers would still bear the cost of unionization, but the marginal benefit would be lower. If the nonwage costs of unionization (less flexible employment policies, for example) are much higher for firms than the corresponding benefits to workers, the firm would find it worthwhile to follow this approach. Of course, the firm must also believe that there is a significant chance that a union will be successfully organized if they do not act. In the right circumstances (not in the middle of an open unionization effort, for example), the firm’s voluntary action could also be interpreted in Akerlof’s gift exchange framework. Workers, receiving the “gift” of higher wages, believe their employer is “fair” and see no need for a union. EMPIRICAL STUDIES Economic theory is most compelling when it provides plausible predictions of widespread phenomena, such as mandatory retirement, that are otherwise difficult to explain. In this section, however, we sample some of the more detailed F E D E R A L R E S E R V E B A N K O F S T. L O U I S 17 SEPTEMBER/OCTOBER 1997 19 Gibbons and Katz (1992) give a bit more equivocal reading of the evidence on inter-industry differentials. Thaler (1989) gives a concise overview. (but often ambiguous) empirical evidence that bears on these theories. The simple competitive supply-anddemand model implies that wages depend only on workers’ productivity and on attributes of firms or jobs that make the job more or less desirable. Characteristics such as the firm’s size or the ease of monitoring employees should not affect compensation. Suppose that a worker at firm A is paid less than a worker with comparable experience, skills, and so forth at firm B. In the supply-and-demand model, the firm A worker will go to firm B and offer to work for slightly less than the current firm B worker. In the competitive supply-and-demand paradigm, then, the law of one price holds, because workers arbitrage away price differences. This observation forms the basis for most econometric tests of the different compensation models. The performance-bonding models predict some additional relationships between wages and characteristics of workers and firms. Lazear’s work–life incentives model, for instance, predicts a positive relationship between wages and job tenure (length of time in present job) after controlling for overall work experience (as well as characteristics such as education-related worker productivity). The evidence on this relationship is supportive on balance, but it is somewhat muddied by technical econometric issues (Hutchens, 1989). Lazear’s theory also predicts that delayed-payment arrangements and collateral phenomena such as mandatory retirement will not be present when employees are easily monitored (a characteristic of the job, not the employee). Hutchens (1987) bases a test on the assumption that jobs involving repetitive tasks are, on average, more easily monitored and should therefore be characterized by absence of high wages for more senior workers, mandatory retirement, pensions, and long job tenures. Despite the fact that his measure of repetitive tasks is a very noisy proxy for ease of monitoring, Hutchens finds in the National Longitudinal Survey that jobs with more repetitive tasks are significantly less likely to exhibit the characteristics predicted by Lazear’s theory. Henry Ford is famous for deciding in 1914 to pay a wage well above the going rate. Raff and Summers (1987), who studied this episode intensively, say that “On balance it seems fair to conclude that Ford was able, by offering the five-dollar day, to reduce the turnover among his workers and to extract much more intensive, and generally productive, effort from them.” Ford’s policy thus had the main hallmarks of an efficiency wage: desirable effects on workers’ behavior brought about by wages above the level necessary to fill vacancies. A study by Krueger and Summers (1988) is one of a number that examine wage differentials across industries. The principle here is that, by and large, the industry in which a worker finds himself should not affect his wages in a competitive model. This observation applies to both the simple supply-and-demand model and the more sophisticated performancebonding models (as long as average age of employees does not differ across industries). They argue that systematically higher wages for workers in one industry than in another constitute evidence of efficiency wages. Krueger and Summers show that there are significant wage differentials across industries and use various types of data to argue that these cannot be attributed to employee demographics, human capital differences, compensating differentials, or unions. Although the existence of inter-industry wage differentials is not direct evidence of efficiency wages, Krueger and Summers seem to take the position that, after all other reasonable explanations have been ruled out, the only possibility left is efficiency wages.19 Murphy and Topel (1990) point out that a fully convincing explanation of interindustry wage differentials would link wages to features of industries that, according to efficiency-wage models, should generate different wages. Similar arguments have been made about the so-called employer-size effect; larger employers, on average, pay higher F E D E R A L R E S E R V E B A N K O F S T. L O U I S 18 SEPTEMBER/OCTOBER 1997 wages, which prove difficult to explain without efficiency wages. Rebitzer and Taylor (1995) pose a challenge to this line of reasoning. They study law firms—organizations in which there are obvious and dramatic promotion tournaments. Associates who are promoted to partner get very large increases in income, creating the presumption that the performance bonds created by the tournament are sufficient to generate high levels of effort, low quit rates, and so on. Rebitzer and Taylor show that the employer-size effect persists even in this environment, where the most common reasons for efficiency wages appear to be absent. Cappelli and Chauvin (1991) test one component of the efficiency-wage model. Using data from a single multi-plant automobile manufacturer, they test directly whether wage premiums result in lower levels of disciplinary action. All workers in their data were covered under the same collective bargaining agreement and the same disciplinary policies. By comparing the wages specified in the contract (the same for all plants) with the average hourly wage for production work in each plant’s Standard Metropolitan Statistical Area, Cappelli and Chauvin measure the wage premium paid at each plant. The premiums varied from 0 to 100 percent. They find fewer shirking-related disciplinary actions at plants with higher wage premiums. Their results provide support for a connection between pay and productivity. Because the firm was unionized, the existence of wage premiums does not imply the presence of efficiency wages, but the result does suggest that a union wage premium, by making the job valuable, acts as an efficiency wage. Of course a union contract that also makes disciplinary actions more difficult would offset that effect. Krueger (1991) compares compensation at company-owned and franchise-owned fast-food restaurants. This comparison controls automatically for different characteristics of workers and jobs. The two groups differ because managers of companyowned restaurants have less incentive to monitor employees than do ownermanagers. Thus the two groups can be presumed to have systematically different levels of monitoring (that is, d is higher for owner-operated restaurants). Krueger finds a small wage premium and steeper tenure-earnings profiles at companyowned outlets, results consistent with the efficiency-wage model. The steeper profile would also be implied by Lazear’s work–life incentives model, but the premium implies that the present value of lifetime wages is higher at companyowned outlets, for which Lazear’s model offers no rationale. Interestingly, the wage premiums are much higher for low-level managers than for regular workers. This finding suggests that the incentive problems faced in this industry are most efficiently solved by paying efficiency wages to supervisors to encourage more effective monitoring of production workers. On the other hand, using data on wages for narrowly defined occupations at 200 plants, Leonard (1987) finds that differing intensity of supervision across plants does not lead to the wage variation predicted by efficiency-wage models. We find this evidence less compelling than Krueger’s because the reason for variation in supervision intensity is unobserved. Without that information, it is difficult to know whether other relevant factors are really being held constant. CONCLUSION Employers and employees are often inclined to pursue goals that are at crosspurposes. The focus of this article is on economists’ hypotheses about how firms resolve this problem, and on the implications of these solutions for the structure of labor markets. Piece rates or incentive pay plans provide powerful direct incentives but have limited applicability. The performancebonding concept adds a valuable general perspective on employment practices such as job ladders, promotion tournaments, mandatory retirement, and pension policy. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 19 SEPTEMBER/OCTOBER 1997 Hart, Oliver. Firms, Contracts, and Financial Structure, Clarendon Press, 1995. These models form an important link between labor economics and the study of firm organization. Still, there are numerous legal, institutional, and economic impediments to the use of performance bonds, so it seems likely that firms’ best efforts to use this approach to motivating employees often fall short of completely resolving fundamental agency problems. Thus, even though efficiency wages are a second-best solution, they may often be needed as a complementary incentive device. Further, efficiency-wage theories present possible explanations for a number of additional labor market features, most notably involuntary unemployment. Holmstrom, Bengt, and Paul Milgrom. “The Firm as an Incentive System,” The American Economic Review (September 1994), pp. 972-92. Hutchens, Robert M. “Seniority, Wages and Productivity: A Turbulent Decade,” Journal of Economic Perspectives (Fall 1989), pp. 49-64. _______. “A Test of Lazear’s Theory of Delayed Payment Contracts,” Journal of Labor Economics (October 1987, part 2), pp. S153-70. Jensen, Michael C., and Kevin J. Murphy. “CEO Incentives—It’s Not How Much You Pay, but How,” Harvard Business Review (May–June 1990), pp. 138-49. Krueger, Alan B. “Ownership, Agency, and Wages: An Examination of Franchising in the Fast Food Industry,” Quarterly Journal of Economics (February 1991), pp. 75-101. _______ and Lawrence H. Summers. “Efficiency Wages and the Inter-Industry Wage Structure,” Econometrica (March 1988), pp. 259-93. REFERENCES Akerlof, George A. “Labor Contracts as Partial Gift Exchange,” Quarterly Journal of Economics (November 1982), pp. 543-69. Lazear, Edward P. “Why Is There Mandatory Retirement?” Journal of Political Economy (December 1979), pp. 1261-84. _______ and Lawrence F. Katz. “Workers’ Trust Funds and the Logic of Wage Profiles,” Quarterly Journal of Economics (August 1989), pp. 525-36. _______. “Agency, Earnings Profiles, Productivity, and Hours Restrictions,” The American Economic Review (September 1981), pp. 606-20. Annable, James. “Another Auctioneer is Missing,” Journal of Macroeconomics (Winter 1988), pp. 1-26. _______. Personnel Economics, The MIT Press, 1995. _______ and Sherwin Rosen. “Rank-Order Tournaments as Optimum Labor Contracts,” Journal of Political Economy (October 1981), pp. 841-64. Baker, George, Robert Gibbons, and Kevin J. Murphy. “Subjective Performance Measures and Optimal Incentive Contracts,” Quarterly Journal of Economics (1994), pp. 1125-56. Leonard, Jonathan S. “Carrots and Sticks: Pay, Supervision, and Turnover,” Journal of Labor Economics (October 1987), pp. S136-52. Bulow, Jeremy I., and Lawrence H. Summers. “A Theory of Dual Labor Markets with Application to Industrial Policy, Discrimination, and Keynesian Unemployment,” Journal of Labor Economics (October 1986), pp. 376-414. Malcomson, James M. “Work Incentives, Hierarchy, and Internal Labor Markets,” Journal of Political Economy (June 1984), pp. 486-507. Cappelli, Peter, and Keith Chauvin. “An Interplant Test of the Efficiency Wage Hypothesis,” Quarterly Journal of Economics (August 1991), pp. 769-87. Medoff, James, and Katharine Abraham. “Experience, Performance, and Earnings,” Quarterly Journal of Economics (December 1980), pp. 703-36. Carmichael, H. Lorne. “Self-Enforcing Contracts, Shirking, and Life Cycle Incentives,” Journal of Economic Perspectives (Fall 1989), pp. 65-83. Murphy, Kevin M., and Robert H. Topel. “Efficiency Wages Reconsidered: Theory and Evidence,” Advances in the Theory and Measurement of Unemployment, Yoram Weiss and Gideon Fishelson, eds., St. Martin’s Press, 1990, pp. 204-40. Dickens, William T., Lawrence F. Katz, Kevin Lang, and Lawrence H. Summers. “Employee Crime and the Monitoring Puzzle,” Journal of Labor Economics (July 1989), pp. 331-47. Raff, Daniel M. G., and Lawrence H. Summers. “Did Henry Ford Pay Efficiency Wages?” Journal of Labor Economics (October 1987), pp. S57-86. Doeringer, P. B., and M. J. Piore. Internal Labor Markets and Manpower Analysis, Heath, 1991. Gibbons, Robert. “Incentives and Careers in Organizations,” National Bureau of Economic Research Working Paper 5705, August 1996. Rebitzer, James, and Lowell J. Taylor. “Efficiency Wages and Employment Rents: The Employer Size Wage Effect in the Job Market for Lawyers,” Journal of Labor Economics (October 1995), pp. 678-708. _______ and Lawrence F. Katz. “Does Unmeasured Ability Explain Inter-industry Wage Differentials?” Review of Economic Studies (July 1992), pp. 515-35. Ritter, Joseph A., and Lowell J. Taylor. “Workers as Creditors: Performance Bonds and Efficiency Wages,” The American Economic Review (June 1994), pp. 694-704. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 20 SEPTEMBER/OCTOBER 1997 _______ and _______. “Seniority-Based Layoffs as an Incentive Device,” Federal Reserve Bank of St. Louis Working Paper 97-17A, November 1997. Shapiro, Carl, and Joseph Stiglitz. “Involuntary Unemployment as a Worker Discipline Device,” The American Economic Review (June 1984), pp. 433-44. Thaler, Richard H. “Anomalies: Inter-industry Wage Differentials,” Journal of Economic Perspectives (Spring 1989), pp. 181-93. Weiss, Andrew. “Job Queues and Layoffs in Labor Markets with Flexible Wages,” Journal of Political Economy (June 1980), pp. 526-38. _______. Efficiency Wages: Models of Unemployment, Layoffs, and Wage Dispersion, Princeton University Press, 1990. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 21 SEPTEMBER/OCTOBER 1997 Christopher J. Neely is an economist at the Federal Reserve Bank of St. Louis. Kent A. Koch provided research assistance. Technical Analysis in the Foreign Exchange Market: A Layman’s Guide Investors are concerned with “beating the market,” earning the best return on their money. Economists study technical analysis in foreign exchange markets because its success casts doubt on the efficient markets hypothesis, which holds that publicly available information, like past prices, should not help traders earn unusually high returns. Instead, the success of technical analysis suggests that exchange rates are not always determined by economic fundamentals like prices and interest rates, but rather are driven away from their fundamental values for long periods by traders’ irrational expectations of future exchange rate changes. These swings away from fundamental values may discourage international trade and investment by making the relative price of U.S. and foreign goods and investments very volatile. For example, when BMW decides where to build an automobile factory, it may choose poorly if fluctuating exchange rates make it difficult or impossible to predict costs of production in the United States relative to those in Germany. Despite the widespread use of technical analysis in foreign exchange (and other) markets, economists have traditionally been very skeptical of its value. Technical analysis has been dismissed by some as astrology. In turn, technical traders have frequently misunderstood what economists have to say about asset price behavior. What can the two learn from each other? This article provides an accessible treatment of recent research on technical analysis in the foreign exchange market. Christopher J. Neely Technical analysis suggests that a long-term rally frequently is interrupted by a short-lived decline. Such a dip, according to this view, reinforces the original uptrend. Should the dollar fall below 1.5750 marks, dealers said, technical signals would point to a correction that could pull the dollar back as far as 1.55 marks before it rebounded. Gregory L. White Wall Street Journal November 12, 1992 T echnical analysis, which dates back a century to the writings of Wall Street Journal editor Charles Dow, is the use of past price behavior to guide trading decisions in asset markets. For example, a trading rule might suggest buying a currency if its price has risen more than 1 percent from its value five days earlier. Such rules are widely used in stock, commodity, and (since the early 1970s) foreign exchange markets. More than 90 percent of surveyed foreign exchange dealers in London report using some form of technical analysis to inform their trading decisions (Taylor and Allen, 1992). In fact, at short horizons—less than a week—technical analysis predominates over fundamental analysis, the use of other economic variables like interest rates, and prices in influencing trading decisions. Investors and economists are interested in technical analysis for different reasons. A PRIMER ON TECHNICAL ANALYSIS IN FOREIGN EXCHANGE MARKETS Technical analysis is a short-horizon trading method; positions last a few hours or days. Technical traders will not hold F E D E R A L R E S E R V E B A N K O F S T. L O U I S 23 SEPTEMBER/OCTOBER 1997 existence of trends: Trends in motion tend to remain in motion unless acted upon by another force. The third principle of technical analysis is that history repeats itself. Asset traders will tend to react the same way when confronted by the same conditions. Technical analysts do not claim their methods are magical; rather, they take advantage of market psychology. Following from these principles, the methods of technical analysis attempt to identify trends and reversals of trends. These methods are explicitly extrapolative; that is, they infer future price changes from those of the recent past. Formal methods of detecting trends are necessary because prices move up and down around the primary (or longer-run) trend. An example of this movement is shown in Figure 1, where the dollar/deutsche mark ($/DM) exchange rate fluctuates around an apparent uptrend.2 To distinguish trends from shorter-run fluctuations, technicians employ two types of analysis: charting and mechanical rules. Charting, the older of the two methods, involves graphing the history of prices over some period—determined by the practitioner—to predict future patterns in the data from the existence of past patterns. Its advocates admit that this subjective system requires the analyst to use judgement and skill in finding and interpreting patterns. The second type of method, mechanical rules, imposes consistency and discipline on the technician by requiring him to use rules based on mathematical functions of present and past exchange rates. Figure 1 Peaks, Troughs, Trends, Resistance and Support Levels Illustrated for the $/DM $ per DM 0.72 0.70 Resistance level Sell signal from a 0.5% filter rule 0.68 0.66 0.64 Local troughs Trendline Local peak 0.62 0.60 May Buy signal from a 0.5% filter rule Support level June July Aug Sept 1992 NOTES: Not all buy and sell signals from the filter rule are identified. 1 2 These principles and a much more comprehensive treatment of technical analysis are provided by Murphy (1986) and Pring (1991). Rosenberg and Shatz (1995) advocate the use of technical analysis with more economic explanation. Figure 1 shows only closing prices. In this, it differs from most charts employed by technical traders, which might show the opening, closing, and daily trading range. positions for months or years, waiting for exchange rates to return to where fundamentals are pushing them. In contrast, fundamental investors study the economic determinants of exchange rates as a basis for positions that typically last much longer, for months or years. Some traders, however, use technical analysis in conjunction with fundamental analysis, doubling their positions when technical and fundamental indicators agree on the direction of exchange rate movements. Three principles guide the behavior of technical analysts.1 The first is that market action (prices and transactions volume) “discounts” everything. In other words, all relevant information about an asset is incorporated into its price history, so there is no need to forecast the fundamental determinants of an asset’s value. In fact, Murphy (1986) claims that asset price changes often precede observed changes in fundamentals. The second principle is that asset prices move in trends. Predictable trends are essential to the success of technical analysis because they enable traders to profit by buying (selling) assets when the price is rising (falling), or as technicians counsel, “the trend is your friend.” Practitioners appeal to Newton’s law of motion to explain the Charting To identify trends through the use of charts, practitioners must first find peaks and troughs in the price series. A peak is the highest value of the exchange rate within a specified period of time (a local maximum), while a trough is the lowest value the price has taken on within the same period (a local minimum). A series of peaks and troughs establishes downtrends and uptrends, respectively. For example, as F E D E R A L R E S E R V E B A N K O F S T. L O U I S 24 SEPTEMBER/OCTOBER 1997 shown in Figure 1, an analyst may establish an uptrend visually by connecting two local troughs in the data. A trendline is drawn below an apparent up trend or above an apparent downtrend. As more troughs touch the trendline without violating it, the technician may place more confidence in the validity of the trendline. The angle of the trendline indicates the speed of the trend, with steeper lines indicating faster appreciation (or depreciation) of the foreign currency. After a trendline has been established, the technician trades with the trend, buying the foreign currency if an uptrend is signaled and selling the foreign currency if a downtrend seems likely. When a market participant buys a foreign currency in the hope that it will go up in price, that participant is said to be long in the currency. The opposite strategy, called shorting or selling short, enables the participant to make money if the foreign currency falls in price. A short seller borrows foreign currency today and sells it, hoping the price will fall so that it can be bought back more cheaply in the future. Spotting the reversal of a trend is just as important as detecting trends. Peaks and troughs are important in identifying reversals too. Local peaks are called resistance levels, and local troughs are called support levels (see Figure 1). If the price fails to break a resistance level (a local peak) during an uptrend, that may be an early indication that the trend may soon reverse. If the exchange rate significantly penetrates the trendline, that is considered a more serious signal of a possible reversal. Technicians identify several patterns that are said to foretell a shift from a trend in one direction to a trend in the opposite direction. An example of the best-known type of reversal formation, called “head and shoulders,” is shown in Figure 2. The head and shoulders reversal following an uptrend is characterized by three local peaks with the middle peak being the largest of the three. The line between the troughs of the shoulders is known as the “neckline.” When the exchange rate penetrates the neckline of a head and Figure 2 The Head and Shoulders Reversal Pattern Illustrated for the $/DM $ per DM 0.66 Head 0.64 Left shoulder 0.62 Right shoulder Exchange rate penetrates the neckline sell signal Neckline 0.60 0.58 0.56 Sept Oct Nov Dec Jan Feb 1991-92 Mar shoulders, the technician confirms a reversal of the previous uptrend and begins to sell the foreign currency. There are several other similar reversal patterns, including the V (single peak), the double top (two similar peaks) and the triple top (three similar peaks). The reversal patterns of a downtrend are essentially the mirrors of the reversal patterns for the uptrend. Mechanical Rules Charting is very dependent on the interpretation of the technician who is drawing the charts and interpreting the patterns. Subjectivity can permit emotions like fear or greed to affect the trading strategy. The class of mechanical trading rules avoids this subjectivity and so is more consistent and disciplined, but, according to some technicians, it sacrifices some information that a skilled chartist might discern from the data. Mechanical trading rules are even more explicitly extrapolative than charting; they look for trends and follow those trends. A wellknown type of mechanical trading rule is the “filter rule,” or “trading range break” rule which counsels buying (selling) a currency when it rises (falls) x percent above (below) its previous local minimum (maximum). The size of the filter, x, which is F E D E R A L R E S E R V E B A N K O F S T. L O U I S 25 Apr May SEPTEMBER/OCTOBER 1997 exchange rate over a given number of previous trading days. The length of the moving average “window”—the number of days in the moving average—governs whether the moving average reflects long- or short-run trends.3 Any moving average will be smoother than the original exchangerate series, and long moving averages will be smoother than short moving averages. Figure 3 illustrates the behavior of a 5-day and a 20-day moving average of the exchange rate in relation to the exchange rate itself. A typical moving average trading rule prescribes a buy (sell) signal when a short moving average crosses a longer moving average from below (above)—that is, when the exchange rate is rising (falling) relatively fast. Of course, the lengths of the moving averages must be chosen by the technician. The length of the short moving average rule is sometimes chosen to equal one, the exchange rate itself. A final type of mechanical trading rule is the class of “oscillators,” which are said to be useful in non-trending markets, when the exchange rate is not trending up or down strongly. A simple type of oscillator index, an example of which is shown in Figure 4, is given by the difference between two moving averages: the 5-day moving average minus the 20-day moving average. Oscillator rules suggest buying (selling) the foreign currency when the oscillator index takes an extremely low (high) value. Note that the oscillator index, as a difference between moving averages, also generates buy/sell signals from a moving average rule when the index crosses zero. That is, when the short moving average becomes larger than the long moving average, the moving average rule will generate a buy signal. By definition, this will happen when the oscillator index goes from negative to positive. Therefore, an oscillator chart is also useful for generating moving average rule signals. Figure 3 5- and 20-Day Moving Averages $ per DM 0.65 Exchange rate 0.64 0.63 5-Day moving average 20-Day moving average 0.62 Sell signal, moving average rule 0.61 0.60 Buy signal, moving average rule 0.59 Feb Mar Apr May Jun 1992 NOTES: These moving averages smooth the exchange rate and can be used to generate buy and sell signals in the foreign exchange market. Figure 4 The Oscillator Index Normalized difference in moving averages 1.0 0.8 0.6 0.4 0.2 0 –0.2 –0.4 –0.6 –0.8 –1.0 Oscillator rule sell signals Moving average rule sell signal Moving average rule buy signal Difference in moving averages Oscillator rule buy signal Feb Mar Apr May Jun 1992 NOTES: The 5-day moving average minus the 20-day moving average can also be used to generate buy and sell signals. 3 For example, the five-day moving average of an exchange rate series is given by: M(5)t = 1 5 4 i =0 S t-i where S t denotes the closing price of the spot exchange rate at day t. chosen by the technician from past experience, is generally between 0.5 percent and 3 percent. Figure 1 illustrates some of the buy and sell signals generated by a filter rule with filter size of 0.5 percent. A second variety of mechanical trading rule is the “moving average” class. Like trendlines and filter rules, moving averages bypass the short-run zigs and zags of the exchange rate to permit the technician to examine trends in the series. A moving average is the average closing price of the Other Kinds of Technical Analysis Technical analysis is more complex and contains many more techniques than those described in this article. For F E D E R A L R E S E R V E B A N K O F S T. L O U I S 26 SEPTEMBER/OCTOBER 1997 example, many technical analysts assign a special role to round numbers in support or resistance levels. When the exchange rate significantly crosses the level of 100 yen to the dollar, that is seen as an indication that further movement in the same direction is likely.4 Other prominent types of technical analysis use exotic mathematical concepts such as Elliot wave theory and/or Fibonacci numbers.5 Finally, traders sometimes use technical analysis of one market’s price history to take positions in another market, a practice called intermarket technical analysis. the investment with the higher expected return. While the U.S. and German interest rates are known, the bank must base its decision on its forecast of the rate of appreciation of the DM. If market participants expect the return to investing in the German money market to be higher than that of investing in the U.S. money market, they will all try to invest in the German market, and none will invest in the U.S. money market. Such a situation would tend to drive down the German return and raise the U.S. return until the two were equalized. The excess return on a German investment over an investment in the U.S. money market (Rt DM), at date t, from the point of view of a U.S. investor is defined as EFFICIENT MARKETS AND TECHNICAL ANALYSIS Technical analysts believe that their methods will permit them to beat the market. Economists have traditionally been skeptical of the value of technical analysis, affirming the theory of efficient markets that holds that no strategy should allow investors and traders to make unusual returns except by taking excessive risk.6 (1) RtDM ; itDM + D St – it $, where itDM is the German overnight interest rate, D St is the percentage rate of appreciation of the DM against the dollar overnight, and it$ is the U.S. overnight interest rate.7 If market participants cared only about the expected return on their investments, and if their expectations about the change in the exchange rate were not systematically wrong, the expected excess return on foreign exchange should equal zero, every day. The assumption that market participants care only about the expected return is too strong, of course. Surely, participants also care about the risk of their investment.8 Risk can come from either the risk of default on the loan or the risk of sharp changes in the exchange rate, or both. If investing in the German market is significantly riskier than investing in the U.S. market, investors must be compensated with a higher expected return in the German market, or they will not invest there. In that case, the expected excess return would be positive and equal to a risk premium. The expected riskadjusted excess return would be equal to zero. That is, Investing in the Foreign Exchange Market To understand the efficient markets hypothesis in the context of foreign exchange trading, consider the options open to an American bank (or firm) that temporarily has excess funds to be invested overnight. The bank could lend that money in the overnight bank money market, known as the federal funds market. The simple net return on each dollar invested this way would be the overnight interest rate on dollar deposits. The bank has other investment options, though. It could instead convert its money to a foreign currency (e.g., the deutsche mark), lend its money in the overnight German money market (at the German interest rate) and then convert it back to dollars tomorrow. This return is the sum of the German overnight interest rate and the change in the value of the DM. Which investment should the bank choose? If the bank were not concerned about risk, it would choose (2) E[RtDM] – RPt = 0, where E[*] is a function that takes the expected value of the term inside the F E D E R A L R E S E R V E B A N K O F S T. L O U I S 27 4 “The 100 yen level for the dollar is still a very big psychological barrier and it will take a few tests before it breaks. But once you break 100 yen, it’s not going to remain there for long. You’ll probably see it trade between 102 and 106 for a while,” said Jorge Rodriguez, director of North American Sales at Credit Suisse, as reported by Creswell (1995). 5 Murphy (1986) discusses Elliot wave theory, Fibonacci numbers, and many other technical concepts. 6 Samuelson (1965) did seminal theoretical work on the modern theory of efficient markets. 7 The excess return may also be considered the return to someone borrowing in dollars and investing those dollars in German investments. 8 Market participants may be concerned about the liquidity of their position as well as the expected return and risk. Liquidity is the ease with which assets can be converted into cash. SEPTEMBER/OCTOBER 1997 brackets [*] and RPt is the risk premium associated with the higher risk of lending in the German market. How do prices move in the hypothetical efficient market? In an efficient market, profit seekers trade in a way that causes prices to move instantly in response to new information, because any information that makes an asset appear likely to become more valuable in the future causes an immediate price rise today. If prices do move instantly in response to all new information, past information, like prices, does not help anyone make money. If there were a way to make money with little risk from past prices, speculators would employ it until they bid away the money to be made. For example, if the price of an asset rose 10 percent every Wednesday, speculators would buy strongly on Tuesday, driving prices past the point where anyone would think they could rise much further, and so a fall would be likely. This situation could not lead to a predictable pattern of rises on Tuesday, though, because speculators would buy on Monday. Any pattern in prices would be quickly bid away by market participants seeking profits. Indeed, there is considerable evidence that markets often do work this way. Moorthy (1995) finds that foreign exchange rates react very quickly and efficiently to news of changes in U.S. employment figures, for example. Because the efficient markets hypothesis is frequently misinterpreted, it is important to clarify what the idea does not mean. It does not mean that asset prices are unrelated to economic fundamentals.10 Asset prices may be based on fundamentals like the purchasing power of the U.S. dollar or German mark. Similarly, the hypothesis does not mean that an asset price fluctuates randomly around its intrinsic (fundamental) value. If this were the case, a trader could make money by buying the asset when the price was relatively low and selling it when it was relatively high. Rather, “efficient markets” means that at any point in time, asset prices represent the market’s best guess, based on all currently available information, as to the fundamental value of the asset. Future price changes, adjusted for risk, will be close to unpredictable. But if any pattern in prices is quickly bid away, how does one explain the Efficient Markets 9 There are a number of versions of the efficient markets hypothesis. This version is close to that put forward by Jensen (1978). 10 For an example of an incorrect interpretation of the efficient markets hypothesis, see Murphy (1986, p. 20-21) who offers, “The theory is based on the efficient markets hypothesis, which holds that prices fluctuate randomly about their intrinsic value. . . . it’s just unrealistic to believe that all price movement is random.” The idea that the expected risk-adjusted excess return on foreign exchange is zero implies a sensible statement of the efficient markets hypothesis in the foreign exchange context: Exchange rates reflect information to the point where the potential excess returns do not exceed the transactions costs of acting (trading) on that information.9 In other words, you can’t profit in asset markets (like the foreign exchange market) by trading on publicly available information. This description of the efficient markets hypothesis appears to be a restatement of the first principle of technical analysis: Market action (price and transactions volume) discounts all information about the asset’s value. There is, however, a subtle but important distinction between the efficient markets hypothesis and technical analysis: The efficient markets hypothesis posits that the current exchange rate adjusts to all information to prevent traders from reaping excess returns, while technical analysis holds that current and past price movements contain just the information needed to allow profitable trading. What does this version of the efficient markets hypothesis imply for technical analysis? Under the efficient markets hypothesis, only current interest rates and risk factors help predict exchange rate changes, so past exchange rates are of no help in forecasting excess foreign exchange returns—i.e., if the hypothesis holds, technical analysis will not work. Malkiel’s summary of the attitude of many economists toward technical analysis in the stock market is based on similar reasoning: The past history of stock prices cannot be used to predict the future in any meaningful way. Technical strategies are usually amusing, often comforting, but of no real value. (Malkiel, 1990, p. 154.) F E D E R A L R E S E R V E B A N K O F S T. L O U I S 28 SEPTEMBER/OCTOBER 1997 apparent trends seen in charts of asset prices like those in Figure 1? Believers in efficient markets point out that completely random price changes—like those generated by flipping a coin—will produce price series that seem to have trends (Malkiel, 1990, or Paulos, 1995). Under efficient markets, however, traders cannot exploit those trends to make money, since the trends occur by chance and are as likely to reverse as to continue at any point. (For example, some families have—purely by chance—strings of either boys or girls, yet a family that already has four girls and is expecting a fifth child still has only a 50 percent chance of having another girl.) published and taken to indicate that trading rule strategies can yield profits. For example, there is a vast literature on pricing anomalies in the equity markets, summarized by Ball (1995) and Fortune (1991), but Roll (1994) has found that these aberrations are difficult to exploit in practice; he suggests that they may be partially the result of data mining. Trading Rules With these considerations, two kinds of trading rules have been commonly tested: filter rules and moving average rules. As a preceding section of this article explained, filter rules give a buy signal when the exchange rate rises x percent over the previous recent minimum. The analyst must make two choices to construct a filter rule: First, how much does the exchange rate have to rise, or what is the size of the filter? Second, how far back should the rule go in finding a recent minimum? The filter rules studied here will use filters from 0.5 percent to 3 percent and go back five business days to find the extrema.12 A moving average rule gives a buy signal when a short moving average is greater than the long moving average; otherwise it gives a sell signal. This rule requires the researcher to choose the lengths of the moving averages. The moving average rules to be tested will use short moving averages of 1 day and 5 days and long moving averages of 10 days and 50 days. Both the filter rules and the moving average rules are extrapolative, in that they indicate that the trader should buy when the exchange rate has been rising and sell when it has been falling. EVALUATING TECHNICAL ANALYSIS The efficient markets hypothesis requires that past prices cannot be used to predict exchange rate changes. If the hypothesis is true, technical analysis should not enable a trader to earn profits without accepting unusual risk. This section examines how two common types of trading rules are formulated and how the returns generated by these rules are measured. Problems inherent in testing the rules, measuring risk, and drawing conclusions about the degree of market efficiency are discussed.11 Finding a Trading Rule A basic problem in evaluating technical trading strategies is that rules requiring judgement and skill are impossible to quantify and therefore unsuitable for testing. A fair test requires fixed, objective, commonly used trading rules to evaluate. An “objective” rule does not rely on individual skill or judgement to determine buy or sell decisions. The rule should be commonly used to reduce the problem of drawing false conclusions from “data mining”— a practice in which many different rules are tested until, purely by chance, some are found to be profitable on the data set. Negative test results are ignored, while positive results are 11 A number of previous studies have documented evidence of profitable technical trading rules in the foreign exchange market: Sweeney (1986); Levich and Thomas (1993); Neely, Weller, and Dittmar (1997). 12 As with most aspects of technical analysis, the choice of filter size and window lengths has been determined by practitioners through a process of trial and error. Profits The trading rules switch between long and short positions in the foreign currency. Recall that a long position is a purchase of foreign currency—a bet that it will go up—while a short position is the reverse, selling borrowed foreign currency now in the hope that its value will fall. Denoting the percentage change in the exchange rate F E D E R A L R E S E R V E B A N K O F S T. L O U I S 29 SEPTEMBER/OCTOBER 1997 Table 1 Technical Trading Rule Results for the $/DM Moving Average Rule Results Monthly Standard Deviation Number of Trades Sharpe Ratio Estimated CAPM Beta Standard Error of Est. Beta Short MA Long MA Annual Return 1 1 5 5 10 50 10 50 6.016 7.546 6.718 6.671 2.979 3.155 3.064 3.236 928 268 576 146 0.583 0.690 0.633 0.595 – 0.022 – 0.135 – 0.144 – 0.134 0.091 0.085 0.084 0.080 Filter Annual Return Monthly Standard Deviation Number of Trades Sharpe Ratio Estimated CAPM Beta Standard Error of Est. Beta 0.005 0.010 0.015 0.020 0.025 0.030 5.739 6.438 3.323 1.934 0.839 – 1.541 3.057 2.951 3.255 3.348 3.236 3.578 1070 584 382 234 142 92 0.542 0.630 0.295 0.167 0.075 – 0.124 – 0.071 – 0.092 – 0.037 – 0.128 – 0.118 – 0.086 0.089 0.093 0.085 0.087 0.082 0.077 Filter Rule Results NOTES: The first two columns of the top panel characterize the length of the short and long moving averages used in the moving-average trading rule. The third column is the annualized asset return to the rule, while the fourth column is the monthly standard deviation of the return. The fifth column is the number of trades over the 23-year sample. The sixth column is the Sharpe ratio, and the last two columns provide the CAPM beta with the S&P 500 and the standard error of that estimate. The lower panel has a similar structure, except that the first column characterizes the size of the filter used in the rule. All extrema for filter rules were measured over the previous five business days. ($ per unit of foreign currency) from date t to t+1 by DSt, and the domestic (foreign) overnight interest rate by it$ (itDM), then the overnight return from a long position is approximately given by Equation 1: 13 14 The estimate of transactions costs used here is consistent with recent figures. Levich and Thomas (1993) consider a round-trip cost of 0.05 percent realistic, as do Osler and Chang (1995). The exchange rate data were obtained from DRI and were collected at 4:00 p.m. local time in London from Natwest Markets and S&P Comstock. Daily overnight interest rates are collected by BIS at 9:00 a.m. London time. Interest rates for Japan were unavailable before 3/1/82, so the interest rates before this date were set to 0 for the $/¥ case. (1) average of daily U.S. dollar bid and ask quotes for the DM, yen, pound sterling, and Swiss franc.14 All exchange rate data begin on 3/1/74 and end on 4/10/97. These four series are called $/DM, $/¥, $/£, and $/SF. Because the results for the four exchange rates were similar, full results from only the $/DM will be reported in the tables. Table 1 shows the annualized percentage return, monthly standard deviation (a measure of the volatility of returns), number of trades per year, and two measures of risk, the Sharpe ratio and the CAPM beta, for each of the 10 trading strategies for the $/DM. The Sharpe ratio and CAPM betas are discussed in some detail in the shaded insert. The mean annual return to the 10 rules was 4.4 percent, and 38 of the 40 trading rules were profitable (had positive excess return) over the whole sample. These results cast doubt on the efficient markets hypothesis, which holds that no trading strategy should be able to consistently earn positive excess RtDM ; itDM + DSt – it$. The return to a short position is the negative of the return to a long position. The return to a trading rule over a period of time is approximately the sum of daily returns, minus transactions costs for each trade. Transactions costs are set at 5 basis points (0.05 percent) for each round trip in the currency. A round trip is a move from a long position to a short position and back or vice-versa.13 Evidence from Ten Simple Technical Trading Rules Six filter rules and four moving average rules were tested on data consisting of the F E D E R A L R E S E R V E B A N K O F S T. L O U I S 30 SEPTEMBER/OCTOBER 1997 returns. The number of trades over the 23-year sample varied substantially over the 10 rules, ranging from 4 trades per year to almost 50 trades per year. The moving average rules were somewhat more profitable than the filter rules. There is little evidence that these excess returns are compensation for bearing excessive risk. The first measure of risk, the Sharpe ratio, is the mean annual return divided by the mean annual standard deviation. The moving average rules had higher Sharpe ratios (0.6 vs. 0.25) than the filter rules. Six of the 10 Sharpe ratios are better than the 0.3 obtained by a buy-and-hold strategy in the S&P 500 over approximately the same period. This result indicates that the average return to the rules is very good compared to the risk involved in following the rules. The second measure of risk, the CAPM betas, reflects the correlation between the monthly trading rule returns and the monthly returns to a broad portfolio of risky assets (the S&P 500). Significantly positive betas indicate that the rule is bearing undiversifiable risk. These CAPM betas estimated from the 10 rules generally indicate negative correlation with the S&P 500 monthly returns. None of them is significantly positive, statistically or economically. In other words, there is no systematic risk in these rules that could explain the positive excess returns. Figure 5 One-Year Moving Average, ForwardLooking Excess Returns to the (1,10) Moving Average Trading Rule Moving Average of Annual Return 50 Excess return to $/DM (1,10) moving average rule 40 30 20 10 0 –10 Excess return to S&P 500 index –20 –30 1974 76 78 80 82 84 86 88 90 March 1974-March 1996 up with the market on a daily basis. How large would transactions costs have to be to eliminate the excess return to the technical rules? If we assume a 6 percent annual excess return to the rule and 230 trades (10 trades a year), round-trip transactions costs would have to be greater than 0.6 percent to produce zero excess returns. In addition to higher transactions costs, individual investors following technical rules also must accept the risk that such a strategy entails. Figure 5 illustrates the risk by depicting, at monthly intervals, the one-year-ahead excess return from 1974 through 1996 for the (1,10) moving average rule on the $/DM and, for comparison, the total excess return on buying and holding the S&P 500 index, a popular measure of returns to a stock portfolio. The figure shows that the excess returns to both portfolios vary considerably at the annual horizon, often turning negative. While the technical trading rule excess return is less variable than the S&P excess return, it can still lead to significant losses for some subperiods. Two ways to measure losses over subperiods are the maximal single-period loss (maximum drawdown) and maximum loss in a calendar year. Over the period from March 1974 through March 1997, the maximum For Whom is Technical Trading Appropriate? The discussion of risk and returns suggests that technical analysis may be very useful for banks and large financial firms that can borrow and lend freely at the overnight interbank interest rate and buy and sell in the wholesale market for foreign exchange, where transactions sizes are in the millions of dollars. Technical trading is much less useful for individuals, who would face much higher transactions costs and must consider the opportunity cost of the time necessary to become an expert on foreign exchange speculating and to keep F E D E R A L R E S E R V E B A N K O F S T. L O U I S 31 92 94 96 SEPTEMBER/OCTOBER 1997 that an investor could have lost by using the moving average trading rule was –28.2 percent; this loss, which would have occurred between March 7, 1995, and August 2, 1995 (a period of 149 days), translates into an annual rate of – 69.2 percent. In other words, an investor using this rule would have lost almost 30 percent of his capital over this five-month period. Similarly, the maximum loss for this technical trading rule in a complete calendar year was – 9.8 percent in 1995, but – 17.8 percent for the S&P 500 in 1981.15 Perhaps the biggest obstacle to exploiting technical rules is that while the returns to stocks depend ultimately on the profitability of the firms in which the stock is held, the source of returns to technical analysis is not well understood; therefore, the investor does not know if the returns will persist into the future or even if they continue to exist at the present. Indeed, Figure 5 shows that the post-1992 return to the (1,10) moving average rule for the $/DM has been negative. the degree of inefficiency. Risk is notoriously difficult to measure. In fact, a major area of study for macro and financial economists for the last 10 years has been to explain why the return on stocks is so much higher than that on bonds, a phenomenon called the equity premium puzzle. Of course, at least part of the answer is that stocks are much riskier than bonds, but there is no generally accepted model of risk that will explain the size of the return difference.16 Defenders of the efficient markets hypothesis maintain that the discovery of an apparently successful trading strategy may not indicate market inefficiency but, rather, that risk is not measured properly. Another problem is that of “data mining”: If enough rules are tested, some— purely by chance—will produce excess returns on the data. These rules may not have been obvious to traders at the beginning of the sample. In fact, the rules tested here are certainly subject to a data-mining bias, since many of them had been shown to be profitable on these exchange rates over at least some of the subsample. Closely related to the data-mining problem is the tendency to publish research that overturns the conventional wisdom on efficient markets, rather than research that shows technical analysis to be ineffective. One solution to the data-mining problem is suggested by Neely, Weller, and Dittmar (1997), who apply genetic programming techniques to the foreign-exchange market. Genetic programming is a method by which a computer searches through the space of possible technical trading rules to find a group of good rules (i.e., rules that generate positive excess return). These good rules are then tested on out-of-sample data to see if they continue to generate positive excess returns. Do These Results Measure the Degree of Market Efficiency? 15 The returns for complete calendar years were available from 1975 through 1995. 16 Kocherlakota (1996) and Siegel and Thaler (1997) discuss the equity premium puzzle extensively. There are a number of problems associated with inferring the degree of market efficiency from the apparent profitability of these trading rules. The first problem is the data. To test the profitability of a trading rule, the researcher needs actual prices and interest rates from a series of simultaneous market transactions. Unfortunately, simultaneous quotes for daily exchange rates and interest rates are not generally available for a long time span. For example, these exchange-rate data were collected late in the afternoon, while the interest rates were collected in the morning. Although most economists judge this problem to be very minor, some argue that the trading rule decisions could not have been executed at the exchange rates and interest rates used. The second problem is that without a good model of how to price risk, positive excess returns resulting from the use of trading rules cannot be used to measure RETHINKING THE EFFICIENT MARKETS HYPOTHESIS Early research in finance on the efficient markets hypothesis was very supportive; little evidence was found of profitable trading rules after transactions costs were accounted for (Fama, 1970). F E D E R A L R E S E R V E B A N K O F S T. L O U I S 32 SEPTEMBER/OCTOBER 1997 Empirical Reasons to Suspect Failure of Efficient Markets The success of technical trading rules shown in the previous section is typical of a number of later studies showing that the simple efficient markets hypothesis fails in important ways to describe how the foreign exchange market actually functions. While these results did not surprise market practitioners, they have helped persuade economists to examine features of the market like sequential trading, asymmetric information, and the role of risk that might explain the profitability of technical analysis. The miserable empirical performance of standard exchange rate models is another reason to suspect the failure of the efficient markets hypothesis. In an important paper, Meese and Rogoff (1983) persuasively showed that no existing exchange rate model could forecast exchange rate changes better than a “no-change” guess at forecast horizons of up to one year. This was true even when the exchange rate models were given true values of future fundamentals like output and money. Although Mark (1995) and others have demonstrated some forecasting ability for these models at forecasting horizons greater than three years, no one has been able to convincingly overturn the Meese and Rogoff (1983) result despite 14 years of research. The efficient markets hypothesis is frequently misinterpreted as implying that exchange rate changes should be unpredictable; that is, exchange rates should follow a random walk. This is incorrect. Equation 2 shows that interest rate differentials should have forecasting power for exchange rate changes, leaving excess returns unpredictable. There is, however, convincing evidence that interest rates are not good forecasters of exchange rate changes.17 According to Frankel (1996), this failure of exchange rate forecasting leaves two possibilities: The Paradox of Efficient Markets Grossman and Stiglitz (1980) identified a major theoretical problem with the hypothesis termed the paradox of efficient markets, which they developed in the context of equity markets. As applied to the foreign exchange market, the argument starts by noting that exchange rate returns are determined by fundamentals like national price levels, interest rates, and public debt levels, and that information about these variables is costly for traders to gather and analyze. The traders must be able to make some excess returns by trading on this analysis, or they will not do it. But if markets were perfectly efficient, the traders would not be able to make excess returns on any available information. Therefore, markets cannot be perfectly efficient in the sense of exchange rates’ always being exactly where fundamentals suggest they should be. Of course, one resolution to this paradox is to recognize that market analysts can recover the costs of some fundamental research by profiting from having marginally better information than the rest of the market on where the exchange rate should be. In this case, the exchange rate remains close enough to its fundamental value to prevent less informed people from profiting from the difference. Partly for these reasons, Campbell, Lo, and MacKinlay (1997) suggest that the debate about perfect efficiency is pointless and that it is more sensible to evaluate the degree of inefficiency than to test for absolute efficiency. • Fundamentals are not observed well enough to allow forecasting of exchange rates. • Exchange rates are detached from fundamentals by (possibly irrational) swings in expectations about future values of the exchange rate. These fluctuations in exchange rates are known as bubbles.18 Which of these possibilities is more likely? One clue is given by the relationship between exchange rates and fundamentals when expectations about the value of the exchange rate are very stable, as they are under a fixed exchange rate F E D E R A L R E S E R V E B A N K O F S T. L O U I S 33 17 Engel (1995) reviews the failure of this theory, called uncovered interest parity. 18 Swings in expectations that are subsequently justified by changes in the exchange rate are known as rational bubbles. Swings that are not consistent with the future path of exchange rates are irrational bubbles. SEPTEMBER/OCTOBER 1997 When Germany and the United States ceased to fix their currencies in March 1973, the variability in the real $/DM exchange rate increased dramatically. This result suggests that, contrary to the efficient markets hypothesis, swings in investor expectations may detach exchange rates from fundamental values in the short run. Figure 6 Monthly Percentage Changes in the $/DM Real Exchange Rate 16 12 8 4 0 Why Do Bubbles Arise? –4 If traders might profit by anticipating swings in investor expectations, then the efficient markets hypothesis needs significant adjustment. The structure of the foreign exchange market has several features that might help drive these swings in expectations that produce bubbles. Most foreign exchange transactions are conducted by large commercial banks in financial centers like London, New York, Tokyo, and Singapore. These large banks “make a market” in a currency by offering to buy or sell large quantities (generally more than $1 million) of currencies for a specific price in another currency (e.g., the dollar) on request. The exchange rates at which they are willing to buy or sell dollars are known as the bid and ask prices, respectively. The market is highly competitive, and transactions occur 24 hours a day over the telephone and automated trading systems. The first feature of this market that might influence technical trading is that specific transactions quantities and prices are not public information; the market is nontransparent. But the bid and ask exchange rates are easy to track, as banks freely quote them to any participant. Second, the trades take place sequentially—i.e., there is time to learn from previous trades. Third, the participants in this market differ from one another in the information they have and their willingness to tolerate risk.19 In other words, the participants are heterogeneous. How might these features combine to produce bubbles? To the extent that some participants are better informed about certain fundamentals than other agents (for instance, they will know more about their own and their customers’ demand for foreign exchange), the trading behavior of –8 –12 1962 66 70 74 78 82 86 90 94 98 NOTES: These changes become much more volatile after the end of the Bretton Woods system of fixed exchange rates in March 1973. The vertical line denotes this break date in the series. Data cover January 1960–February 1997. 19 It has long been assumed that there is little or no private information in foreign exchange markets, but this view has been forcefully challenged with respect to intraday trading by Ito, Lyons, and Melvin (1997). regime. A fixed exchange rate regime is a situation in which a government is committed to maintaining the value of its currency by manipulating monetary policy and trading foreign exchange reserves. Fixed exchange rate regimes are contrasted to floating regimes, in which the government has no such obligation. For example, most countries in the European Union had a type of fixed exchange rate regime, known as a target zone, from 1979 through the early 1990s. Fixed exchange rates anchor investor sentiment about the future value of a currency because of the government’s commitment to stabilize its value. If fundamentals, like goods prices, or expectations based on fundamentals, rather than irrationally changing expectations, drive the exchange rate, the relationship between fundamentals and exchange rates should be the same under a fixed exchange rate regime as it is under a floating regime. This is not the case. Countries that move from floating exchange rates to fixed exchange rates experience a dramatic change in the relationship between prices and exchange rates. Specifically, real exchange rates (exchange rates adjusted for inflation in both countries) are much more volatile under floating exchange rate regimes, where expectations are not tied down by promises of government intervention. Figure 6 illustrates a typical case: F E D E R A L R E S E R V E B A N K O F S T. L O U I S 34 SEPTEMBER/OCTOBER 1997 the informed participants will reveal some of their private information to the uninformed agents. For example, if the informed agents know of fundamental forces that are likely to make the exchange rate rise in the future, they are likely to buy the foreign currency and thereby bid up the publicly observed bid and ask prices. The uninformed agents might infer from the rise that the rate will continue to rise and, as a result, they might buy more foreign exchange, pushing the rate up themselves in a self-fulfilling prophecy.20 This inference from past price behavior is extrapolative technical analysis: It assumes that the exchange rate will continue moving as it has in the recent past. The uninformed traders may continue to buy foreign exchange past the point where it is supported by fundamentals. Although this story is most plausible for very high-frequency (intraday) trading, it might also generate longer-term swings in the exchange rate. There are other explanations for extrapolative trading that jettison the assumption of rational behavior in favor of the study of how people really make decisions. This field, called behavioral finance, has concentrated on examples of seeming irrationality in decision making. Two findings of this field are that (1) experimental participants seem unusually optimistic about their chances for success in games and (2) the behavior and opinions of members of a group tend to reinforce common ideas or beliefs.21 For example, members of a jury may become more confident about their individual verdicts if the other members of the group agree. Either explanation for extrapolative trading implies that bubbles may be produced by slow dissemination of private information into the market, coupled with extrapolative trading rules. There is some evidence to support this explanation. Eichenbaum and Evans (1995) found that foreign exchange markets reacted gradually to money supply shocks, over a period of many months, instead of instantly incorporating the new information. Surveys revealed that foreign exchange market participants’ expectations are extrapolative at horizons up to six months. That is, if the exchange rate has risen recently, market participants expect it to continue to rise in the near future (Frankel and Froot, 1987). Also, the success of extrapolative traders tends to feed on itself. Frankel and Froot (1990) argue that extrapolative traders’ success during the early part of the large dollar appreciation of 1981-1985 convinced many other traders to follow extrapolative rules, driving the dollar up even further. Central Bank Intervention The other popular explanation for the apparent profitability of technical trading rules is that technical traders are able to profit consistently from central bank intervention. Some central banks frequently intervene (buy and sell currency) in the foreign exchange market to move the exchange rate to help influence other variables like employment or inflation.22 Because these actions are designed to control macroeconomic variables rather than to make money, central banks may be willing to take a loss on their trading. Trading rule profits may represent a transfer from central banks to technical traders. Lebaron (1996) found that most trading rule profits were generated on the day before a U.S. intervention. Neely and Weller (1997) find that “intelligent” trading rules tend to trade against the Fed; that is, they tend to buy dollars when they find out the Fed is selling dollars. This tantalizing story does not fit all the facts, however. For example, Leahy (1995) finds that U.S. foreign exchange operations make positive profits, on average.23 This finding is inconsistent with the idea that central banks are giving money away to technical traders. 20 Treynor and Ferguson (1985), Brown and Jennings (1989), Banerjee (1992), and Kirman (1993) construct models of behavior in which information is inferred from the actions of others. One easily understood example is the problem of consumers who must choose between two restaurants. One seemingly sensible strategy for choosing would be to go to the more crowded restaurant on the theory that it is likely to be crowded because it has better food. This phenomenon depends on asymmetric information. 21 Shiller (1988) and Shleifer and Summers (1990) discuss behavioral finance in more detail. Ohanian (1996) considers the reasons for the collapse of bubbles. 22 In the United States, the Federal Reserve and the U.S. Treasury generally collaborate on foreign exchange intervention decisions, and operations are conducted by the Federal Reserve Bank of New York on behalf of both. 23 See Szakmary and Mathur (1996) for more on central bank intervention and trading rule profits. Why Are the Profits Not Arbitraged Away? Whether the trends or inefficiencies in exchange rates are created by swings in expectations or by central bank intervention, efficient market advocates would ask why any predictable returns in exchange rates F E D E R A L R E S E R V E B A N K O F S T. L O U I S 35 SEPTEMBER/OCTOBER 1997 HOW TO MEASURE RISK? The simplest widely used measure of risk is the Sharpe ratio or the ratio of the average annual excess return to a measure of excess return volatility called the standard deviation. Higher Sharpe ratios are desirable because they indicate either higher average excess returns or less volatility. A commonly used benchmark of a good Sharpe ratio is that of the S&P 500, which Osler and Chang (1995) estimated to be about 0.32 from March 1973 to March 1994. A major drawback to Sharpe ratios is that they ignore an important idea in finance: An investment is risky only to the extent that its return is correlated with the return to a broad measure of the investments available. To see this, consider the risk associated with holding a portfolio of assets whose returns are each individually volatile but completely independent of each other. Each year, the assets in the portfolio that do unusually well will tend to offset those that do unusually poorly. The portfolio as a whole will be much less risky than any of the individual assets. The more assets in the portfolio, the less risky it will be. In fact, if enough of these independent assets are grouped together into a portfolio, the return on this portfolio becomes certain. This means that investors do not need to be compensated for holding risky assets that are not correlated with all the other assets they can buy (the market portfolio), because the risk of each uncorrelated asset can be reduced to zero if the portfolio contains a large enough variety of these assets. On the other hand, assets for which returns are positively correlated with those of the other assets on the market need a higher expected return to convince investors to hold them. This idea motivates the second measure of riskiness, the CAPM beta: the coefficient from the linear regression of an asset’s (or trading rule’s) excess return on the excess return of a proxy for the market portfolio, the return to a broad equity index like the S&P 500. An estimated beta equal to zero means that the trading rule is bearing no systematic risk, while significantly positive betas indicate that a trading strategy is bearing some risk, and a beta equal to one means that the trading rule moves closely with the market, so that following it requires the investor to accept significant risk. 24 Both Shleifer and Summers (1990) and Shleifer and Vishny (1997) discuss the importance of risk in speculating against bubbles. 25 Essentially the same argument is presented more simply in Shleifer and Summers (1990). should not be arbitraged away. One answer to this question is that speculators have short horizons and are deterred from speculating against the trends by the risk that such a strategy would incur. There are several reasons for this: First, traders typically operate on margin, borrowing some of the money with which they trade. With a limited line of credit, the borrowing costs would add up if traders were not able to turn a quick profit. Second, a trader’s performance is typically evaluated on relatively short horizons (less than a year). Third, there may be institutional or legal restrictions that prevent some types of enterprises from taking on “excessive” exchange risk. And finally, traders do not know the equilibrium value of the exchange rate with any certainty, so they cannot distinguish bubbles from movements in fundamentals. Investors who bet on longrun reversion to fundamental values in exchange rates may be wiped out by shortrun deviations away from those values.24 Explaining the success of technical trading rules with bubbles begs one more question: Why do destabilizing extrapolative traders not lose their money? Friedman (1953) showed that destabilizing speculation is doomed to lose money and so drive the speculators out of the market. Friedman argued that speculation can only destabilize asset prices if the speculators consistently buy when the asset price is above its equi- F E D E R A L R E S E R V E B A N K O F S T. L O U I S 36 SEPTEMBER/OCTOBER 1997 librium value (driving the price up further) and sell when the asset price is below its equilibrium value; as the destabilizing speculators lose their money, he maintained, they will have less effect on the market. The corollary to this argument is that all successful speculation is stabilizing. Delong, Schleifer, Summers, and Waldman (1989) constructed a “noise trader” model that questioned this logic, however.25 They showed that irrational (“noise”) traders could create so much risk in asset markets that the returns to those assets would have to be unusually high for rational traders to trade in them at all. In other words, the irrational traders make unusually high returns (on average) by foolishly pursuing risky strategies. Some go out of business, but, on average, this group increases its market position. transactions data or experimental work on expectations formation may provide a better understanding of market behavior. CONCLUSION Eichenbaum, Martin, and Charles L. Evans. “Some Empirical Evidence on the Effects of Shocks to Monetary Policy on Exchange Rates,” Quarterly Journal of Economics (November 1995), pp. 975-1009. REFERENCES Ball, Ray. “The Theory of Stock Market Efficiency: Accomplishments and Limitations,” Journal of Applied Corporate Finance (Spring 1995), pp. 4-17. Banerjee, Abhijit V. “A Simple Model of Herd Behavior,” Quarterly Journal of Economics (August 1992), pp. 797-817. Brown, David P., and Robert H. Jennings. “On Technical Analysis,” The Review of Financial Studies (1989), pp. 527-51. Campbell, John Y., Andrew W. Lo, and Archie Craig MacKinlay. The Econometrics of Financial Markets, Princeton University Press, 1997. Creswell, Juli. “Currency Market Expects Rate Cut By Bank of Japan,” Wall Street Journal, September 5, 1995, p. C16. DeLong, J. Bradford, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann. “Noise Trader Risk in Financial Markets,” Journal of Political Economy (August 1990), pp. 703-38. Technical analysis is the most widely used trading strategy in the foreign exchange market. Traders stake large positions on their interpretations of patterns in the data. Economists have traditionally rejected the claims of technical analysts because of the appealing logic of the efficient markets hypothesis. More recently, however, the discovery of profitable technical trading rules and other evidence against efficient markets have led to a rethinking about the importance of institutional features that might justify extrapolative technical analysis such as private information, sequential trading, and central bank intervention, as well as the role of risk. The weight of the evidence now suggests that excess returns have been available to technical foreign exchange traders over long periods. Risk is hard to define and measure, however, and this difficulty has obscured the degree of inefficiency in the foreign exchange market. There is no guarantee, of course, that technical rules will continue to generate excess returns in the future; the excess returns may be bid away by market participants. Indeed, this may already be occurring. Continued research on high-frequency Engel, Charles. “Why is the Forward Exchange Rate Forecast Biased? A Survey of Recent Evidence,” Federal Reserve Bank of Kansas City Working Paper 95-06, September 1995. Fama, Eugene F. “Efficient Capital Markets: A Review of Theory and Empirical Work,” Journal of Finance (May 1970), pp. 383-417. Fortune, Peter. “Stock Market Efficiency: An Autopsy?,” New England Economic Review, Federal Reserve Bank of Boston (March/April 1991), pp. 18-40. Frankel, Jeffrey. “How Well Do Foreign Exchange Markets Function: Might a Tobin Tax Help?,” National Bureau of Economic Research Working Paper 5422, January 1996. _______ and Kenneth A. Froot. “Using Survey Data to Test Standard Propositions Regarding Exchange Rate Expectations,” The American Economic Review (March 1987), pp. 133-53. _______ and _______. “Chartists, Fundamentalists and the Demand for Dollars,” Private Behavior and Government Policy in Interdependent Economies, Anthony S. Courakis and Mark Taylor, eds., Clarendon Press, 1990. Friedman, Milton. “The Case for Flexible Exchange Rates,” Essays in Positive Economics, University of Chicago Press, 1953. Grossman, Sanford J., and Joseph E. Stiglitz. “On the Impossibility of Informationally Efficient Markets,” The American Economic Review (June 1980), pp. 393-408. Ito, Takatoshi, Richard K. Lyons, and Michael T. Melvin. “Is There Private Information in the FX Market? The Tokyo Experiment,” National Bureau of Economic Research Working Paper 5936, February 1997. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 37 SEPTEMBER/OCTOBER 1997 Paulos, John Allen. A Mathematician Reads the Newspaper, Basic Books, 1995. Jensen, Michael C. “Some Anomalous Evidence Regarding Market Efficiency,” Journal of Financial Economics (June/September 1978), pp. 95-101. Pring, Martin J. Technical Analysis Explained, Third Edition, McGraw-Hill, 1991. Kirman, Alan. “Ants, Rationality and Recruitment,” Quarterly Journal of Economics (February 1993), pp. 137-56. Kocherlakota, Narayana R. “The Equity Premium: It’s Still a Puzzle,” Journal of Economic Literature (March 1996), pp. 42-71. Roll, Richard. “What Every CFO Should Know About Scientific Progress in Financial Economics: What is Known, and What Remains to be Resolved,” Financial Management (Summer 1994), pp. 69-75. Leahy, Michael P. “The profitability of US intervention in the foreign exchange markets,” Journal of International Money and Finance (December 1995), pp. 823-44. Rosenberg, Michael R., and Eric A. Shatz. The Merrill Lynch Guide to Understanding and Using Technical Analysis, Merrill Lynch and Co., Global Securities Research & Economics Group, 1995. Lebaron, Blake. “Technical Trading Rule Profitability and Central Bank Intervention,” National Bureau of Economic Research Working Paper 5505, March 1996. Samuelson, Paul A. “Proof that Properly Anticipated Prices Fluctuate Randomly,” Industrial Management Review (1965), pp. 41-49. Shiller, Robert J. “Fashions, Fads and Bubbles in Financial Markets,” Knights, Raiders and Targets: The Impact of the Hostile Takeover, J. Coffee, S. Ackerman, and L. Lowenstein, eds., Oxford University Press, 1988. Reprinted in Market Volatility, by Robert J. Shiller, MIT Press, 1989. Levich, Richard M., and Lee R. Thomas. “The Significance of Technical Trading-Rule Profits in the Foreign Exchange Market: A Bootstrap Approach,” Journal of International Money and Finance (October 1993), pp. 451-74. Malkiel, Burton G. A Random Walk Down Wall Street, Fifth Edition, W. W. Norton & Company, 1990. Shleifer, Andrei, and Lawrence Summers. “The Noise Trader Approach to Finance,” Journal of Economic Perspectives (Spring 1990), pp. 19-33. Mark, Nelson C. “Exchange Rates and Fundamentals: Evidence on Long-Horizon Predictability,” The American Economic Review (March 1995), pp. 201-18. _______ and Robert W. Vishny. “The Limits of Arbitrage,” Journal of Finance (March 1997), pp. 35-55. Meese, Richard A., and Kenneth Rogoff. “Empirical Exchange Rate Models of the Seventies: Do They Fit out of Sample?,” Journal of International Economics (February 1983), pp. 3-24. Siegel, Jeremy J., and Richard H. Thaler. “Anomalies: The Equity Premium Puzzle,” Journal of Economic Perspectives (Winter 1997), pp. 191-200. Moorthy, Vivek. “Efficiency Aspects of Exchange Rate Response to News: Evidence from U.S. Employment Data,” Journal of International Financial Markets, Institutions and Money (1995), pp. 1-18. Sweeney, Richard J. “Beating the foreign exchange market,” Journal of Finance (March 1986), pp. 163-82. Murphy, John J. Technical Analysis of the Futures Markets, New York Institute of Finance, Prentice-Hall, New York, 1986. Szakmary, Andrew C., and Ike Mathur. “Central Bank Intervention and Trading Rule Profits in Foreign Exchange Markets,” Journal of International Money and Finance (August 1997), pp. 513-35. Neely, Chris, and Paul Weller. “Technical Analysis and Central Bank Intervention,” Federal Reserve Bank of St. Louis Working Paper 97002A, January 1997. Taylor, Mark P., and Helen Allen. “The use of technical analysis in the foreign exchange market,” Journal of International Money and Finance (June 1992), pp. 304-14. _______, _______, and Robert Dittmar. “Is Technical Analysis Profitable in the Foreign Exchange Market? A Genetic Programming Approach,” Forthcoming in Journal of Financial and Quantitative Analysis (December 1997). Treynor, Jack L., and Robert Ferguson. “In Defense of Technical Analysis,” Journal of Finance (July 1985), pp. 757-73. Ohanian, Lee E. “When the Bubble Bursts: Psychology or Fundamentals?,” Business Review, Federal Reserve Bank of Philadelphia (January/February 1996), pp. 3-13. Osler, Carol L., and P. H. Kevin Chang. “Head and Shoulders: Not Just a Flaky Pattern,” Federal Reserve Bank of New York Staff Paper 4, August 1995. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 38 SEPTEMBER/OCTOBER 1997 Michael R. Pakko is an economist at the Federal Reserve Bank of St. Louis. Kelly M. Morris and Eran Segev provided research assistance. The Business Cycle and Chain-Weighted GDP: Has Our Perspective Changed? Figure 1 Quarterly Real GDP Growth 1987 Dollars and Chain-Weighted 1992 Dollars Percent (ar) 8 6 4 2 0 –2 –4 Michael R. Pakko –6 1989 I n early 1996 the Bureau of Economic Analysis released data from an extensive revision to its measures of aggregate economic activity in the United States. The new data and methodology, which represent the first comprehensive revision of the National Income and Product Accounts (NIPA) since 1991, reflect more substantial changes than have many previous revisions.1 While past revisions have changed some definitions and statistics and have incorporated newly available source data, the most recent revision includes an even more important change: the move from a fixed base-year measure to a chained index, which is said to have significantly altered the way we view recent economic performance. One analysis in the New York Times even speculated that the outcome of the 1992 election might have been different if we had known then what we know now about the 1990-91 recession: 1990 1991 1987 Dollars 1992 1993 1994 Chain-Weighted 1992 dollars it actually did, might have changed the impression among many voters that President Bush did not care.2 Figure 1 illustrates the effect of the changes on the 1990-91 recession and subsequent recovery. During the three quarters of negative economic growth, real GDP registered a 1.8 percent average annual rate of decline under the old methodology. The revised figures show an average rate of decline of more than 2.7 percent. Moreover, the new figures indicate a slower growth rate in the period of economic recovery that followed the recession. The old figures showed a 2.8 percent growth rate from the first quarter of 1991 to the first quarter of 1994, while the new figures indicate 2.5 percent growth. Two limitations of the old system of measurement have been cited to support the claim of superiority for the new chainweighting approach: First, fixed weights fail to measure the effects of shifting demand in response to relative price changes—known as “substitution bias.” Second, the periodic re-basing required by the fixed-weight system of measurement implies a “rewriting of economic history.”3 The revisions suggest that the downturn of 1990-91 was nearly twice as deep as Washington knew at the time, raising the possibility that the Federal Reserve and the White House would have reacted more aggressively to dig the economy out of a recession. Such a response, coming much earlier than F E D E R A L R E S E R V E B A N K O F S T. L O U I S 39 1 This was the tenth such “comprehensive” revision. See Parker and Seskin (1996). 2 Passell (1996). 3 E.g., Samuelson (1996). SEPTEMBER/OCTOBER 1997 But has the revision to NIPA data really changed our view of the fluctuations and economic co-movements known collectively as the “business cycle”? Both casual observation and more rigorous empirical examination reveal that the revisions to GDP and its components have had very little effect on the empirical regularities that characterize economic fluctuations. After first describing the methodological changes and rationales underlying the recent NIPA data revisions, this article discusses the differences between the revised data and the previously reported data. It then presents a comparison of the basic business cycle regularities as measured under the old and new systems. explicitly reflected in the national accounts. More importantly, government expenditures have been separated into categories for current expenditures and government investment expenditures. The addition of the latter category—which classifies government expenditures for structures and equipment as investment— provides a more complete measure of investment, including purchases by both the public and private sectors. The change also makes figures for U.S. investment more comparable to those of other countries by treating government investment in a way that is more consistent with the International System of National Accounts. By far the most significant and widely discussed revision to the NIPA data is the new methodology for reporting real output and its components: the “chained” dollar series. The previously used “constantdollar” measures were based on prices for a specific base year. (Prior to the most recent revision, the base year was 1987.) The new method uses contemporaneous price data to produce estimates of growth rates for real output and its components. The prices are then chained together to form a series that is independent of the choice of any particular base year. This approach mitigates the two related weaknesses of the constant-dollar approach: “substitution bias” and “rewriting economic history.” REVISIONS TO NIPA DATA 4 For a complete description of the statistical and methodological revisions to the NIPA accounts, see United States Department of Commerce– Bureau of Economic Analysis (1995). The revisions to NIPA included new and revised source data, new methodological procedures, the new chain-weighted methodology for reporting real output and its components, and an updated base year.4 Some of the more routine changes reflect the incorporation of new and revised source data. For example, the revisions to the data on non-durable consumption expenditures for 1993 and 1994 are based on newly available information on retail sales from the 1993 Annual Retail Trade Survey, while the revisions to data on nondurable goods are based on the results of a 1987 input-output analysis constructed by the Commerce Department. Revisions to the services consumption data are based on direct estimates of rental payments for tenant-occupied dwellings, taken from a 1991 Residential Finance Survey. In addition, the Commerce department has incorporated some new methodological procedures. For example, it has replaced the previously used straight-line method of estimating depreciation with one based on studies of the prices of used equipment and structures in resale markets. There are also two significant changes in the classification of government expenditures. Federal Government contributions to civilian and military retirement programs (rather than benefits paid out) are now Substitution Bias A fundamental purpose of the new chain-weighting methodology is to avoid the problem known as “substitution bias.” As relative prices rise and fall, purchasers tend to substitute less expensive items for ones on which prices have gone up. In a fixed-weight system, items with falling prices continue to constitute a large share of the total, because this system is based on their historically higher prices. Similarly, although sales tend to decline for goods and services items with rising prices, these items continue to have low weights because of their historically lower prices. Consequently, a fixed-weight index tends to overstate growth in periods after the base F E D E R A L R E S E R V E B A N K O F S T. L O U I S 40 SEPTEMBER/OCTOBER 1997 CONSTRUCTION OF CHAIN-WEIGHTED INDEXES The purpose of the change to chain-weighting in the NIPA accounts is to improve the separation of price changes and quantity changes in the overall measurement of nominal (current dollar) magnitudes. Specifically, real and nominal measures are related by Real GDP = Nominal GDP , P where P is some measure of the aggregate price level. Previously, this separation was accomplished by using implicit price deflators to measure P. Implicit price deflators are examples of what is known as a Paasche price index, calculated as follows: N ∑p q it it Pt = P i =1 N ∑p , ib qit i =1 where p and q denote quantities and prices, respectively; and the subscripts t and b denote measures in the current period and in a base period, respectively. An alternative price measure is known as a Laspeyres price index, the formula for which is N ∑p q it ib Pt = L i =1 N ∑p . ib qib i =1 Each of these indexes suffers from substitution bias, giving an imperfect measure of the price level. The Laspeyeres index tends to overstate price changes, while the Paasche index tends to understate them. One way of adjusting for these errors is to take an average of the two. The geometric average of a Laspeyres and a Paasche index is known as a Fisher Index: Pt F = Pt P x Pt L . Although this measure is largely free of systematic substitution bias, it is still subject to the problem of base-year sensitivity. History is rewritten with changes in the base year, since each of the component price indexes uses a fixed base year. The new chain-weighted GDP accounts use a variant of the Fisher index, in which the base year is not fixed. Instead, the previous year is taken as the base. This makes the price index interpretable as a percentage rate of change from the previous year. The percentage changes can then be “chained” together and expressed either as an index number or in terms of some particular reference period. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 41 SEPTEMBER/OCTOBER 1997 prices, caused growth to be overstated. Computers and related equipment were not the only items contributing to this effect, but their role was significant: For instance, it has been estimated that computers accounted for about three-fifths of the 1.1 percentage point overstatement of GDP in the fourth quarter of 1994.5 The tendency of the old 1987-dollar GDP series to understate growth before 1987 and to overstate growth after 1987 can be seen in Figure 2, which shows GDP growth from 1960 to 1995. Measured according to the old fixed-baseyear measure, GDP growth was 3.1 percent before 1987, and 2.3 percent after. With the new chain-weighted measure, the growth rate measures 3.4 percent prior to 1987, and only 1.8 percent after 1987. Figure 2 Real GDP 1987 Dollars and Chain-Weighted 1992 Dollars Trillions of Dollars 8 1.8% 7 6 3.4% 5 4 2.3% 3 3.1% 2 1 1960 1967 1974 1987 Dollars 1981 1988 1995 Chain-Weighted 1992 dollars Figure 3 A Changing View of the 1974-75 Recession Rewriting Economic History Quarterly Real GNP Growth Percent Change (ar) 10 5 0 –5 –10 1972 1973 1972 Dollars 5 Landefeld (1995). 6 Jain (1996). 7 Prowse (1996). 8 Ulan (1994). 9 The BEA switched from GNP to GDP as the featured measure of output in the 1991 comprehensive revisions. Because figures for GDP are unavailable for previous base-year series, GNP is used in Figure 3. 1974 1982 Dollars 1975 1987 Dollars 1976 1977 Chain-Weighted 1992 dollars year—and understate growth in periods before the base year. The move to a “Fisher Ideal” index, implicit in the new chained figures, largely solves this problem. (See shaded box p. 41: "Construction of •••: Chain-Weighted Indexes.”) The problem of substitution bias became prominent in the context of computer equipment. Compared to the 1987 base year, prices of computers and peripherals fell dramatically in the early 1990s, while sales of such equipment rose sharply. As a result, the volume of computer sales, valued at high 1987 An additional problem of the old fixed-weight methodology was that rebasing the weights resulted in “rewriting macroeconomic history every 5 years.”6 For example, one columnist has observed that “each time [the base year] shifts, the 1973-75 recession looks a little different.”7 In general, the effect of the old rebasing has been characterized as providing “a very distorted view of economic history.”8 Although the new measure continues to be expressed relative to a specific base year—now 1992—the chain-weighting approach makes the choice of a base year nothing more than a matter of normalization. Price and quantity changes are calculated as percentage changes from one year to the next and then “chained” together in a sequence. As a result, future “Construction of base year will no longer updates of the result in a revised view of the past, because the growth rates on which the chain-weighted figures are based will not be subject to subsequent revision. Figure 3 illustrates the rewriting of history for one episode—the 1973-75 recession and subsequent recovery. It shows a comparison between the growth rates of real GNP reported at the time of F E D E R A L R E S E R V E B A N K O F S T. L O U I S 42 SEPTEMBER/OCTOBER 1997 the 1973-75 recession and the revised figures after each of three revisions to the base year.9 It is clear that each time the NIPA accounts were rebased, the pattern of output growth was revised. However, the most significant change took place in the initial comprehensive revision, which rebased from the original 1972 dollars to 1982 dollars. The subsequent rebasing to 1987 dollars had little additional effect on the pattern of real GNP growth. Perhaps surprisingly, even the chainweighted measure of real GNP growth does not differ markedly from the 1982 or 1987 fixed-weighted measures, in spite of the fact that the chained figures use prices and quantities from the early 1970s, much as the original 1972-based figures did. Although rebasing changes the pattern of real GDP growth, another important factor in the periodic revision of the NIPA is the incorporation of improved source data. New source data is typically integrated by the time of the first rebasing (or in the annual benchmark revisions), implying noticeable changes to the series. Relative to the original 1972-based data shown in Figure 3, the first base year shows the most significant alteration in the pattern of measured economic activity. Subsequent updates show modest changes. Hence, Figure 3 suggests that revisions to the base year are less important than to the incorporation of new source data in initial benchmark revisions. Figure 3 also shows that although the pattern of GNP growth changed from one revision to another, the overall pattern did not change markedly. In particular, the timing of the onset of recession in late 1973 and the recovery in early 1975 was not affected by revisions. History was re-written with each revision, but the fundamental features of that episode in the business cycle remained essentially unchanged. This observation leads naturally to a basic question: Do data revisions—even those as fundamental as the recent switch to chain-weighted GDP—alter our overall perspective on business cycles? Figure 4 Real GDP and H-P Trend Chain-Weighted 1992 Dollars Trillions of Dollars 7 6 5 4 3 2 1 1960 1965 1970 1975 1980 H-P Trend 1985 1990 1995 1990 1995 Real GDP H-P Deviations from Trend Percent 6 4 2 0 –2 –4 –6 1960 1965 1970 1975 1980 1985 NOTE: Shaded bars indicate recessions. CYCLICAL PATTERNS UNDER THE OLD AND NEW METHODS Each business cycle seems to have its own unique features, yet certain regularities characterize the general nature of economic fluctuations. For various measures of economic activity, key issues about cyclical behavior revolve around the degree of variability, the persistence of fluctuations, and co-movement with other economic indicators. Some economic indicators are more variable than total output, some are less variable; some tend to move in the same direction as output (i.e., they are procyclical) while others tend to move in the opposite direction (countercyclical). These empirical F E D E R A L R E S E R V E B A N K O F S T. L O U I S 43 10 Formally, the trend component isolated by the H-P filter is given by the equation: ∑ {( y T min t t =0 [( − ytτ ) + λ ytτ+1 − ytτ ( ) )] 2 − ytτ − ytτ−1 , where ytτ is the trend component of the time series yt , and λ is a smoothing parameter. For applications with quarterly economic data, a value of λ=1600 is typically used. SEPTEMBER/OCTOBER 1997 Figure 5 Real GDP Fluctuations 1987 Dollars and Chain-Weighted 1992 Dollars Percent 4 2 0 -2 -4 1960 1965 1970 1975 1987 Dollars 11 This approach of comparing various “second moments” of time series has been popularized in the “real business cycle” literature; e.g. Kydland and Prescott (1982); King, Plosser, and Rebelo (1988). As in Figures 5 and 6, the data are logged and detrended by means of the H-P filter. Two series—investment and net exports—are not logged, but taken as a ratio to GDP in order to convert to proportionate measures before detrending. 12 The countercyclicality of prices is contrary to conventional wisdom and has been an issue of some controversy. See Kydland and Prescott (1992). 1980 1985 1990 1995 Chain-Weighted 1992 dollars regularities of business cycles are fundamental to economists’ attempts to understand and explain them. One way of approaching an answer to the question of how revisions affect our understanding of business cycles is to examine various measures and components of output after adjusting for long-term growth trends. This section uses a method of isolating the cyclical components, known as the Hodrick-Prescott (H-P) filter, to examine the cyclical behavior of economic variables. The H-P filter represents a statistical method of fitting a smooth trend line to an economic time series. The deviations from the trend line are then interpreted as the cyclical component of the series.10 Figure 4 (p. 43) illustrates a way that the H-P filter separates the trend components from the cyclical components of GDP. Shaded sections in the lower panel of Figure 4 indicate periods of recession as identified by the National Bureau of Economic Research. Note that the periods identified as economic downturns by the H-P filtering technique correspond closely to the official recession episodes. Figure 5 compares the H-P cyclical component of fixed-weighted and chainweighted GDP. It is clear from the figure that the broad pattern of cyclical fluctuations in real GDP is affected very little by the change from fixed-weighted 1987 dollars F E D E R A L R E S E R V E B A N K O F S T. L O U I S 44 to chain-weighted 1992 dollars, particularly for the 1970s and early 1980s. To the extent that differences do appear, they occur mostly in the late 1980s and 1990s (where recent benchmark and data-source revisions are important) and in the early 1960s (where fixed 1987 weights are likely to be particularly misleading due to the passage of time). Figure 6 provides a similar comparison for the major components of GDP. For all four components, the change from fixedweighted to chain-weighted measures makes little discernable difference. In all four cases, revisions to data for the late 1980s and 1990s are most prominent. Table 1 provides a more formal examination of the behavior of various components of GDP, presenting standard deviations, autocorrelations, and cross correlations with GDP.11 The first row of numbers for each variable shows the previously reported statistics based on the fixed-weight measures in 1987 dollars. The second number is the corresponding statistic for chain-weighted 1992 dollars. The overall patterns are clearly robust to the methodological changes. As measured by standard deviation, consumption of nondurables and services is only about half as variable as output, while durablegoods consumption and investment are about three times as variable. Each of these measures shows a strong positive correlation with output—i.e., each is procyclical. Government spending is about as variable as GDP, and largely uncorrelated to the business cycle. Net exports are countercyclical, due primarily to the strong procyclicality of imports. Under either measurement system, the price level is countercyclical.12 These overall patterns are clear for both the fixedweighted and chain-weighted data. Table 1 (pp. 46-47) also reports probability measures (in italics) for testing the hypothesis that there is any significant difference between the statistics for fixed-weight and chain-weighted figures. 13 A probability value near zero would indicate that the chain-weighted statistic was significantly less than the fixed-weight statistic, while a value near one SEPTEMBER/OCTOBER 1997 Figure 6 Fluctuations in Components of Real GDP Consumption Expenditures Percent Fixed Investment Percent 3 15 2 10 1 5 0 0 –1 –5 –2 –0 –3 1960 65 70 75 80 85 90 95 Government Spending Percent –5 1960 65 70 75 80 85 90 95 65 70 75 80 85 90 95 Net Exports Percent 6 2 4 1 2 0 0 –2 –4 –1 –6 –8 1960 –2 65 70 75 80 85 90 95 1987 Dollars would indicate the opposite. There are no cases at all for which the probability values indicate a statistically significant difference between the measures at conventional significance levels (< 0.05 or > 0.95). Table 2 (p. 48) provides information on the relationship of the two measures of GDP and other non-NIPA economic indicators. As in Table 1, the differences that arise as a result of the switch to chain-weighting are minor. Measures of employment remain strongly procyclical, while wages and prices are countercyclical. With the exception of non-borrowed reserves, measures of money (including total reserves, the monetary base, M1 and M2) are procyclical. None of the comparisons between the fixed-weight method and chain-weighting in Table 2 is statistically significant. 1960 1992 Chain-Weighted Dollars index. Most importantly, the chain-weighting approach corrects for the “substitution bias” in which measures of long-term growth become increasingly distorted as they move away from a fixed base year. It also resolves the issue of “rewriting history” every time there is a comprehensive revision to the national accounts. However, the switch to the new approach has little or no effect on an overall evaluation of the fluctuations and comovements among economic variables that constitute the business cycle. 13 The probability values for differences in standard deviations are derived from standard F-tests. For autocorrelations and crosscorrelations with GDP, the reported statistic is based on Z= N −3 2 2 1 + ρi x ln 1 − ρi 1+ ρ j − ln , 1 − ρ j which is normally distributed under the hypothesis that the correlations, ρ i and ρ j , are identical. CONCLUSION The new chain-weighted GDP measure is conceptually superior to the old fixed-weight F E D E R A L R E S E R V E B A N K O F S T. L O U I S 45 SEPTEMBER/OCTOBER 1997 Table 1 Comparisons of the Cyclical Properties of NIPA Data 1987 Dollars and Chained 1992 Dollars — H-P Filtered Variable GDP Standard Deviations Percent Ratio to GDP 1.63 1.00 1.71 1.00 0.611 0.85 Autocorrelations 2 3 0.65 4 0.43 0.23 0.84 0.62 0.39 0.19 0.614 0.661 0.655 0.635 Consumption Nondurables plus services 0.84 0.52 0.86 0.67 0.50 0.28 0.87 0.51 0.86 0.68 0.50 0.29 0.500 0.440 0.500 0.464 Consumption Services 0.68 0.42 0.80 0.62 0.46 0.29 0.71 0.41 0.582 0.600 Consumption Nondurables 0.82 0.65 0.50 0.32 0.315 0.339 0.333 0.392 1.20 0.74 0.84 0.63 0.43 0.20 1.24 0.72 0.84 0.62 0.41 0.18 0.500 0.554 0.580 0.568 0.76 0.60 0.40 0.23 0.576 Consumption Durables 5.01 3.08 5.01 2.93 0.500 0.77 0.60 0.41 0.24 0.421 0.500 0.461 0.465 Fixed nonresidential investment 5.22 3.21 0.89 0.71 0.48 0.23 5.00 2.93 0.89 0.70 0.46 0.21 0.500 0.565 0.584 Change in business inventories 0.52 0.32 0.43 0.17 0.01 – 0.15 0.47 0.27 – 0.14 Government purchases 1.65 1.63 0.400 0.46 0.18 0.02 0.466 0.467 0.466 1.02 0.86 0.72 0.58 0.43 0.95 0.85 0.70 0.58 0.43 0.621 0.631 0.500 0.500 0.64 0.52 0.38 0.22 4.37 2.69 4.38 2.56 0.505 Imports 0.51 0.37 0.19 0.545 0.538 0.602 3.16 0.75 0.54 0.35 0.18 5.29 3.09 0.74 0.54 0.37 0.21 0.574 0.500 0.425 0.398 0.89 0.76 0.64 0.49 0.50 0.31 0.44 0.26 0.226 Price deflator 0.62 0.608 5.14 0.567 Net exports 0.569 0.378 0.276 0.471 Exports NOTE: The first row of numbers for each variable refers to 1987 dollars, the second to chainweighted 1993 dollars, and the third (in italics) is a normal test statistic, extreme values of which represent rejection of the hypothesis that the two measures are equal (see text). 1 0.87 0.73 0.62 0.46 0.769 0.712 0.608 0.626 0.89 0.55 0.91 0.79 0.64 0.46 0.89 0.52 0.93 0.81 0.66 0.48 0.139 0.323 0.387 0.416 0.500 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 46 SEPTEMBER/OCTOBER 1997 Table 1 continued Comparisons of The Cyclical Properties of NIPA Data 1987 Dollars and Chained 1992 Dollars — H-P Filtered Cross-Correlations with Y(t+ j) 12 GDP 0 –1 –2 –4 –8 – 12 0.23 0.65 0.85 1.00 0.85 0.65 0.23 – 0.33 – 0.47 – 0.32 0.19 0.62 0.84 1.00 0.84 0.62 0.19 – 0.32 – 0.42 0.635 0.661 0.614 NA 0.614 0.661 0.635 – 0.41 – 0.32 0.17 0.56 0.73 0.86 0.83 0.71 0.37 – 0.24 – 0.42 – 0.40 – 0.32 0.12 0.52 0.71 0.84 0.81 0.69 0.37 – 0.20 – 0.39 0.664 0.680 0.635 0.725 0.693 0.627 0.500 0.463 0.500 . 0.463 0.364 0.303 0.383 – 0.46 – 0.24 0.21 0.51 0.65 0.80 0.79 0.69 0.36 – 0.24 – 0.45 – 0.46 – 0.24 0.18 0.49 0.64 0.76 0.77 0.67 0.38 – 0.19 – 0.41 0.602 0.587 0.556 0.802 0.664 0.621 0.424 – 0.34 – 0.35 0.12 0.53 0.71 0.82 0.77 0.65 0.34 – 0.20 – 0.35 – 0.30 – 0.34 0.06 0.49 0.68 0.80 0.75 0.63 0.32 – 0.19 – 0.33 0.692 0.673 0.685 0.685 0.652 0.610 0.574 0.356 Consumption Durables 1 – 0.33 0.500 Consumption Nondurables 2 – 0.42 0.461 Consumption Services 4 – 0.47 0.303 Consumption Nondurables plus services 8 0.500 0.463 0.332 0.466 0.342 0.426 – 0.43 – 0.52 – 0.08 0.38 0.59 0.80 0.77 0.68 0.47 – 0.05 – 0.29 – 0.40 – 0.50 – 0.08 0.39 0.61 0.81 0.78 0.67 0.44 – 0.06 – 0.29 0.500 0.461 0.398 0.407 0.418 0.560 Fixed nonresidential investment – 0.53 0.382 – 0.28 0.49 0.79 0.85 0.80 0.60 0.36 – 0.05 – 0.40 – 0.33 – 0.50 – 0.28 0.45 0.78 0.85 0.81 0.60 0.35 – 0.07 – 0.40 – 0.32 0.665 0.585 0.500 0.407 0.500 0.538 0.566 0.500 Change in business inventories – 0.17 – 0.41 – 0.15 0.20 0.42 0.61 0.44 0.33 0.12 0.00 – 0.14 – 0.13 – 0.40 – 0.16 0.18 0.42 0.65 0.50 0.37 0.12 0.00 – 0.12 Government purchases – 0.02 0.368 0.367 – 0.08 Net exports Price deflator 0.461 0.534 0.568 0.500 0.291 0.40 0.34 0.18 0.13 0.08 0.262 0.353 0.500 0.533 0.500 0.500 0.463 0.433 – 0.01 – 0.07 – 0.14 – 0.18 – 0.12 – 0.03 – 0.12 – 0.22 – 0.16 0.35 0.37 0.26 0.22 0.16 0.05 0.388 0.243 0.221 0.251 0.310 – 0.03 0.29 0.47 0.45 0.41 0.27 0.05 – 0.18 – 0.43 – 0.49 – 0.10 – 0.09 0.23 0.42 0.45 0.42 0.29 0.10 – 0.12 – 0.36 – 0.45 – 0.10 0.691 Imports 0.500 0.623 0.685 0.691 Exports 0.411 0.703 0.370 0.433 0.635 0.335 0.632 0.697 0.500 0.460 0.429 0.339 0.306 0.246 – 0.42 – 0.52 0.06 0.48 0.70 0.75 0.65 0.50 0.28 – 0.09 – 0.30 0.500 – 0.38 – 0.48 – 0.06 – 0.30 0.03 0.45 0.68 0.75 0.66 0.51 0.29 0.347 0.329 0.598 0.624 0.624 0.500 0.442 0.456 0.464 0.39 0.64 0.24 – 0.12 – 0.31 – 0.44 – 0.51 – 0.53 – 0.52 – 0.29 0.14 0.33 0.57 0.21 – 0.12 – 0.31 – 0.43 – 0.49 – 0.50 – 0.48 – 0.26 0.15 0.716 0.820 0.51 0.49 – 0.04 – 0.34 – 0.46 – 0.58 – 0.67 – 0.72 – 0.64 – 0.10 0.35 0.48 0.49 – 0.07 – 0.40 – 0.55 – 0.66 – 0.70 – 0.70 – 0.56 – 0.06 0.34 0.629 0.500 0.603 0.598 0.500 0.718 0.500 0.459 0.842 F E D E R A L R E S E R V E B A N K O F S T. L O U I S 47 0.860 0.413 0.680 0.368 0.369 0.329 0.150 0.401 0.394 0.369 0.500 0.466 0.537 SEPTEMBER/OCTOBER 1997 Table 2 Cross-Correlations of Non-NIPA Variables with GDP Variable 12 Employment – 0.45 – 0.17 – 0.43 – 0.18 0.419 Average weekly hours 8 –2 –4 0.55 0.83 0.90 0.84 0.64 0.42 0.03 – 0.44 – 0.4 0.51 0.81 0.88 0.82 0.63 0.41 0.02 – 0.41 – 0.44 0.693 0.788 0.703 0.555 0.540 0.533 – 0.09 0.34 0.60 0.77 0.71 0.55 0.23 – 0.25 – 0.23 – 0.24 – 0.48 – 0.12 – 0.24 – 0.20 0.32 0.58 0.77 0.72 0.57 0.23 0.599 0.574 0.600 0.500 0.433 0.405 0.500 0.381 – 12 – 0.48 0.500 0.677 –8 – 0.26 0.430 0.534 Cross-Correlations with Y(t+ j) 2 1 0 –1 4 0.465 0.377 0.397 Total work effort – 0.44 – 0.3 0.41 0.76 0.89 0.91 0.75 0.52 0.09 – 0.45 – 0.43 (Employment x Hours) – 0.42 – 0.31 0.38 0.74 0.87 0.9 0.75 0.53 0.09 – 0.43 – 0.41 Average hourly earnings CPI Nonborrowed reserves Total reserves 0.420 0.536 0.29 0.16 – 0.04 – 0.17 – 0.23 – 0.28 – 0.39 – 0.47 – 0.49 – 0.26 0.500 0.455 0.500 0.419 – 0.05 – 0.16 – 0.21 – 0.25 – 0.35 – 0.43 – 0.46 – 0.25 0.420 0.10 0.12 0.533 0.46 0.54 0.11 – 0.25 – 0.42 – 0.59 – 0.71 – 0.77 – 0.70 – 0.18 0.33 0.44 0.50 0.11 – 0.23 – 0.39 – 0.55 – 0.68 – 0.73 – 0.67 – 0.17 0.30 0.582 0.675 0.09 0.19 – 0.21 – 0.29 – 0.27 – 0.18 – 0.01 0.08 0.21 – 0.17 – 0.29 – 0.29 – 0.20 – 0.05 0.533 0.432 0.500 0.366 0.466 0.430 0.431 0.383 0.395 0.312 0.351 0.315 0.339 0.374 0.465 0.11 0.467 0.224 0.320 0.466 0.609 0.15 0.27 0.27 0.02 0.10 0.22 0.24 0.04 0.500 0.571 0.568 0.630 0.663 0.670 0.605 0.434 – 0.07 0.04 – 0.03 0.03 0.09 0.17 0.24 0.27 0.23 0.04 – 0.07 – 0.05 0.07 – 0.01 0.03 0.07 0.13 0.19 0.21 0.17 0.01 – 0.03 0.402 0.434 0.500 0.566 0.633 0.668 0.701 0.698 – 0.19 0.04 0.11 0.25 0.30 0.36 0.32 0.26 0.16 – 0.09 – 0.20 – 0.18 0.08 0.12 0.23 0.26 0.31 0.26 0.19 0.14 – 0.08 – 0.20 0.370 0.467 0.570 0.640 0.679 0.706 0.729 0.567 0.467 0.598 0.370 0.500 – 0.09 – 0.06 – 0.04 0.09 0.19 0.30 0.36 0.36 0.28 0.05 – 0.19 – 0.06 – 0.01 – 0.03 0.08 0.16 0.25 0.30 0.30 0.21 0.03 – 0.15 0.533 0.601 0.673 0.711 0.711 0.731 0.566 – 0.14 – 0.40 – 0.26 0.03 0.20 0.39 0.52 0.57 0.48 0.06 – 0.26 – 0.13 – 0.39 0.339 0.467 0.367 – 0.26 – 0.26 0.04 0.22 0.40 0.53 0.57 0.48 0.05 0.461 0.500 0.467 0.431 0.461 0.455 0.500 0.500 0.533 – 0.04 0.22 0.48 0.48 0.45 0.34 0.08 – 0.21 – 0.55 – 0.58 – 0.09 – 0.05 0.18 0.43 0.46 0.43 0.34 0.10 – 0.17 – 0.52 – 0.55 – 0.08 0.635 0.699 0.584 0.581 0.500 0.434 0.466 3-month T-bill rate 0.676 0.632 0.401 M2 0.769 0.27 0.466 M1 0.647 0.571 0.434 Monetary base 0.616 0.533 0.366 NOTE: See note to Table 1. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 48 0.364 0.358 0.500 0.467 SEPTEMBER/OCTOBER 1997 REFERENCES Jain, P. “Feature: The Chain-weighted GDP Measure,” Prudential Economics (January 1996), pp. 14-17. King, Robert G., Charles I. Plosser, and Sergio T. Rebelo. “Production, Growth and Business Cycles: I. The Basic Neoclassical Model,” Journal of Monetary Economics (1988), pp. 195-232. Kydland, Finn E., and Edward C. Prescott. “Business Cycles: Real Facts and a Monetary Myth,” Quarterly Review, Federal Reserve Bank of Minneapolis (Spring 1990), pp. 3-18. _____ and _____. “Time to Build and Aggregate Fluctuations,” Econometrica (1982), pp. 1345-70. Landefeld, J. Steven. “BEA’s Featured Measure of Output and Prices,” NABE News 113 (September 1995). Parker, Robert P., and Eugene P. Seskin. “Improved Estimates of the National Income and Product Accounts for 1959-95: Results of the Comprehensive Revision,” Survey of Current Business, (January/February 1996), pp. 1-31. Passell, Peter. “Maybe It Wasn’t the Economy in the 1992 Election,” New York Times, Jan. 20, 1996. Prowse, Michael. “Lies, Damned Lies, and the US Commerce Department’s New Way of Measuring GDP,” Financial Times, Jan. 20-21, 1996. Samuelson, Robert J. “Rewriting Economic History,” The Washington Post, Feb. 28, 1996. Ulan, Michael. “Is the Current Business Cycle Different? Does How We Measure Matter?” Business Economics (April 1994), pp. 41-47. United States Department of Commerce–Bureau of Economic Analysis, “Gross Domestic Product: Third Quarter 1995 (preliminary); Corporate Profits, Third Quarter 1995 (preliminary); and Revised Estimates, 1959-95,” Survey of Current Business (November/December 1995), pp. 1-47. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 49 SEPTEMBER/OCTOBER 1997 Katrin Wesche is an assistant professor of economics at the Institut für Internationale Wirtschaftspolitik, Universität Bonn in Bonn, Germany. Cindy Gleit provided research assistance. The author is grateful to Dan Thornton for his invaluable help, and to Barry Jones and Travis Nesmith, whose comments substantially improved this paper. The Demand for Divisia Money in a Core Monetary Union constructed for each country, and the European aggregate is taken to be the average of the national indexes. In the direct method, the components are added across countries, and weighted averages of national interest rates are used to obtain the user cost for each component. Neither approach is strictly consistent with aggregation theory, because aggregation by averaging national Divisia implicitly assumes perfect substitutability across indexes, and the summation of monetary assets across countries requires that assets, denominated in different currencies, be perfect substitutes. Aggregation over different national moneys should employ appropriate methods.3 European monetary aggregation that uses indexes for monetary services is particularly attractive because such indexes can account for the different paces of financial innovation in the countries of Europe. The main contribution of this paper is to apply the aggregation theoretic framework consistently to money holdings of European residents.4 The first section presents the definition of the Divisia index. The second derives a European Divisia index. In the third, the Divisia index and a simple-sum measure of European money are compared and analyzed. Katrin Wesche T he ratification of the Maastricht Treaty and the agreement on the constitution of the European Central Bank have given rise to a number of papers investigating the demand for money in Europe. In most of this work, conventional simplesum aggregates have been used to measure the quantity of money in the European Union.1 However, proponents of the aggregation theoretic approach to the demand for money argue that simple-sum measures lack adequate theoretical foundations and fail to capture the theoretical notion of money. This is especially true for broad monetary aggregates, which include components that are imperfect substitutes for transactions media. The use of simple-sum aggregates in the investigation of European money demand, therefore, is questionable—especially since the European Central Bank will presumably target a broad monetary aggregate. Some research on European money demand has considered aggregation theory. For example, Fase and Winder (1994) and Fase (1996) compute European Divisia monetary indexes2 for different groups of countries in the European Union and find that European money demand is fairly stable. A similar result is obtained by Monticelli and Papi (1996), who construct a currency-equivalent index proposed by Rotemberg, Driscoll, and Poterba (1995). Both studies construct indexes by using the direct and the indirect methods. In the indirect method, an index is THE DIVISIA MONETARY INDEX Most writers define money according to the functions it performs.5 Monetary assets serve as a medium of transaction, a store of value, and a unit of account, with the medium-of-transaction function being crucial for distinguishing monetary assets from other financial assets. It has, however, become commonplace for monetary aggregates to include financial assets that are not mutually exchangeable.6 For example, savings and time deposits are included in the M2 monetary aggregate, despite the fact that they cannot be used to make transactions. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 51 1 See, e.g., Kremers and Lane (1990), Monticelli and StraussKahn (1991), Artis et al. (1993). 2 The term Divisia index is used throughout the paper to refer to the Törnqvist-Theil discrete time approximation to the continuous time index suggested by Divisia (1926). 3 Marquez (1987) tackles this problem by applying the aggregation approach to money demand in an open economy. But since his focus is on the holdings decisions of residents, only residents’ holdings of foreign currency are included. The same is the case in the study by Ewis and Fisher (1984), who find strong substitutability between domestic and foreign monetary assets with a translog utility model. 4 This paper focuses on the consumer’s problem. For models applicable to firms and financial intermediaries, see Barnett (1987). 5 See Osborne (1992) for a more detailed survey on different approaches to the definition of money. 6 In accordance with aggregation theory, a monetary aggregate is defined over a weakly separable block in the utility function. This definition is rarely implemented because tests for blockwise weak separability are biased towards rejection; a single rejection in the data renders the formation of a separable group impossible. For a discussion of separability tests and applications to U.S. monetary data, see Swofford and Whitney (1986, 1987, 1988, 1994). SEPTEMBER/OCTOBER 1997 7 Barry Jones and Travis Nesmith point out that the aggregation approach provides a “negative” definition of monetary services by specifying that the store-ofvalue function is not a monetary service. Nevertheless, as the Divisia index intends to measure a monetary services flow (and given the difficulties with testing for weak separability), some a priori idea of which assets contain a monetary services component is useful for determining which assets to include in a monetary aggregate (see Barnett, 1982, p. 697). 8 Aggregation of real monetary assets is equivalent to aggregating nominal assets and deflating the monetary services index afterwards (see Anderson, Jones, and Nesmith, 1997b). 9 Actually, the single-period utility function is a special case. The solution to the single-period utility function is equivalent to an intertemporal optimization if current-period monetary assets are weakly separable from the other decision variables in the consumer’s utility function. 10 Money is included in the utility function, since it provides services such as convenience, liquidity, and information. For the equivalence of putting money into the utility function or solely into the budget constraint, see Barnett, Fisher, and Serletis (1992, p. 2093). Barnett (1987) shows that with money entering the production function as durable physical capital, the user cost formula also applies for firms. The user costs also have the same form for financial intermediaries, if no reserve requirements are present. The idea in the aggregation approach is to extract the store-of-value function from all financial assets, so that what remains are the “monetary services” for the assets.7 It is assumed that the store-ofvalue characteristic of an asset is reflected by its investment yield and that one asset, called the benchmark asset, provides only the store-of-value function and no other. In addition, instead of simply adding such assets together, as it is done in simple-sum aggregation, the theoretical approach of aggregation creates an index of monetary services that has microeconomic foundations. While conventional monetary aggregates are derived in a simple accounting procedure from the banking sector’s balance sheet, the theoretical approach, or Divisia index, is based on the optimizing behavior of economic agents. One way to see how this aggregation theory approach compares with simple-sum aggregation is to assume that individuals maximize a utility function composed of a number of real monetary assets, Mi /p*, 8 and commodities that are directly consumed, Cj.9 That is, consumers maximize the optimal quantities of consumption goods and optimal total expenditures on monetary assets. In the second stage, the monetary expenditures are allocated among specific monetary assets. The solution to the maximization problem that uses the two-step approach is identical to the one that uses a one-step approach so long as the marginal rate of substitution between any two monetary assets does not depend on the quantities of commodities consumed (Barnett, Fisher, and Serletis, 1992). This condition, referred to as blockwise weak separability, is necessary for economic aggregation.12 If this condition is satisfied, the monetary aggregate behaves like a single economic good for which a demand function exits.13 Under these assumptions, M is the monetary aggregate that we desire to measure. The discrete-time approximation to the continuous-time Divisia index is exact for a function that can provide a second-order approximation to any arbitrary aggregator function, M, and therefore belongs to the class of superlative indexes, as defined by Diewert (1976).14 The growth rate of the Divisia index is defined as M M M U = U 1 , 2 ,…, I , C1, C2 ,…, CJ , p* p* p* log Qt - log Qt - 1 N subject to a budget constraint, where p* is a true cost-of-living index.10 The I monetary assets commonly include assets that are used directly in transactions—i.e., cash and checkable deposits—but may include other financial assets such as saving and time deposits as well. The aggregation approach assumes that there exists an aggregator function, = ∑ s (log it i =1 sit = M Mit – log i*,t - 1 , * pt pt - 1 ( 1 (sit + si ,t - 1 ). 2 with the expenditure shares sit = M M M 11 M = f 1 , 2 , …, I . p* p* p * π it Mit K ∑ π kt Mkt . k =1 In the aggregation approach, money is regarded as a durable good that yields services in facilitating transactions and providing liquidity. The user cost, pit , for monetary services therefore can be derived in a fashion analogous to that used to derive the user cost for a durable consumption good (see Donovan, 1978; Barnett, 1978, 1987).15 For a durable consumption good, the one-period holding cost, or rental price, The utility function can be rewritten as U = F (M , C1, C2 ,…, CJ ), so that the demand for money can be separated from the demand for consumption goods. Consumers can be seen as allocating their budget in two stages (Green, 1964). In the first stage, the consumer chooses F E D E R A L R E S E R V E B A N K O F S T. L O U I S 52 SEPTEMBER/OCTOBER 1997 is given by the cost of the purchase of the good in the current period less the discounted expected resale value of the depreciated good in the next period.16 The opportunity cost of a component monetary asset is measured by the user cost, pit , of the ith monetary asset, defined by π it = pt* financial assets across countries. Namely, the value of component assets changes as exchange rates vary. Hence, the aggregation approach must be modified to account for expected changes in the exchange rate. The stock of monetary assets is redefined to account for currencies of different denominations. That is, the representative consumer is assumed to hold real monetary assets, denominated in different European currencies, Rt - rit , 1 + Rt which is a function of the difference between the own rate of return on the ith asset, ri , and the return on a so-called benchmark asset, R. The benchmark asset is assumed to provide no monetary services and is used only to transfer wealth between periods. The user cost is larger, the smaller is the own rate of return. The own rate of return on cash is taken to be zero and therefore cash has the highest user cost. On the other hand, a monetary asset earning the benchmark’s rate of return would not contribute to the growth of the index in that period.17 The aggregation approach does not consider aggregation over a diverse population of individuals. To deal with the problem, it uses the concept of a representative consumer. In essence, the behavior of the representative consumer is assumed to reflect the average behavior of the population. Researchers frequently employ the representative agent methodology to avoid the problems that can arise from aggregation over a diverse group of individuals (see Phlips, 1974, p. 100). The assumption of a representative consumer is very restrictive, but it is assumed in simple-sum aggregation as well. Mik,t / ek,t pt* , where Mik is the ith monetary asset denominated in the kth country’s currency and e k is the kth country’s exchange rate relative to a weighted currency basket like the ECU (see Wesche, 1996 for details). As it is assumed that the representative consumer allocates his consumption expenditure on European consumption goods, the true cost-of-living price index, p*, is defined in terms of this bundle of European consumption goods. In addition, the own rate of return, rik , of a component monetary asset has to take account of the expected depreciation or appreciation of the respective currency relative to the weighted exchange rate. The nominal user cost for the European Divisia index thus becomes R − r +δe π ik,t = p t ik,t k,t , 1 + Rt * t with EUROPEAN MONETARY AGGREGATION δ ek,t = eke,t +1− ek,t ek,t +1 being the expected depreciation of the kth country’s currency and Rt = max(Rk,t – dke,te) the European benchmark yield, which is the highest yield on a portfolio of European bonds, corrected for expected depreciation of the exchange rate. The main difference between the user cost in the multiplecountry framework and the single-country case is that the user cost reflects the expected capital gain (or loss) on money To derive a European monetary aggregate, researchers assume that consumers hold a diversified portfolio of European currencies with different degrees of liquidity.18 Nevertheless, in contrast to the computation of a Divisia index for a single country, an additional difficulty arises when the Divisia index is applied to F E D E R A L R E S E R V E B A N K O F S T. L O U I S 53 11 Technically, M is the function evaluated at its optimal point. 12 The precise conditions under which aggregation is valid are stated in Anderson, Jones, and Nesmith (1997a). 13 The condition of blockwise weak separability can be tested by examining the consistency of consumer choices. A violation of consistency would occur if consumers chose a different market basket even if prices remained unchanged. 14 In fact, the continuous time Divisia index is always exact. In contrast, the simple-sum is the exact index if, and only if, all of the component assets are perfect substitutes. Moreover, the simple-sum has no statistical properties in any case. 15 The user-cost formula is derived from a dynamic budget constraint. User costs are correct even if aggregation is not valid. 16 Even if the good is held more than one period, it can be assumed that the holder sells the good to himself at the end of each period (see Donovan, 1978). 17 This implies that the asset does not provide any marginal utility in that period (Barnett, 1996). 18 This says nothing about the substitutability of different national moneys. Indeed, failure of the representative consumer to react in his portfolio composition to exchange-rate changes indicates that different currencies are not close substitutes. SEPTEMBER/OCTOBER 1997 19 Unless otherwise indicated, all data are from the International Monetary Fund’s International Financial Statistics. 20 The weighted exchange rate uses GDP weights, converted with purchasing power parities from the OECD (1990). holdings that results from exchange-rate fluctuations. A capital gain caused by an appreciation of the exchange rate is treated like the interest yield of a monetary asset. Though national currencies have different user costs, consumers hold all of them because they are imperfect substitutes. If they were perfect substitutes, the representative consumer would hold only the currency with the lowest user cost. the benchmark rate, although even longterm bonds are not completely illiquid. To construct a European benchmark rate, we assume that bonds denominated in different currencies perform the same function — i.e., the transfer of wealth between periods. So the benchmark rate becomes the highest national interest rate, corrected for expected depreciation.23 In theory, the benchmark yield is the maximum expected holding-period yield in the economy (Barnett, Fisher, and Serletis, 1992).24 Any asset that yields monetary or liquidity services must earn less interest than the benchmark asset. In reality, however, interest rates on time deposits are often higher than long-term rates. This would cause the user costs to become negative if the long-term rate is taken to be the benchmark rate. To avoid negative user costs, which make no sense, two types of adjustments have been used. In the first, the user cost is augmented by its minimum value. This approach can be interpreted as a “liquidity mark-up,” since data on the theoretically correct benchmark yield are difficult to identify. This method is arbitrary, however, as the particular minimum value depends on the sample period. In the second approach, the asset yielding the highest return in the period is taken to be the benchmark asset. Fisher, Hudson, and Pradhan (1993) argue that in principle the benchmark asset should not provide monetary services and, therefore, an asset that is included as money in a previous time period should not be used later as the benchmark. Thus in some periods an asset will have a zero user cost and a resulting zero contribution to the growth rate of the monetary index. Only results for the index obtained with the second method are presented here.25 It is generally assumed that M1 earns no interest. For the interest rate on quasi money, I use the money market rate.26 Figures 1a to 1c show the user cost for narrow money and quasi money for each of the three countries. The user costs for M1 are very similar for all countries after 1987 because of the convergence of nom- 21 Here I construct the Divisia index in a single stage. One could also (with the appropriate separability assumptions) use a two-stage aggregation approach in which the consumer first allocates his expenditures among monetary assets in different national currencies, and then among monetary assets denominated in the same currency with different degrees of liquidity. Though this approach has more restrictive assumptions, it would be advantageous if higher-quality national Divisia indexes could be used. 22 Following the literature, aggregation is performed over the components of the official aggregates and implicitly assumes that weak separability is satisfied (see also Thornton and Yue, 1992; Fisher, Hudson, and Pradhan, 1993; or Gaab and Mullineux, 1996). 23 Expected depreciation is proxied by actual depreciation, assuming zero uncertainty and agents with rational expectations. These assumptions are restrictive and do not hold in practice. However, because expected depreciation enters in the numerator and denominator of the user cost, errors may cancel out partially. To capture uncertainty, the user cost could be adjusted by an additional term reflecting the interest-rate and exchange-rate risk (see Barnett and Liu, 1995). This aspect, however, is neglected here. Construction of the Index The countries investigated are Germany, France, and the Netherlands, the most likely candidates for a core monetary union. A currency union without Germany and France is inconceivable, since these two countries are the driving forces behind European unification. The Netherlands, being the only country for which the narrow exchange rate targets currently apply, has close economic relations with Germany as well as with France. Data are quarterly from 1973:1 to 1994:4.19 The simple-sum European money stock is converted with current exchange rates and expressed in a weighted currency.20 As in Fase and Winder (1994) and Monticelli and Papi (1996), aggregation is performed over two different groups of monetary assets: narrow money (M1) and quasi money (M3-M1) as defined in the International Financial Statistics.21,22 The income variable, gross domestic product (GDP), is also converted into a weighted currency. The European price index, used to deflate the simple-sum and the Divisia aggregates, is obtained through aggregation of national consumer price indexes with GDP weights, based on current exchange rates. Identifying the benchmark asset is difficult. Conceptually, the benchmark asset offers no transactions services and can be used only to transfer wealth between periods. Moreover, in order to be comparable to monetary assets, the benchmark asset should be capital-certain, and its yield should not include a risk premium (see Fisher, Hudson, and Pradhan, 1993). The yield on government bonds is taken as F E D E R A L R E S E R V E B A N K O F S T. L O U I S 54 SEPTEMBER/OCTOBER 1997 Figures 1a-1c Figures 2a-2c User Cost for Narrow Money and Quasi Money (Percent) Annual Growth Rate of Narrow Money and Quasi Money (Percent) Germany Germany 25 25 20 20 15 15 10 10 5 5 0 –5 0 74 76 78 80 82 84 86 88 90 92 74 94 France France 25 20 20 15 76 78 80 82 84 86 88 90 92 94 10 15 5 10 0 5 24 To –5 0 –10 74 76 78 80 82 84 86 88 90 92 74 94 76 Netherlands Netherlands 20 25 78 80 82 84 86 88 90 92 94 78 80 82 84 86 88 90 92 94 20 15 15 10 10 5 5 0 0 74 76 78 80 82 84 86 88 90 92 –5 94 74 — UC Quasi Money — UC Narrow Money 76 — Quasi Money inal interest rates during the “hard” period of the European Monetary System following the Basle-Nyborg agreement in 1987. Even the widening of the exchange rate bands in 1993 had almost no effect on the user cost, since neither the French franc nor the Dutch guilder depreciated significantly against the German mark. The user cost of quasi money is surprisingly low for France because the short-term interest rate in France is relatively high—often higher than the government bond yield. Consequently, the French money market rate is frequently — Narrow Money the benchmark rate. After the establishment of the European Monetary System, the user costs for the three countries narrowed considerably, indicating progress in monetary and financial integration. Figures 2a to 2c show the growth of the monetary components in the three countries. The German Unification is denoted by the sharp spike in money growth rates in 1990, with M1 growing more rapidly than quasi money. Money growth in France declined steadily after the beginning of the ’80s. After German Unification, France had to follow a very F E D E R A L R E S E R V E B A N K O F S T. L O U I S 55 be comparable, interest rates should be holding-period adjusted; liquidity premia are generally higher on longer maturity assets. This can be done by estimating a yield curve adjustment (see e.g., Anderson, Jones, and Nesmith, 1997b, or Farr and Johnson, 1985). Unfortunately, in the International Financial Statistics no data on the yield curve for government bonds are available. The own rates on monetary assets, however, are comparable as they refer to the same holding period. 25 Which of these two adjustments for negative user costs is used makes no qualitative difference for the empirical results. To avoid taking logarithms of zero, a very small constant of less than a basis point was further added to the user costs (see Anderson, Jones, and Nesmith, 1997b). 26 A deposit rate would have been preferable but was not available for all countries over the sample period. SEPTEMBER/OCTOBER 1997 Table 1 Figures 3a, 3b Nominal Growth Rates of Monetary Aggregates Annual Growth Rates Divisia and M1 QM3 M3 M3-M1 M1 Sample 73:1-94:4 Mean Standard deviation 7.39 2.56 7.69 2.30 7.66 2.43 7.74 3.08 Sample 73:1-78:4 Mean Standard deviation 8.79 2.44 9.36 1.26 9.01 1.41 10.04 3.49 Sample 79:1-86:4 Mean Standard deviation 8.17 2.12 8.56 1.97 8.81 2.25 8.07 2.11 Sample 87:1-90:2 Mean Standard deviation 5.80 0.98 5.57 1.34 5.27 1.62 6.15 1.22 Sample 90:3-94:4 Mean Standard deviation Percent 20 M1 15 10 5 Divisia 0 –5 74 76 78 80 82 5.56 1.57 5.67 1.55 86 88 90 92 94 86 88 90 92 94 Divisia and M3 Percent 14 M3 12 5.38 2.50 84 10 5.34 2.63 8 6 NOTES: “QM3” denotes the Divisia aggregate, “M3” the simple-sum monetary aggregate, “M1” narrow money, and “M3–M1” quasi money for Germany, France, and the Netherlands. Annual growth rates in percent. 4 Divisia 2 0 74 27 These characteristics are also found by Gaab and Mullineux (1996), and by Issing et al. (1993) for Germany, and by Gaiotti (1994) for Italy. restrictive monetary policy to support its exchange rate. This is reflected in the sharp drop in M1 growth in 1990, and in the relatively slow money growth and quasi money growth thereafter. Money growth slowed over the sample period in the Netherlands, although no clear effect of German Unification is seen. This is not surprising, since the Netherlands did not experience an exchange-rate crisis. Substitutability between narrow money and quasi money appears to be high for all these countries, but particularly so for the Netherlands. This is especially true at the beginning of the sample period when the growth rates of narrow money and quasi money moved in opposite directions. The user costs for non-interest-bearing money are highest in France because, on average, France had higher inflation in the first part of the sample leading to exchange-rate depreciation against both of the other currencies. Figures 3a and 3b compare the annual growth rates of the European Divisia indexes and the traditional simple-sum aggregates. From 1982 on, the growth 76 78 80 82 84 rates of the Divisia index and M1 were very close. As the money market rate is used as own rate on quasi money, the user cost for quasi money is presumably too low because time and savings deposits in general earn an interest rate below the money market rate. Consequently, the share of quasi money in the index is biased downwards, and the Divisia aggregate behaves much like M1. Table 1 shows descriptive statistics for the whole sample period as well as for different subperiods. Like the national Divisia indexes for 10 European countries computed by Fase and Winder (1994), nominal Divisia money shows a lower average growth rate and a higher standard deviation than simple-sum M3.27 Differences between the growth rates of the Divisia index and the traditional aggregates are not significant for any sample period. From 1987 onwards, the growth rate of all aggregates fell considerably. Even after German Unification, money growth was lower than in every other subsample, despite the rise in the German F E D E R A L R E S E R V E B A N K O F S T. L O U I S 56 SEPTEMBER/OCTOBER 1997 Table 2 Unit Root Tests Variable QM3R M3R EXPR GDPR PQ GBY ADF Level ADF 1. Diff. –2.858 –3.075 –4.409 –3.306 –1.954 –2.257 –3.781 –3.303 –5.631 –3.318 –5.433 –3.619 Regression T T T T C C C C C C N N Conclusion unit root unit root trend stationary unit root unit root unit root NOTES: “QM3R“ is real Divisia money, “M3R” real simple-sum money. “EXPR” denotes real expenditure on consumption and monetary services, “GDPR” real gross domestic product, “PQ” the price dual to the Divisia index, and “GBY” the government bond yield. Except for the government bond yield, all variables are in logs. The sample period is 1973:1 to 1994:4. “ADF Level” and “ADF 1. Diff.” are the Augmented Dickey-Fuller test statistics for the levels and the first differences of the variables, respectively. The “Regression” column shows the specification of the test, with the first entry referring to the test of the levels, and the second entry the test of the first differences of the variables. “T” indicates the inclusion of a trend and a constant, “C” the inclusion of a constant only, and “N” a test without trend and constant. All tests include four lags. Critical values are –3.464 for the tests including a trend and a constant, –2.896 for the tests with a constant only, and –1.946 for the tests without trend and constant (MacKinnon, 1991). money stock. This rise was compensated by a very slow money growth in France. In general, all monetary aggregates show the same picture and are consistent with the lower inflation and the more stable exchange rates that have prevailed in Europe since the mid-’80s. Furthermore, expenditures on monetary services are not included in GDP but would be included in the representative household’s budget constraint. Similar considerations apply to the opportunity cost variable frequently used in money demand estimations. Modeling the demand for Divisia money in the conventional way is justifiable from a policymaker’s perspective. A measure of money is useful to the policymaker only insofar as it conveys information about the behavior of objective variables, such as prices and output (see Pill and Pradhan, 1994). Two different money demand equations for Divisia money are estimated here: The first one uses expenditures on consumption and monetary services as the income variable 29 and the Divisia price dual 30 as opportunity cost. The second uses GDP and an interest rate as regressors. These regressions are compared to a conventional simple-sum money demand function. Money Demand Analysts often assess the performance of a Divisia index by estimating a demand function for Divisia money and comparing it to the money demand function for a simple-sum aggregate.28 Money demand functions generally include real income and an interest rate as explanatory variables. However, Barnett (1996) argues that these variables are inconsistent with demand theory. The Divisia index is derived from a utility maximization framework; hence, the demand for Divisia money should be modeled according to demand theory as the first stage of the budget allocation, in which the agent allocates his expenditures among consumption goods and monetary services. However, national income does not correspond to the representative agent’s income as it appears in the budget constraint. For example, GDP contains components such as investment that do not appear in the budget constraint. 28 See, e.g., Gaab and Mullineux (1996), and Barnett (1982). 29 Data on private consumption expenditures are from the OECD National Accounts. Data were converted to 1990 prices for France (1980 prices) and Germany (1991 prices). For the Netherlands, consumption data from 1973:1 to 1976:4 were extrapolated with GDP data. As for Germany, pre-unification data are seasonally adjusted and post-unification data are not; these were adjusted by a regression on three seasonal dummies and a constant. 30 The EMPIRICAL RESULTS Before the model is specified, the time series are tested for their order of integration. Table 2 presents the results of the unit root tests. Most variables are integrated of order one; only real expenditures seem to be trend-stationary in levels. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 57 price dual is computed by dividing total expenditure on monetary assets by the Divisia quantity index. As the Divisia index is computed over real monetary assets, the corresponding price dual is nominal (see Anderson, Jones, and Nesmith, 1997b). SEPTEMBER/OCTOBER 1997 Divisia aggregate than the interest rate elasticity of simple-sum M3. Nevertheless, if the Divisia aggregate is regressed on GDP and the government bond yield, the results are almost identical to those obtained with M3. Stationarity of the residuals is tested with a Dickey-Fuller test. Residuals are stationary at the 5 percent level for both of the Divisia regressions but not for M3. Next, the dynamic adjustment to the long-run relationship is modeled. Dynamic models are specified according to the general-to-specific approach, starting with four lags of each variable. Insignificant terms have been eliminated. Table 4 shows the final specifications, including the errorcorrection term, a dummy for German monetary union, and four seasonal dummies. Each of the dynamic models is satisfactory. For M3, lagged changes of real GDP have no significant effect. In all equations, the error-correction term is highly significant. For the Divisia aggregate, about 20 percent of the deviation from equilibrium is corrected each quarter, whereas the error-correction term for M3 is slightly lower. Table 3 Estimation Results for the Long-Run Relation Variable Constant EXPR/GDPR PQ/GBY Adj. R2 Durbin-Watson Dickey-Fuller QM3R QM3R M3R –22.002 0.875 –0.224 –23.519 0.869 –0.015 –8.722 0.971 –0.013 0.975 0.514 –3.935 0.977 0.468 –4.203 0.988 0.326 –3.228 NOTES: All regressions include a dummy for German Unification, which takes the value of 1 from 1990:3 on, and seasonal dummies. The first column shows the regression of the real Divisia aggregate on real expenditure and the price dual. The second column regresses the real Divisia aggregate on real GDP and the government bond yield. The third column gives the results for real M3, real GDP, and the government bond yield. The last line shows the Dickey-Fuller test statistic for stationarity of the residuals. The critical value for the 5 percent level is –3.840 (McKinnon, 1991). See also notes to Table 2. 31 In fact, parameter estimates are superconsistent: they converge asymptotically against the true parameter values at an even faster rate than in usual OLS regressions. Small sample bias, however, can be severe (see Banerjee et al., 1993). 32 The Engle-Granger method is less efficient than the Johansen approach, since the long-run relation is estimated without the information in the dynamic adjustment. Moreover, with more than two variables, testing for the existence of multiple cointegrating vectors is impossible. On the other hand, the Johansen method is often very sensitive to the lag choice. 33 T-values are not shown, since their distribution is nonstandard. Money demand is estimated with the Engle-Granger method, which uses ordinary least squares to estimate the long-run relation. Though non-stationary variables are involved in the regression, parameter estimates remain consistent.31 If cointegration exists—that is, if the variables move together in the long run—the residuals must be stationary. Nevertheless, they may exhibit autocorrelation or non-normality because the dynamic adjustment is not modeled in the first step.32 Results for the long-run relations are shown in Table 3. Three different equations are estimated. The first column shows the results for the Divisia money demand regression, including the real Divisia quantity index, expenditures on consumption and monetary services, and the price dual. The second column regresses the Divisia index on GDP and the government bond yield. The third column gives the results for a conventional money demand equation for M3. All regressions include four seasonal dummies and a dummy for German Unification that takes the value of one from the third quarter of 1990 onwards and zero elsewhere. The income elasticity is close to unity in all three regressions, though the point estimate for the Divisia equations is slightly lower than that of simple-sum M3.33 The price dual elasticity is much higher for the CONCLUSION The Divisia index has microeconomic foundations and empirically performs better than the simple-sum M3. While the aggregation approach regards money as a durable consumption good yielding a flow of services, simple-sum aggregation treats money as a component of wealth in a simple accounting procedure. In this paper, a consistent framework for the aggregation of monetary assets in different currencies has been developed. With completely fixed exchange rates, the European Divisia index equals the conventional Divisia index, since depreciation vanishes. If a common currency is introduced, monetary assets of the same degree of liquidity become indistinguishable for the consumer and can be aggregated across countries by simplesum aggregation. The advantage of the Divisia index is likely to be important during the transition to monetary union, because this index can F E D E R A L R E S E R V E B A N K O F S T. L O U I S 58 SEPTEMBER/OCTOBER 1997 take account of increased exchange-rate stability. Moreover, it can cope better with financial innovation. The move to a currency union will liberalize financial markets and increase competition in the banking sector, and will presumably lead to new financial products in those countries where markets are still regulated. As payments systems still differ among the European countries, the Divisia index may give a more appropriate indication of liquidity in Europe until a completely integrated financial market has developed (see Spencer, 1995). Even after the financial markets have been completely integrated, the Divisia index would continue to be more valid than simple-sum measures, because substitution effects between assets with different degrees of liquidity will remain. Though the Divisia index performs slightly better, the empirical differences between the Divisia index and simple-sum M3 with regard to money demand are small. This lack of striking findings is probably a result of the degree of disaggregation, since the breakup into narrow money and quasi money is a very crude one. Nevertheless, the Divisia index of European monetary services may provide additional insight into money demand during the period of transition to monetary union. With more disaggregated data on monetary assets and the corresponding interest rates, the performance of the Divisia index relative to simple-sum indexes would likely improve. Therefore, the European Monetary Institute should monitor Divisia aggregates in addition to M3 during the transition to a monetary union. Table 4 Dynamic Equations Variable Constant DM3R(–1) DQM3R(–4)/ DM3R(–4) DEXPR DGDPR(–1) DPQ/DGBY RES(–1) Adj. R2 SEE Durbin-Watson DQM3R DQM3R DM3R –0.170) (–4.994) –0.019) (–4.369) 0.410) (5.060) 0.314) (3.508) 0.391) (4.582) –0.007) (–2.209) 0.194) (2.179) 0.189) (2.140) –0.077) (–4.023) –0.206) (–4.011) 0.305) (1.772) –0.009) (–2.988) –0.221) (–4.070) –0.006) (–4.023) –0.144) (–4.011) 0.824) 0.013) 1.902) 0.832) 0.012) 1.931) 0.773) 0.008) 2.003) NOTES: The dynamic equations include a dummy for German Unification, which takes the value of 1 in 1990:3, and seasonal dummies. “D” means first differences, (–1) and (–4) indicate that the variable is lagged one and four quarters, respectively. See also notes to Tables 2 and 3. Banerjee, Anindya, Juan J. Dolado, John W. Galbraith, and David F. Hendry. Co-Integration, Error Correction, and the Econometric Analysis of Non-Stationary Data, Oxford University Press, 1993. Barnett, William A. “The User Cost of Money,” Economics Letters (vol. 1, 1978), pp. 145-49. _______. Consumer Demand and Labor Supply: Goods, Monetary Assets, and Time, Amsterdam: North-Holland, 1981. _______. “The Optimal Level of Monetary Aggregation,” Journal of Money, Credit, and Banking (November 1982), pp. 687-711. _______. “The Microeconomic Theory of Aggregation,” New Approaches to Monetary Economics: Proceedings of the Second International Symposium in Economic Theory and Econometrics, William A. Barnett and Kenneth J. Singleton, eds., Cambridge University Press, 1987, pp. 115-68. REFERENCES _______. “Which Road Leads to Stable Money Demand?,” forthcoming in The Economic Journal. Anderson, Richard G., Barry E. Jones, and Travis D. Nesmith. “Monetary Aggregation Theory and Statistical Index Numbers,” this Review (January/February 1997a), pp. 31-51. _______, Douglas Fisher, and Apostolos Serletis. “Consumer Theory and the Demand for Money,” Journal of Economic Literature, (December 1992), pp. 2086-119. Anderson, Richard G., Barry E. Jones, and Travis D. Nesmith. “Building New Monetary Services Indices: Concepts, Data and Methods,” this Review (January/February 1997b), pp. 53-82. _______, and Yi Liu. “The CAPM-Extended Divisia Monetary Aggregate with Exact Tracking Under Risk,” Working Paper No. 194, Washington University, St. Louis, 1995. Artis, Michael J., Robin C. Bladen-Hovell, and Wenda Zhang. “A European Money Demand Function,” Policy Issues in the Operation of Currency Unions, Paul R. Masson, and Mark P. Taylor, eds., Cambridge University Press, 1993, pp. 240-65. Diewert, W. Erwin. “Exact and Superlative Index Numbers,” Journal of Econometrics, (May 1976), pp. 115-45. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 59 SEPTEMBER/OCTOBER 1997 Divisia, François. “L’indice Monétaire et la Théorie de la Monnaie,” Revue d’Economie Politique (1925), pp. 980-1008. Phlips, Louis. Applied Consumption Analysis, Amsterdam: North-Holland, 1974. Donovan, Donald J. “Modeling the Demand for Liquid Assets: An Application to Canada,” International Monetary Fund Staff Papers (December 1978), pp. 676-704. Pill, Huw, and Mahmood Pradhan. “Monetary Aggregation: A Reconciliation of Theory and Central Bank Practice,” International Monetary Fund Working Paper 118 (October 1994). Ewis, Nabil A., and Douglas Fisher. “The Translog Utility Function and the Demand for Money in the United States,” Journal of Money, Credit, and Banking (February 1984), pp. 35-52. Rotemberg, Julio J., John C. Driscoll, and James M. Poterba. “Money, Output, and Prices: Evidence from a New Monetary Aggregate,” Journal of Business and Economic Statistics (January 1995), pp. 67-83. Farr, Helen T., and Deborah Johnson. “Revisions in the Monetary Services (Divisia) Indexes of the Monetary Aggregates,” Staff Study 147, Board of Governors of the Federal Reserve System, December 1985. Spencer, Peter. “Should the ECB Adopt a Divisia Monetary Aggregate?,” paper presented at the CEPR/Irving Fisher Society Conference: What Monetary Policy for the European Central Bank?, 1995. Fase, Martin M. G. “Divisia Aggregates and the Demand for Money in Core EMU,” De Nederlandsche Bank Staff Reports No. 5, 1996. Swofford, James L., and Gerald A. Whitney. “Flexible Functional Forms and The Utility Approach to the Demand for Money: A Nonparametric Analysis,” Journal of Money, Credit, and Banking (August 1986), pp. 383-89. _______, and C. C. A. Winder. “Money Demand within EMU: An Analysis with the Divisia Measure,” De Nederlandsche Bank Quarterly Bulletin (September 1994), pp. 25-55. Fisher, Paul G., Suzanne Hudson, and Mahmood Pradhan. “Divisia Measures of Money,” Quarterly Bulletin, Bank of England, (May 1993), pp. 240-52. _______, and _______. “Nonparametric Tests of Utility Maximization and Weak Separability for Consumption, Leisure and Money,” Review of Economics and Statistics (August 1987), pp. 458-64. Gaab, Werner, and Andy Mullineux. “Financial Innovation, Monetary Aggregates and Monetary Policy in the UK and Germany,” Finanzmärkte, Finanzinnovationen und Geldpolitik, Schriften des Vereins für Socialpolitik, Bd. 242, D. Duwendag, ed., Berlin: Duncker und Humblot, 1996, pp. 73-110. _______, and _______. “A Comparison of Nonparametric Tests of Weak Separability for Annual and Quarterly Data on Consumption, Leisure, and Money,” Journal of Business and Economic Statistics (April 1988), pp. 241-46. Gaiotti, Eugenio. “Measuring Money with a Divisia Index: An Application to Italy,” Temi di Discussione del Servizio Studi 223, Banca d’Italia, 1994. _______, and _______. “A Revealed Preference Test for Weakly Separable Utility Maximization with Incomplete Adjustment,” Journal of Econometrics (January/February 1994), pp. 235-49. Green, H. A. John. Aggregation in Economic Analysis: An Introductory Survey, Princeton University Press, 1964. Thornton, Daniel L., and Piyu Yue. “An Extended Series of Divisia Monetary Aggregates,” this Review (November/December 1992), pp. 35-52. Issing, Otmar, Karl-Heinz Tödter, Heinz Herrmann, and Hans-Eggert Reimers. “Zinsgewichtete Geldmengenaggregate und M3 – ein Vergleich,” Kredit und Kapital (Vol. 26, Issue 1, 1993), pp. 1-21. Wesche, Katrin. “Aggregating Money Demand in Europe with a Divisia Index,” Discussion Paper No. B-393, SFB 303, Universität Bonn, 1996. Kremers, Jeroen J. M., and Timothy D. Lane. “Economic and Monetary Integration and the Aggregate Demand for Money in the EMS,” International Monetary Fund Staff Papers (December 1990), pp. 777-805. MacKinnon, James G. “Critical Values for Cointegration Tests,” Long-Run Economic Relationships, Robert F. Engle and Clive W. J. Granger, eds., Oxford University Press, 1991, pp. 267-76. Marquez, Jaime. “Money Demand in Open Economies: A Divisia Application to the U.S. Case,” New Approaches to Monetary Economics, Proceedings of the Second International Symposium in Economic Theory and Econometrics, William A. Barnett, and Kenneth J. Singleton, eds., Cambridge University Press, 1987, pp. 183-99. OECD. Purchasing Power Parities and Real Expenditures – GK Results, Vol. 2, Paris: Organisation for Economic Co-operation and Development, 1993. Osborne, Dale K. “Defining Money,” The New Palgrave Dictionary of Money and Finance, Peter Newman, Murray Milgate, and John Eatwell, eds., Macmillan, 1992, pp. 602-06. F E D E R A L R E S E R V E B A N K O F S T. L O U I S 60