The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Economic r a m .................... M I C R E V I E W Vol. 2 4 , No. 3 2 Rules Versus Discretion: Making a Monetary Economic Review is Rule Operational published quarterly by the Research D e p a rt b y John B. Carlson m ent of the Federal Reserve Bank of C leveland. Copie s of the Review The question of whether monetary policy should be conducted by rules known are available through our Public in advance to all or by policymaker discretion is an enduring debate. This Inform ation D epa rtm en t, paper traces the evolution of rule advocacy from the time of the Federal 216 /579 -215 7. Reserve Act and examines the problems associated with making a rule operational. The author concludes that some degree of discretion m ay be necessary and that it is useful to analyze policies on the continuum between a C oordinating Eco n o m is t: Randall W . Eb erts pure rule and pure discretion. Recent changes in the operating procedure illustrate how actual policy has varied within this continuum. Editor: W illiam G . M urm a nn A s sista n t Editor: Robin Ratliff Design: M ichael Galka Actual Competition, Potential Competition, 14 and Bank Profitability in Rural Markets Typesetting: L iz H an n a Opinions stated in Review Economic are those of the authors b y Gary W halen and not necessarily those of the Despite a great deal of previous empirical work, disagreement remains about Cleveland or of the Board of Federal Reserve B ank of the relationship between market concentration and bank performance. Further, Governors of the Federal Reserve the performance impacts of potential competition remain virtually unexplored. S y s te m . In this study, the author examines the relationship between bank profitability and both actual and potential competition. Risk is explicitly incorporated into the analysis, and possible simultaneity is also explored. The results suggest Material m ay be reprinted pro that n o n -M S A markets are contestable or that potential competition has a vided that the source is credited. significant influence on banks’ performance. Please send copies of reprinted material to the editor. Getting the Noise Out: Filtering Early 24 GN P Estimates b y John S cadding The accuracy of the U .S . Department of Commerce’s estimates of G N P is vital in order for policymakers and private agents to make informed economic decisions. The early, or provisional, G N P estimates for each quarter can be seen as rational forecasts of the final numbers. This article takes an alternative view: that the early G N P numbers are estimates of the final number, but estimates that are contaminated with error, or “noise.” The author investigates a w ay to adjust the numbers to remove the error and to produce more accurate predictions of the final G N P growth estimates. Comment 32 is s n 0013-0281 R u le s V e rs u s D is c r e tio n : M a k in g a M o n e ta r y R u le O p e r a tio n a l by John B. Carlson John B. Carlson is an economist at the Federal Reserve Bank of Cleveland. The author would like to thank Charles Carlstrom, Edward Gambier, William Gavin, Jam es Hoehn, Mark Sniderman, and E .J . Stevens for their helpful comments. Introduction The rules-versus-discretion debate is the most enduring, if not the most central, issue in mone tary policy. It concerns whether monetary policy should be conducted by rules known in advance to all or by policymaker discretion. For many years, the case for a monetary rule was associated with a particular proposal by Mil ton Friedman (1959). Building on a tradition initiated by Henry Simons (1936), Friedman introduced the idea that the effects of monetary policy were uncertain, occurring with long and variable lags. In short, he argued that discretion ary management of the money supply in the face of such uncertainty actually amplified economic fluctuations. Hence, Friedman argued for a constant-money-growth rule. The case for rules has changed fundamentally since an important paper by Kydland and Pres cott (1977). They show that precommitment to a rule could have beneficial effects that discretion ary policies cannot. Unlike Friedman’s argument, the Kydland-Prescott case was not specific to any one view of the world, but could be applied to a very general class of models. In principle, one cannot deny that a policy rule can have poten tially stabilizing effects. The example of Kydland and Prescott, how ever, trivialized an important concern of policy makers: how to account for uncertainty in the link between policy instruments and ultimate objectives. Once one allows for uncertainty, there is a potential role for flexibility to deal with variability in the links. To the extent that some variation is systematic and can be predicted, it is possible to incorporate feedback into a rule. However, some contingencies cannot be fore seen. When such events are potentially destabil izing, discretion may not be ruled out a priori. This suggests that it is reasonable to consider the idea of rules with discretion. Fischer (1988) has concluded that the dichotomy between rules and discretion should be seen as a continuum, in which the extent of the monetary authority is determined by the immediacy of the link between its actions and the attainment of the objectives. The actual practice of monetary policy can be viewed as a point on the continuum. Moreover, the rise of monetary targeting in the 1970s, which led to alternative operating procedures with differing degrees of commitment, illustrates that the degree of commitment to any rule can vary over time. Changes in the degree of com mitment are best understood when one confronts the difficulties in making rules operational. This paper reviews the historical development of the rules-versus-discretion debate and exam ines the problems associated with making rules operational. Section I traces the evolution of rule advocacy from the time of the Federal Reserve Act. Section II describes the actual operating procedures from the early 1970s to the present. The operational problems facing rule advocates are highlighted in Section III, and Section IV dis cusses how two recently proposed rules address the operational problems. Section V offers some concluding comments. I. Rule Advocacy in the United States After the Federal Reserve The original Senate bill to create the Federal Reserve System in 1913 contained a provision that the system should This provision was stricken by the House Committee on Banking and Currency and was not included in the original Federal Reserve Act, reflecting the dominant influence of the real bills doctrine at that time. By the late 1 9 2 0 s, however, several bills had been proposed to amend the Federal Reserve Act explicitly to include a provi sion for price stability. 1 Advocates of these bills essentially sought to legislate a rule establishing the primacy of the price-level objective. These efforts culminated in the Strong Hear ings, held by the House Banking Committee in 1926-1927. 2 The hearings initially considered a bill by Representative James G. Strong including a provision that “all the powers of the Federal Reserve System should be used for promoting a stable price level.” Specifically, Congressman Strong did not want the Federal Reserve to have the discretion to vary the price level for the pur suit of any other objective. While the bill instructed that the Federal Reserve’s discount-rate policy was to be deter mined with “the view of promoting price stabil ity,” no formula was specified. Thus, there was a certain vagueness about how the rule would be implemented . 3 It left open the role for discre tion in determining how much the discount rate should be altered when the price level deviated level. ■ 1 promote a stable price For a thorough review of the debate, see Fisher (1934). It should be from its objective. A subsequent version of the bill was even more ambiguous about the objec tive of price stability. Eventually, Congressional interest in establishing the primacy of the objec tive of price stability faded. The Simons Tradition In a widely celebrated article of 1936, Henry Simons initiated a case for rules that was to become known as the Chicago view. Specifically, Simons contrasted two sharply distinct ways to conduct monetary policy: one, to assign in advance specific responsibilities to a monetary authority to be carried out in accordance with well-defined operational rules; the other, to specify a general goal while allowing the mone tary authority wide discretionary powers to achieve the goal. The essential distinction is that the first regime defines the authority’s objective in terms of the means, while the second defines the objective in terms of the ends. Simons argued for rules in terms of means. His case was predicated on liberal (19th-century sense) principles. “The liberal creed demands organization of our economic life largely through individual participation in a game It calls upon the state to provide a stable framework of rules within which enter prise and competition may effectively control and direct the production and distribution of goods.” (Simons [1936], p. 1) The essential notion is that government is necessary for establishing laws that would define the rules for a “game” in which competitive free enterprise could flourish, but that government should not be a player in the game. The idea that government would manage the currency to manipulate aggregate economic outcomes meant that government would be a player and thus violated the liberal creed. An ideal rule according to Simons would be one that fixed the quantity of the money supply. He did not believe, however, that such a rule could be made operational without radical reform of the financial structure. Essentially, he believed that an unregulated financial sector was a source of great instability in money demand. This instability was reflected in the perverse behavior of velocity which, he argued, necessi tated a role for discretionary actions. Simons definite rules. with noted here that a provision for purchasing power was eventually incorporated in the Employment Act of 1946. However, the price-stability goal was not included as the primary objective as most advocates of price stability in the 1920s had sought. ■ 3 Hetzel (1985) notes that Congressman Strong and his supporters wanted to institutionalize the policy of Governor Strong (no relation) of the ■ 2 For an excellent discussion of the background and events surrounding the Strong hearings, see Hetzel (1985). New York Federal Reserve Bank, which they credited for the considerable price stability that existed after 1922. therefore suggested a number of ideal reforms to reduce the variability of velocity to levels condu cive to successful implementation of a fixedmoney-supply rule. That is, government would need to redefine the rules of the game to avoid having to manage the money supply. One proposed reform was the elimination of fractional-reserve banking. By requiring 100percent reserves on all demand deposits, Simons sought to reduce greatly the threat of bank runs and the consequent effects on hoarding money (velocity changes). Such a reform would also give the monetary authority direct control over the total money supply by making it equivalent to the monetary base. Simons recognized, however, that fixing the supply of deposits might merely serve to encourage the creation of effective money substi tutes that would also affect velocity. Thus, another “ideal” (but even more radical) reform would be to prohibit fixed-money contracts. Re stricting claims to residual equity or commonstock form would essentially drive a wedge between money and other assets and would tend to minimize the variability of velocity. In sum, Simons believed that a monetary rule in terms of means could be made operational only under a highly regulated financial system. Simons was not naive about the kind of assent that could be gained for such radical reforms in modern democratic societies. He thought that adoption of an appropriate framework could be implemented only after decades of “gradual and systematic reordering of financial practices.” Iron ically, liberal principles also seem to support the notion that financial institutions should be largely unregulated and free to offer any instruments they choose. Indeed, institutional reform has moved in the opposite direction of Simons’s ideal. Recognizing the practical difficulties of sharp changes in velocity and that his ideal reforms might be unattainable, Simons argued for a rule for price stability in the interim. Because this is a rule of ends rather than means, the opera tional procedures were not well defined. His basis for this practical solution was that it was the “least illiberal” of the alternatives he consid ered. Thus, he recognized that for immediate purposes a certain amount of discretionary lati tude was necessary. While Simons may have misjudged society’s willingness to adopt his ideal reforms (new rules), his liberal view of economic agents participating in a game was prescient about the future state of the debate. The Simons tradition was subsequently m odi fied and popularized by Milton Friedman (1948, 1959, 1969). Initially, Friedman offered detailed proposals much in the spirit of Simons. They included the 1 0 0 -percent reserves reform applied to both time and savings deposits at banks. Subsequently, however, Friedman changed tack, taking the position that the behavior of velocity, particularly the velocity of the M2 aggregate, was not so perverse in a relative sense, even under a fractional-reserve banking system. He argued that the discretionary actions of the Federal Reserve (albeit well-intentioned) were likely to be a more perverse source of economic instability. Thus, adherence to a constant-money-growth rule would lead to greater economic stability than would a rule with feedback, with or without discretion. In essence, Friedman maintained the idea that the monetary authority should not be a player in the game, but he eventually rejected the need for wholesale reform of the financial system. Friedman’s case for a constant-growth rule was based less on the liberal creed and more on pragmatism. His premise was that the economic impact of monetary policy occurs with a long and variable lag. Feedback, especially of the dis cretionary type, would have effects at the inap propriate time more often than not. Moreover, Friedman argued that political pressures and accountability problems under discretion are likely to exacerbate the problem. While Friedman’s case has intuitive appeal, it is difficult to justify in principle. Potentially sta bilizing effects of policy feedback could be ruled out a priori only if money were the exclusive determinant of nominal GNP in the short run. If other identifiable factors also have significant explanatory power, then judicious use of feed back can, in principle, reduce the variability of nominal GNP, even if the coefficients on lagged money are stochastic. O n the other hand, the sta bilizing effects of policy feedback with parameter uncertainty are smaller than when parameters are nonstochastic (see Brainard [1967] ) . 4 By eventually abandoning 100-percent reserves, Friedman also allowed a control prob lem: how to make a constant-growth rule opera tional for measures of inside money. Under 100percent reserves there would be virtually no distinction between money and monetary base. Since Friedman also proposed closing the dis count window, all money would essentially be outside money, and hence directly controlled by the Fed. ■ 4 When effects of monetary policy occur with a lag, there is a potential for instrument instability. The prospect of dynamic instability can be reduced with appropriate modifications to the objective function. As advocates for constant money growth dropped the idea of 1 0 0 -percent reserves, how ever, the issue of monetary control became relevant. When the measure of money is endo genous, the problem of making a constantmoney-growth rule operational is far from triv ial. Such was an important lesson of monetary targeting in the early 1980s. Perhaps recognizing this fact, advocates for money-growth rules now typically propose closing the discount window and adopting a constant-growth rule for the monetary base. Arguments for a monetary rule in the Simons tradition remain highly controversial in princi ple. One cannot rule out the possibility that an intelligent policymaker could effectively take account of incomplete information when decid ing optimal monetary policy. As Barro (1985) notes, “if the policymaker were also wellmeaning, then there was no obvious defense for using a rule in order to bind his hands in advance.” Moreover, Fischer argues, “At a formal level Friedman’s analysis suffered from the logi cal weakness that discretion seemed to dom i nate rules: if a particular rule would stabilize the economy, then discretionary policymakers could always behave that way— and retain the flexibility to change the rule as needed.” Kydland and Prescott The idea that discretion could always replicate a preferred policy rule seemed to provide a highly influential argument in which intelligent, wellmeaning policymakers should not be bound by rules. However, in a widely recognized paper, Kydland and Prescott (1977) demonstrate a fal lacy in this argument. It is now well understood that if economic outcomes depend on expecta tions about future policies, then credible pre commitment to a rule could have favorable effects on the economic outcomes that discre tionary policies cannot have. Applications of the Kydland and Prescott result to monetary policy are often developed in famil iar (and highly abstract) models of output and inflation. 5 These models assume that wagesetters and the monetary authority are engaged in a noncooperative game. In this game, wagesetters must specify the nominal wage rate in a ■ 5 before contract (their play) prices are deter mined (the policymaker’s play). Firms’ decisions to hire are made after prices are determined, so that the real wage is known. Since firms are assumed to be profit maximizers, the real wage determines the level of output for the economy. An essential feature of the game is that by determining the price level, the policymaker’s play determines the real wage and level of real output. Moreover, expectational errors of wagesetters determine the deviation of output from its full employment levels. Thus, the game yields the familiar output supply function ( 1) y Fischer (1988). The use ot a static model to illustrate dynamic inconsistency has been criticized as inadequate. The basic concept, however, has been devel oped in the context ot a dynamic model (see Roberds [19 86 ]). Since it is the concept we want to convey here, the static model suffices. + b( n - 7Te), n y* where and are output and inflation, is full employment output, and is the expected inflation rate. The policymaker is assumed to have a loss function quadratic in the deviations of inflation and output from target levels. Here, desired inflation is assumed to be zero. (2) ne L = an2 + (y - ky*)2 The target rate of output is assumed to be above the natural rate, that is, 1. One motivation for this assumption is that tax distortions and unemployment policy cause the natural rate to be too low from a social point of view. Alterna tively, one might argue that the labor market is dominated by large unions (see Canzoneri [1985]). He assumes that the labor supply curve includes only union members and that wagesetters’ behavior systematically excludes other workers. By contrast, the loss function includes all workers. Others have argued that equation ( 2 ) is not really a measure of social utility, but reflects the bias of policymakers to underesti mate the natural rate of unemployment. To illustrate the advantage of a rule, consider the case in which the policymaker has discre tion in a one-period game. Because the policy maker chooses policy the wage-setters know that the policymaker has the incentive to take the and to induce higher employment with additional inflation, if possible. Given the known loss func tion, there is only one strategically rational expectation (that is, Nash solution) for inflation: k> specify the wage rate, after the wage-setters expected inflation rate as given (3) The particular example presented here is the compact static model in y = y* rre = a ~ lb ( k - )y*. 1 Under this solution, the policymaker has no incentive to choose an inflation rate higher than expected. The gains from the additional output would be more than offset by the loss of the additional inflation. Note also that if the policy maker had an objective for the inflation rate less than the expected inflation rate wagesetters acted, it would be inconsistent That a zero-inflation objective is not cred ible with discretion is an example of the problem of The value of the loss function evaluated at the solution is denoted as and is given by before ward. after time inconsistency. Ld (4) Ld = u - i ) y 2[i + * V ] . If the policymaker could credibly precommit to a policy of zero inflation, that is, a dynami cally consistent inflation objective, the loss func tion would be (5) Lp = ( k - l ) 2y * 2. Lp Ld, Since < precommitment to a zero-inflation objective affects expectations in a way that leads to a more favorable outcome than pure discretion would allow. Essentially, discretion buys nothing in terms of output, which is the same under both policies, but leads to an inflationary bias. To be sure, the basic result of Kydland and Prescott demonstrates in a very precise way a benefit to precommitment to a policy rule. Although developed in a highly abstract model, the result has been widely influential in aca demic research. A major shortcoming of the anal ysis, however, is that it trivializes the control problem. Specifically, it presumes that the policy maker has a deterministic operating procedure that enables precise control of inflation. Once disturbances are introduced into the model, the precommitment solution does not necessarily dominate the discretion solution. To analyze the control problem, Canzoneri considers a stochastic disturbance to money demand such that velocity follows a random walk. In his game, wage-setters cannot see the disturbance at the time they specify their wage, but the Federal Reserve has some forecast of money demand before it chooses its policy for money growth. If the Fed is left with some flex ibility, it can accommodate the predictable component of the change in velocity. As Can zoneri notes, this practice benefits both wagesetters and society as a whole. Thus, the policy problem becomes one of trading off flexibility needed for stabilization with the constraint needed for eliminating the inflation bias. 6 The discussion thus far has been in the con text of a one-period framework. In reality, how ever, the central bank has a horizon that extends beyond one period. Indeed, this may explain why central banks are typically isolated from political pressures by design. It is now widely understood that in a multiperiod context, the Fed may be able to establish a reputation that serves the same purpose as a monetary rule. This possi bility has been investigated by Barro and Gordon ( 1983a, 1983b). They find that under certain con ditions, reputation-building can lead to a result that is superior to pure discretion, although not as good as precommitment to a rule. Barro and Gordon assume, however, that wagesetters eventually have access to the same infor mation as the Fed. Canzoneri shows that when the Fed has its own private forecast of money demand, it has an incentive to misrepresent its intentions. 7 He further demonstrates that no sta ble resolution of the credibility problem can rely on the Fed’s own announcement of its forecast. When the Barro and Gordon model is modified to account for asymmetric information, the Fed cannot build sufficient credibility by simply run ning a noninflationary policy for a few periods. Rogoff (1985) has shown that other solutions may mitigate the problem of dynamic inconsis tency. One such solution is that society can benefit by choosing a “conservative” central banker— one that places a high cost on inflation. In the context of the simple model above, this means that the central bank places a high value on parameter in its loss function. Equation (4) reveals that as gets large, the value of the loss function diminishes, ultimately approaching the value of the precommitment solution given in equation ( 5 ). Like Barro and Gordon, Rogoff assumed sym metric information. When the Fed has private information, it has the incentive to appear more conservative than it actually is; the wage-setters have no way of telling. The implication is that there could be periodic inflationary breakdowns followed by sustained periods where credibility builds and wage-setters learn the true intentions of their central bank. Unfortunately, Canzoneri shows that it is no simple matter to legislate incentive-compatible rules that would remedy the problem posed by private information. a a ■ 6 Fischer (1988) demonstrates in a formal model that when control error exists, the ordering of the loss functions under precommitment and discretion is ambiguous. ■ 7 If the money demand forecast were predicated on a stable model over time, it would be preferable for the Fed to commit to a contingent rule based on that model forecast. Thus, while the rule would allow flexibility, it would not admit discretion. Given the absence of evidence of stability in money demand, such a rule seems infeasible. Rogoff also demonstrates that under certain conditions, intermediate targeting may also pro vide a reasonable solution to the problem of dynamic inconsistency. By providing the central bank with incentives to hit an intermediate target, it is possible to induce fewer inflationary wage bargains in the context of his model. While the Rogoff result demonstrates some a priori basis for intermediate targeting, his analysis abstracts from many problems the policymaker faces in practice. Nevertheless, the literature since 1977 suggests there is a reasonable basis for some precommitment— if not to a rule for all time— to some monetary policy on a continuum between a pure rule and pure discretion. II. The Operating Strategy of the Federal Reserve The operating strategy of the Federal Reserve can be viewed as a commitment to a policy on the continuum between a pure rule and pure discre tion. The rule-like elements are embedded in the Fed’s monetary targeting procedure. Mone tary targets are not ends in themselves, but are intermediate variables between the instrument variables that the Fed directly controls, such as the federal funds rate or nonborrowed reserves, and ultimate goals, such as price stability and stable output growth. Thus, intermediate target variables must be closely linked to both ultimate objectives and instruments. The use of intermediate targets has been criticized as redundant and inefficient from a control-theoretic perspective (see B. Friedman [1975]). These objections, however, are based on the assumption that policymakers have pre cise, reliable knowledge about the relationships between instruments and final objectives. In practice, policymakers see great uncertainty in these links and doubt that such relationships could be captured by econometric models accu rately enough to be operationally useful (see Black [1987]). In contrast, intermediate target variables are seen as relatively more controllable than ultimate variables. Moreover, policy decisions are made by major ity rule. It is therefore difficult, if not impossible (Arrow’s theorem) to obtain a consensus for adopting a particular social objective function, which is necessary under direct targeting of final objectives. Under an intermediate targeting strat egy, the Fed does not need to specify numerical objectives for goal variables. Intermediate targeting strategies can vary sub stantially in degree of flexibility or commitment. In principle, intermediate targets may or may not be designed to allow feedback. For example, a target could be specified for a five-year horizon without allowing for revisions, or for a three month horizon to accommodate frequent adjustments based on new information. Also, the operating procedure used to control the target variable may or may not allow for a high degree of discretion. Thus, operating rules could be highly automatic with infrequent discretionary input or be judgmentally modified day-to-day, based on the latest information. Actual practice of monetary7 targeting indicates that the degree of flexibility and discretion incorporated into the strategy7 is influenced by two key factors. The first is evidence concerning the stability of the relationship on which the strat egy is based. If there is a broad consensus about the reliability of the relationship between the intermediate target and ultimate goals, then it is more likely that a central bank would be willing to commit to closer targeting of the variable with less feedback from other sources, whether dis cretionary or not. The other key factor is the cen tral bank’s credibility or reputation in containing inflationary expectations. If the central bank establishes its credibility by avoiding inflationary policies, then the public and Congress are gen erally more willing to accept a greater degree of discretion in strategy and tactics. The interplay of these factors may well account for the increased reliance on monetary aggregates as intermediate targets during the early 1970s. Before the mid-1960s, there was scant evidence that discretion exercised by the Federal Reserve provided a substantive basis for inflationary expectations. Nominal interest rates were, on average, too low to indicate expecta tions of rising inflation. The public apparently believed that the Fed would “take the punch bowl away just as the party got going,” a percep tion consistent with Rogoff s notion of a conser vative central bank. Although the Federal Reserve had intermediate targets for interest rates— a strat egy that is now widely viewed as potentially defective for avoiding inflation— the Fed seemed to use its discretion judiciously in avoiding infla tion and hence in assuaging public doubt about the efficacy of its operating strategy. By the early 1970s, however, a basis for doubt was beginning to emerge, as inflation had accelerated to new and persistently high levels. Over that decade the Fed gradually strengthened its reliance on monetary aggregates as a source of information about its ultimate objectives. While the process was initially internal only, the Fed began to announce publicly its desired annual growth ranges for selected monetary aggregates in response to a Congressional reso lution in 1975. Evidence in the early 1970s con vinced many that the relationship between money and nominal GNP— as summarized by velocity—was sufficiently reliable to choose monetary targets over annual, or even longer, horizons. Also, the parallel rise in the price level offered simple but persuasive evidence that infla tion could be slowed by slowing growth of the monetary aggregates. In 1979, the Fed adopted a strategy for disinflation by gradually reducing the rate of money growth from year to year. The strategy was coupled with an automatic feedback rule to enhance monetary control and demonstrate a commitment to the strategy. Over most of the 1970s, the Fed used the federal funds rate— the interest rate banks charge one another on overnight loans of reserves— as its operating target for controlling money growth. Specifically, it sought to influence the quantity of money the public demanded by altering the opportunity cost of money. For example, if money growth was too rapid, it attempted to raise the federal funds rate, and thereby raise other short-term rates. The higher rates were expected to slow money growth by inducing the public to shift from monetary assets to other financial assets. Over longer horizons, higher interest rates might also be expected to slow spending growth and hence the transactions demand for money. In practice, however, there is always substantial pressure for the Fed to minimize interest-rate movements, particularly interestrate increases. For this reason and others, the Federal Reserve did not respond sufficiently promptly or intensively to keep monetary growth from accelerating in the 1970s. By late 1979, the inflation rate had accelerated to double-digit levels. Financial markets, espe cially foreign markets, began reacting strongly to the inflationary developments. The dollar was falling rapidly as foreign investors appeared to doubt the Fed’s resolve to contain inflation. In response to the evident inflationary pressures, the Federal Reserve adopted a new set of tactics “as a sign of its commitment to longer-run re straint on money growth” (Lindsey [1984], p. 12). These tactics in effect eliminated a substan tial degree of discretion that the Fed had used to smooth short-term interest-rate movements. The new procedures sought to control money growth by maintaining a short-run target path for nonborrowed reserves. As Lindsey describes, “holding to a nonborrowed reserves path essen tially introduces in the short run an upward slop ing money supply curve on interest rate and money space” (p. 12). In effect, the nonborrow ed reserves target created an selfcorrecting mechanism that would partially resist all deviations of money from target. If money growth in a given week moved above target, the prespecified level of nonborrowed reserves vir tually assured that the federal funds rate would move upward. In sum, the Federal Reserve gave up its discretion to minimize federal funds rate movements to assure financial markets of its commitment to the disinflation strategy. While the new procedure involved substantial commitment at the tactics level, it permitted sig nificant discretionary feedback at the strategy level. Under the strategy, the FOMC was free to change its short-term monetary target to take account of new information— a practice that led to significant deviations of money from announced annual targets. Such discretionary feedback was deemed necessary as evidence mounted that the velocity of money was not as reliable as expected. It was well understood at the time that dereg ulation in financial markets, changes in transac tions technology, and disinflation were having a substantial impact on individual portfolios and hence on the velocity of money. While such fac tors could account for the target misses in a qual itative sense, policymakers lacked means to pre dict the impact on money growth in order to specify reliable target values. By August 1982 the evidence was compelling that the behavior of velocity had been altered in some permanent way. Because time was needed to identify the new patterns of velocity behavior, attempts to con trol monetary aggregates closely appeared futile. Consequently, the Fed abandoned its operat ing procedure and hence its commitment to a fixed path of nonborrowed reserves in the short run. It de-emphasized the role of M l and adopted a more flexible operating strategy. Since the fall of 1982, the Fed’s operating target has been the aggregate level of seasonal plus adjustment borrowings at the discount window. Under this procedure, the FOMC specifies a short-term objective for this variable at each of its regularly scheduled meetings (at approxi mately five- to six-week intervals). Unlike with the nonborrowed reserves oper ating target, the current procedure does not produce automatic self-correcting federal funds rate responses to resist divergences of money from its long-run path. Substantial changes in the federal funds rate are largely a consequence of judgmental adjustments to the borrowings target. Thus, the Fed has regained much of the automatic leeway to smooth short-term interest rate changes that it had prior to 1979It is important to note that by the end of 1982 the disinflation process had become credible to most of the public. Financial markets, particu larly those for fixed-income securities, reacted favorably to the procedural change. Long-term interest rates continued to decline substantially after the Fed announced abandonment of the nonborrowed-reserves procedure. Moreover, over the long term, wage demands moderated to pre-1970s levels and have been persistently moderate to this day. 8 Such would seem strong evidence that wage setters haven’t suspected the Fed of “cheating” on its goal of reducing and maintaining lower inflation. The evolutionary cycle of the Federal Reserve’s operating procedure provides a useful illustration of how the degree of discretion has varied in response both to evidence concerning the reli ability of the money-income relationship and to the reputation of the Fed. As the Fed’s credibil ity on inflation appeared to wane in the 1970s, it adopted procedures that increased reliance on monetary aggregates as intermediate targets and limited its discretion to smooth interest rates. As evidence suggested a breakdown in the behavior of velocity, the degree of commitment to mone tary control diminished to allow the necessary operational flexibility. By that time the Fed’s commitment to maintaining lower rates of infla tion had become credible. While the actual strat egy can be characterized as a monetary rule with varying degrees of discretion, it never incorpo rated the degree of commitment that most monetarists had hoped for— one that would have not altered monetary targets at all. III. Problems with Making Rules Operational The review of the Federal Reserve’s actual oper ating strategy also serves to highlight a number of potential problems with making rules opera tional. Poole (1988), a longtime monetary rule advocate, recently concludes that “there is a serious and probably insurmountable problem to designing a predetermined money growth path to reduce inflation.” Essentially, he argues that it is not possible to reliably quantify the effects of disinflation on money demand and, hence, on velocity. 9 Thus, managed money is ■ 8 For evidence concerning moderation in compensation demands, see Groshen (1988). unavoidable during the transition to lower infla tion. While Poole accepts the eventual efficacy of a constant-growth rule, he believes there is no formula to determine when the discretionary mode should terminate. Presumably, it would only be after inflation has been eliminated. Even if the transition to lower inflation were no longer operationally relevant, the experience of the early 1980s makes it clear that money demand and velocity have also been independ ently affected by regulatory change and by devel opments in transactions technology. McCallum (1987) has recently argued that a rule should not rely on the presumed absence of the effects of such changes. This principle of rule design pre cludes simple, fixed rules like the constant growth rate of money (or monetary base). Oper ational feasibility demands that a monetary rule should at least be flexible enough to accommo date the effects of such changes on velocity. Recognizing a need for some form of flexibil ity, some pure-rule advocates now propose nondiscretionary feedback rules. Nondiscretionary feedback requires specification of a formula link ing goal (or target) variables to policy instru ments. The formula presumes the existence of some reasonably stable and hence reliable model, that is, one that characterizes sufficiently well the relationship between instruments and objectives. The absence of a consensus in macroeconom ics about an appropriate model poses a serious obstacle for gaining assent for any feedback rule in practice. While most economists adopt perspective, few seem willing to accept the notion that a particular (especially simple) characterization of the economy would be suffi ciently reliable for long periods. Even among rule advocates sharing a common perspective, there are likely to be subtle differences about the formula specification that may splinter support for a given rule. This problem of model uncertainty is com pounded by the important demonstration by Lucas (1976) that “structural” models are in gen eral not invariant to the way in which policy is implemented. Since this critique, there has been no widely accepted means of evaluating opera tionally concrete policy proposals. 1 0 While many large-scale econometric models have met the market test, few economists seem convinced by policy evaluations based on particular economet ric models. particular a ■ 10 Advocates of rules sometimes argue that if a nondiscretionary rule were to be implemented, relationships would stabilize, leading to more favora ble outcomes than suggested by simulations based on historical relationships. ■ 9 This point is an example of a more general result of Lucas (1976), which is discussed below. While this purely a priori theoretical argument is consistent, it does not appear to be convincing to most economists. Without a consensus about how monetary pol icy affects aggregate economic outcomes, it is not compelling to argue that expectations of economic agents (for example, wage-setters) are based on any one model of the economy. Any given rule could possibly be perceived as unsus tainable by a sufficient number of agents such that the rule would not be credible in an aggre gate sense. If agents believed the rule was unsus tainable, the game between agents and policy makers would become extremely complicated, with no apparent solution. Thus, it would not be clear that commitment to a rule would be bene ficial. It would seem useful that a rule advocate demonstrate that favorable consequences of a proposed rule would be robust to alternative models of the economy. IV. Two Recently Proposed Rules Two recently proposed rules by McCallum (1987, 1988) and Hall (1984) illustrate how the debate over rules versus discretion has evolved to a more operationally concrete level. Both authors appeal to the result of Kydland and Prescott as a justification for implementing their rules. Both also recognize a need for flexibility and address operational problems. In sharp contrast, how ever, is the way they incorporate flexibility. McCallum proposes a nondiscretionary feed back rule for nominal income using the mone tary base as the instrument. The target path of nominal income is fixed and grows at a pre specified rate of 3 percent per year. The feed back formula is (6 ) A b, = 0.00739 - ( / I + X(x* - .*,_]), 1 bt - 6 )k M - vt_17] t), where log of monetary base (for period = log of base velocity, log of nominal GNP, and target path for nominal income. The constant term 0.00739 is simply a 3 per cent annual growth rate translated into quarterly logarithmic units. The second term subtracts the average growth rate of velocity, approximated by the average difference in the logarithm of velocity over the previous four years. This term can be thought of as a simple time-series estimate of trend velocity growth. The third term specifies how policy is to respond to deviations of nom i nal income from its target path. The moving average of velocity growth is a simple statistical filter designed to detect per manent changes in velocity growth. As such, it provides a mechanism to maintain a long-term vt x* - xt = correspondence between the current base growth path and the long-term nominal objec tive to account for changes in transactions tech nology. Given the length of the moving-average period (four years) and the absence of any sys tematic feedback from interest rates, however, the rule provides virtually no adjustment in response to the current state of the business cycle or to financial conditions. 11 The third term provides feedback to assure that nominal income ultimately returns to its trend path. The choice of parameter A incorporates some degree of flexibility to deal with the poten tial problem of instrument instability. This prob lem arises when effects of policy7 occur over time as they do in actual economies, particularly those with sticky prices. Large responses to maintain a target path in the near term could lead to longerterm effects in the opposite direction, requiring even greater offsetting policy responses in later periods. This sequence would be unstable if responses and effects were to become ever in creasing. The value of A (presumably less than zero) should be chosen to minimize the poten tial for this dynamic instability, under the con straint that it be sufficiently large to provide ade quate responsiveness of base growth to target misses. McCallum suggests that a value of 0.25 appears to be somewhat robust for this objective over alternative models of the economy. If velocity growth were constant, and if nom i nal GNP were on its target path for a sustained period, the policy prescribed by McCallum’s rule would be the same as a 3 percent growth rule for the monetary base. Thus, McCallum’s rule is essentially a generalization of the constantmoney-growth rule. Because it is more general, it allows for flexibility to deal with some of the problems of making monetarist rules operational. Moreover, McCallum claims that because the monetary base is “controllable,” the rule can be accomplished with no operational discretion. 12 ■ 11 Recent evidence suggests that velocity has become increasingly interest-sensitive in the 1980s. To the extent that systematic effects of inter est rates could be reliably estimated, additional flexibility could be introduced into the rule as feedback to compensate for short-run variability in velocity. McCallum expresses doubt, however, that economists know enough to base policy on any one short-run empirical model. In this sense he defends, if only indirectly, the monetarist dictum of Friedman, in which monetary policy affects the economy with long and variable lags. ■ 12 Under current institutional arrangements, the total monetary base can be controlled only indirectly, working through effects of changes in interest rates on the demand for base components. Advocates of base targeting often call for institutional reforms— such as exactly contemporaneous reserve accounting and closure of the discount window— to enable direct control of the base. Alternatively, McCallum’s rule can be applied to the nonborrowed base, which is directly controllable under existing institutions. D In this sense McCallum’s proposal is a flexible version of a rule for means. The flexibility is extremely limited, however, involving only feed back from simple statistical models to maintain long-run relationships. No role is given to struc tural models that might allow feedback for short term economic stabilization. Such a rule shows little faith in macroeconomic models or in dis cretionary decisions of the Fed. Some rule advocates, on the other hand, pro pose a much greater role for economic models and judgment of the Fed. An example is an endsoriented rule advanced by Hall (1984). Under Hall’s strategy7, the Federal Reserve is instructed to stabilize the price level around a constant long-run average value. To make this strategy7 elastic in the short run, Hall proposes giving the Fed some prespecified leeway in achieving the target depending on the amount of unemploy ment. The permissible deviation of the actual price level, from its target, is defined by the simple numerical rule linking it to the devia tion of the unemployment rate, from its nor mal rate, presumed to be 6 percent: p, p* u, (7) 100(p - p*)/p* = A (u - 6 ). A The coefficient is to be specified by the Fed eral Reserve. Based on simulations, Hall tenta tively recommends that it equal eight. Specifically, this relationship is to be imposed as a constraint on policy instrument settings. In formal terms: “Monetary7 policy is on track when the deviation of the price level from its constant target is eight times the deviation of unemploy ment times its normal level Policy is too tight if the price deviation is less than eight times the employment devia tion; it is too expansionary when the price devia tion is more than eight times the employment deviation. The elasticity7 of 8 in this statement is a matter for policymakers to choose.” (Hall [1984], p. 140) Policy formulation under this approach would be prospective. Thus, the Fed would need to employ a model that links instrument variables to the price level and to the unemployment rate over the criterion period . 13 It would be free to use whatever model and instruments it chooses. Instrument settings would be determined by an iterative process. To begin, an initial forecast for the unemployment rate and price level would be compared against the rule formula to be judged for appropriateness— for example, too tight, too percent\. [presumed to be 6 easy, or on track. This process would thereby determine the direction in which instrument set tings should be changed, if necessary. A second round of forecasts would then be obtained and compared. The process would continue until the instrument settings yielded price-level and unemployment forecasts consistent with the rule. To impose discipline, Hall would require the Fed to be explicit about its forecasts, defending them publicly at the semiannual Congressional review and in comparison with private forecasts. Hall argues that forecasting errors of good pri vate forecasters would provide a sufficiently reliable standard to maintain unbiased out comes. If the Fed’s forecasts were consistently different from reputable private forecasts, and if the outside forecasts were more often correct, then the Fed would be under public pressure to modify its way of setting policy instruments. For Hall, the problem with discretion lies not with the use of faulty econometric models but with the absence of a commitment to an explicit rule for the price level. Both Hall and McCallum employ small empir ical models to generate simulations under their rules. McCallum uses a variety of models based on competing views to examine the robustness of his rule’s performance. His simulations sug gest that his rule would have produced a root mean square error (RMSE) of nominal income of around 2 percent from 1954 to 1985. This is approximately one-third the RMSE of actual GNP around its trend over the same period. He con cludes that his rule would have worked rela tively well in the United States. To address the criticism that his simulations are subject to the Lucas critique, McCallum notes that his rule relates nominal demand to nominal policy instruments. He argues that the sensitivity of parameters to policy regime changes is likely to be quantitatively less important for such rules than for rules that relate real to nominal varia bles, for example, based on Phillips curve m od els. Hall’s simulations, on the other hand, are based on the presumption that there is a relia ble (policy invariant) relationship between the of the inflation rate and the of the price level. 14 His simulation results suggest that price level variability7 and unemployment variability would have been less than actually experienced from 1 9 5 2 to 1 9 8 3 under the elastic-price rule. variability ity ■ 14 ■ 13 Based on the assumption that monetary policy affects the unemploy variabil both The analysis of policy in terms of the variability of unemployment and price level was developed by Taylor (1980, 1981). It is important to note ment rate reliably only after a yearlong lag, Hall argues that the criterion that there is no implied trade-off in this model between the inflation rate and period should be the forecast horizon for the year beginning six months ahead. trend output growth. H While the results presented under both rules appear favorable, few analysts seem convinced by small-model simulations. Experience with large-scale econometric models, for example, suggests that interest rates would vary sharply under McCallum’s rule. His models, which do not allow for interest-rate interactions, cannot account for the economic consequences of such interest-rate variation. Fischer (1988) argues that the natural vehicles for studying policy rules are the large-scale econometric models, many of which have met the market test. Nevertheless, he notes that it would be difficult to justify legis lating any nondiscretionary rule given the vari ety and inadequacies of existing models. O n the other hand, existing models may be no more reliable for discretionary decisions, particularly when policymakers may use them selectively to support their own prior beliefs. V. Some Concluding Comments The success of the U.S. disinflation strategy early in this decade helped reestablish the Federal Reserve’s credibility as an inflation fighter. Much of the reputational capital surely persists today. Recently, however, some analysts have questioned whether the current strategy is adequate to extend and maintain the progress against infla tion (see Black [1987]). A key concern is that the strategy may lack suf ficient institutional discipline to assure that short-term objectives—such as interest-rate smoothing— do not interfere with the achieve ment of longer-term price stability. This fear has led to a renewed interest in alternative strategies that are closer to a pure rule on the continuum between a pure rule and pure discretion. Ideally, a policy strategy should perform ade quately well under alternative views about aggregate economic relationships so that suffi cient numbers of agents believe that the rule could be credibly implemented. Rule advocates might well follow the example of McCallum and examine the robustness of their rule’s perfor mance, simulating with alternative models of the economy. The choice of criteria for “adequate performance” is of course a difficult and contro versial matter. We conclude here, as does Fischer ( 1 9 8 8 ), that the discussion of alternative policies is too important to be suppressed by the econ ometric evaluation critique. References Barro, Robert J. “Recent Developments in the Theory of Rules Versus Discretion.” Confer ence Papers: Supplement to the Economic foumal. 96(1985): 23-37. ------------ , and David B. Gordon. “A Positive Theory of Monetary Policy in a Natural Rate Model.” 91(1983a): 598-610. Journal of Political Economy. ------------ , and David B. Gordon. “Rules, Discre tion, and Reputation in a Model of Monetary Policy.” 1 2 ( 1 9 8 3 b): 1 0 1 2 1 . foum al of Monetary Economics. Black, Robert P. “The Fed’s Anti-inflationary Strat egy: Is It Adequate?” Federal Reserve Bank of (September/ Richmond. October 1987): 1-9. Economic Review Brainard, William. “Uncertainty and the Effec tiveness of Policy.” 57(May 1967): 411-25. American Economic Review Papers and Proceedings. Canzoneri, Matthew B. “Monetary Policy Games and the Role of Private Information.” 75( December 1985): 1056-70. can Economic Review. Ameri Fischer, Stanley. “Rules Versus Discretion in Monetary Policy.” National Bureau of Eco nomic Research Working Paper No. 2518 (February 1988). Fisher, Irving. Movement. Stable Money: A History of the New York: Adelphi Company, 1934. Friedman, Benjamin M. “Targets, Instruments, and Indicators of Monetary Policy.” 1(October 1975): 443-73. Monetary Economics. foum al of Friedman, Milton. “The Lag in Effect of Monetary Policy. ” 69( Oc tober 1961): 447-66. Reprinted in foum al of Political Economy The Opti mum Quantity o f Money and Other Essays. Chicago: Adeline Publishing Company, 1969: 237-60. A Program for Monetary Stability. ------ . New York: Fordham University Press, 1959. ------ . “A Monetary and Fiscal Framework for Economic Stability.” 38(June 1948): 245-64. Reprinted in Chicago: The University of Chicago Press, 1953- American Economic Review Essays in Positive Economics. D Groshen, Erica L. “What’s Happening to Labor Compensation?” Federal Reserve Bank of Cleveland. May 15, Taylor, John B. “Stabilization, Accommodation, and Monetary Rules.” 71 ( May 1981): 145-49. Hall, Robert E. “Monetary Strategy with an Elastic Price Standard.” A symposium sponsored by the Fed eral Reserve Bank of Kansas City, August 1984: 137-59. ________ . “Output and Price Stability: An Inter national Comparison.” 2(February 1980): 109-32. Economic Commentary, 1988. Price Stability and Public Policy. Hetzel, Robert L. “The Rules Versus Discretion Debate Over Monetary Policy in the 1920s.” Federal Reserve Bank of Richmond. (November/December 1985): 3-14. Economic Review Kydland, Finn E., and Edward C. Prescott. “Rules Rather than Discretion: The Inconsis tency of Optimal Plans.” 85(June 1977): 473-91. Economy. Journal of Political Lindsey, David E. “The Monetary Regime of the Federal Reserve System.” Conference on Alternative Monetary Regimes sponsored by Ellis L Phillips Foundation and Dartmouth College (August 1984). Lucas, Robert E. “Econometric Policy Evaluation: A Critique.” Supplementary Series 1(1976): 19-46. Journal of Monetary) Economics. McCallum, Bennett T. “Robustness Properties of a Rule for Monetary Policy.” CarnegieRochester Conference Series on Public Policy>. Carnegie Mellon University and National Bureau of Economic Research (February 1988). ________ . “The Case for Rules in the Conduct of Monetary Policy: A Concrete Example.” Fed eral Reserve Bank of Richmond. (September/October 1987): 10-18. Economic Review Poole, William. “Monetary Policy Lessons of Recent Inflation and Disinflation.” 2(Summer 1988): 73-100. nal of Economic Perspectives. The Jour Roberds, William. “Models of Policy Under Sto chastic Replanning.” Federal Reserve Bank of Minneapolis Research Department Staff Report 104 (March 1986). Rogoff, Kenneth. “The Optimal Degree of Commitment to an Intermediate Monetary Target.” C4(November 1985): 1169-89. Quarterly Journal of Economics. Simons, Henry C. “Rules Versus Authorities in Monetary Policy.” 44(February 1936): 1-30. my. Journal of Political Econo American Economic Revieu>Papers and Proceedings. Journal of Economic Dynamics and Control. A c t u a l C o m p e titio n , P o te n tia l C o m p e titio n , a n d B a n k P r o fita b ility in R u r a l M a r k e ts by Gary Whalen Gary Whalen is an economic advisor at the Federal Reserve Bank of Cleveland. The author wishes to acknowledge the helpful comments of Kelly Eakin. Introduction The nature of the relationship between the struc ture of the market in which banks operate— the number and size distribution of actual competi tors in a market— and their performance has been examined in a considerable number of empirical studies over the past 20 years. 1 Industrial organi zation economists have investigated the structure/ performance relationship for a wide variety of intra- and interindustry samples of firms. The typical maintained hypothesis has been that explicit or tacit collusion is more likely in markets with a limited number of large competi tors and should result in a statistically significant positive relationship between market concentra tion and the profitability of firms operating in the market. Definitive support for this hypothesis implies that an activist antitrust policy aimed at limiting merger-related increases in concentra tion is an appropriate public policy goal. A positive concentration/profits relationship has been found in some, but far from all, of the empirical studies investigating bank market structure and performance. The mixed results of this body of empirical work have been inter preted in widely different ways. ■ 1 For reviews of this work, see Rhoades (1982), Gilbert (1984), and Osborne and Wendel (1983). Some researchers, predisposed to accept the reasonableness of the concentration/collusion hypothesis, have concluded that the weight of the evidence supports this position and have advanced a number of reasons to discount the lack of consistent empirical support for the expected relationship between concentration and bank profitability. 2 One is that the equations estimated in many of these studies have been misspecified, possibly biasing the estimated coefficient on the concentration variable. In par ticular, several researchers have suggested that market concentration might impact bank man agement’s risk-return preferences or opportuni ties. 3 Specifically, bank management operating in concentrated markets might trade off potential monopoly profits for lower risk. If this is the case, significant concentration-related differences in profitability might not be evident in studies that fail to explicitly control for risk. Other researchers have argued that the single equation estimation techniques typically used in previous empirical work, even those where risk measures have been included as additional ■ 2 This is the conclusion of Rhoades (1982). ■ 3 See Heggestad (19 77), Rhoades and Rutz (1982), Clark (1986b), and Liang (1987). □ explanatory variables, may have biased the results. 4 In their view, profitability and risk are determined simultaneously, so we should rely only on the results of studies where the relation ships between these variables and concentration are investigated using simultaneous equation estimation techniques. Yet another group of researchers argue that the concentration/collusion hypothesis is unreason able because it embodies a questionable implicit assumption: that technological conditions, regu lation, other barriers to entry, or the threat of predation allow colluding firms in concentrated markets to disregard potential competitors. Concentration-related monopoly power and profits can exist and persist only when there is no threat of entry by potential competitors. 5 Mar kets in which this type of behavior can occur have been given the label “noncontestable.” In theoretical work, researchers have shown that when entry and exit are not precluded, or a mar ket is contestable, then outcomes can approxi mate those of perfect competition even if the number of actual competitors is quite small or if concentration is high . 6 Consequently, firm prof itability should not be expected to vary with concentration. The possibility that potential competitors may significantly affect the prices charged and profits earned by incumbent firms has been recognized for some time . 7 Until quite recently, however, banks and other financial intermediaries faced numerous regulatory and legislative constraints on geographic location, on permissible products and services they could offer, as well as on the prices they could charge. Thus, few of the geo graphic and product markets in which banks operated approximated the contestable ideal. This situation has changed dramatically in the past 10 years. A large number of states have reduced intrastate and, more recently, interstate ■ 4 This is the conclusion of Clark (1986b) and Liang (1987). ■ 5 See Brozen (1982) and Baumol, Panzar, and Willig (1982). ■ Actually, researchers have differentiated markets according to the 6 barriers to geographic expansion by commercial banks and by savings and loan institutions. In addition, the repeal of usury laws and removal of Regulation Q ceilings on deposit rates have left financial intermediaries basically free to compete on a price basis. Empirical investigations of scale and scope economies in banking suggest that small-scale entry is not precluded by cost conditions. 8 A negligible amount of the costs of branching appears to be sunk. These circumstances suggest that banking markets— at least in states that have liberalized branching to some extent, facilitating entry by out-of-market firms— have become contestable. Alternatively, potential competition may have become an effective disciplinary force, which could explain the absence of a strong positive concentration/profitability relationship in some of the more recent empirical studies. 9 Researchers who do not subscribe to the con centration/collusion hypothesis have offered an alternative explanation for the significant posi tive relationship between concentration and profitability reported in some previous studies. They argue that such a finding need not neces sarily signal collusion or indicate causation run ning from concentration to profitability. In their view, labeled the “efficient structure hypothe sis” (ESH), superior efficiency, management, or luck could result in increased firm profitability and market share and, ultimately, in higher con centration. 1 0 If the ESH is correct, then the posi tive relationship between concentration and profitability detected in empirical work where a market share variable is not included is spurious and simply reflects the correlation between market share and concentration. At present, then, there continues to be a great deal of uncertainty and disagreement about the relationship between market concentration, potential competition, and bank performance. Very few of the numerous previous studies have incorporated risk, controlled for market share, and investigated possible simultaneity. More important, virtually no empirical work on the impact of potential competition in bank ing, or in any other industry for that matter, has degree to which they are contestable. A t one extreme are noncontestable markets. A t the other extreme are perfectly contestable markets. In essence, perfectly contestable markets are ones in which entry and exit are costless. This, in turn, implies no barriers of any kind to entry and exit. In particular, ■ 8 See Berger, Hanweck, and Humphrey (1986). zero sunk costs are required to enter the market. Markets in which entry and exit can occur but are not costless have been labeled imperfectly contestable. ■ 9 In such markets, potential competition is expected to influence the perfor concentration/profitability relationship for a subsample of banks drawn from For example, Evanoff and Fortier (1988) find evidence of a positive mance of incumbent firms. For a more detailed discussion of these issues, see unit banking states but not for the subsample drawn from states where Schwartz (1986), pp. 37-48, and Morrison and Winston (1987), pp. 53-60. branching is permitted. ■ 7 ■ 10 This possibility was noted in Bain (1949) more than 30 years ago. See Smirlock (1985). been done to date. 11 A number of circumstances make banking an ideal subject for such research. The partial, gradual elimination of geographic barriers to market entry, cost conditions, and the local nature of banking markets mean that entry can occur if market conditions warrant and that the number of potential bank entrants for each local market can be determined. This paper attempts to provide more defini tive evidence on the relationship between com petition and bank profitability. The relationship between bank profitability and both actual and potential competition is examined in a frame work that explicitly includes market share and risk variables. Further, the impact of possible simultaneity is also explored. The sample consists of 159 banks drawn from non-MSA (metropolitan statistical area) counties in Ohio. The focus is on non-MSA counties for several reasons. First, the number of actual bank competitors in a typical non-MSA county is gen erally small, and concentration is high relative to MSAs in the state. Second, economic and demo graphic characteristics of rural counties generally make them less attractive for entry than urban counties. Finally, actual and potential competi tion from out-of-market and nonbank suppliers of financial services is likely to be limited. Thus, if the concentration/collusion hypothe sis is correct and if potential competition is a rel atively unimportant determinant of firm perfor mance, supporting empirical evidence is likely to be obtained from this data set. Conversely, ab sence of support for the concentration/collusion hypothesis and the finding that potential com petition impacts bank performance in rural markets is strong evidence that local banking markets, both rural and urban, are contestable. The time interval examined is from 1979 to 1981. This particular period was chosen because the bank branching law in Ohio was liberalized in January 1979. Before then, de novo branching was limited to a bank’s home office county. Under the new law, banks could branch de novo into all counties contiguous to the county in ■ 11 The only explicit empirical test to date is Hannan (1979). In many structure/performance studies, the sign and statistical significance of coeffi which their head office was located. Thus, the partial removal of geographic restrictions on branching created an identifiable number of po tential bank entrants for each county in the state. The choice of a three-year time period appears somewhat arbitrary. However, a period of this length should be short enough to ensure that ongoing expansion activity by banks does not materially affect the measure of potential com petition used in the study. It should also be long enough to allow any performance impacts attributable to potential competition to be detected statistically. In the following sections, we discuss the model to be estimated, describe the sample and estimation techniques, and present the results. A summary and conclusions follow. I. Model Specification Unfortunately, there continues to be no strong consensus about the “best” microeconomic model of the banking firm. As a result, researchers disagree about how the profitability equation to be estimated—whether a single reduced-form equation or a structural equation in a simultaneous system— should be specified. No attempt is made here to resolve the theoreti cal debate. Our approach is simply to estimate versions used in previous studies, with market share, risk, and potential competition variables explicitly included. Thus, the profitability equations estimated had the following general form: (1) where PROF, = f(A C i, PC,, MSr RISKj , Z,) PROFi : a measure of the profitability of bank i AC): a proxy for actual competition in the market in which bank i operates PCt: a proxy for potential competition faced by bank i MSt: the market share of bank i RISK;: a measure of the overall risk of bank i Z; : a vector of additional control variables cients on branching law dummies in estimated profitability equations are used to draw inferences about the intensity of potential competition. In others, the statistical significance (or lack of significance) of the estimated coefficient on the concentration term is used to obtain insight on this issue. In fact, very few explicit empirical tests of contestability/potential competition have been done for any industry, including the airline industry, which Baumol, et al. cited as an example of one with contestable markets. The study by Morrison and Winston (1987) m ay be the only one published to date. The profitability measure employed as the dependent variable in this study is rate of return on equity (net income after taxes, excluding se curities gains and losses, divided by book equity, both measured at year-end) averaged over the three years from 1979 to 1981. This profitability measure best reflects the efforts of managers interested in shareholder wealth maximization. The determinants of profitability of primary interest in this study are actual and potential competition. The former is proxied in two alternative ways: by incumbent firm market concentration and by the number of actual competitors. The latter is proxied only by the number of potential competitors. 12 The precise form of the relationship between the proxies for actual competition, potential competition, and profitability are unclear and could take a number of different forms. The consensus view is that actual competition will be more intense and incumbent profitabil ity will be lower, the greater the number of actual competitors or the lower the market con centration. The relationship between these proxies, the likelihood of collusion, and the intensity of competition and ultimately profita bility might not be linear, however. 13 For example, the marginal impact of additional actual competitors might not be constant, but could decline as the number of competitors increased. As a result, we also investigate non linear relationships between the proxies for actual competition and profitability. As long as entry into rural banking markets is not precluded, the prices and profits of incum bents should also vary systematically with the number of potential entrants. However, there is some uncertainty about the precise form of the relationship between incumbent profitability and the of potential competitors because the relationship between the number of potential competitors and the intensity of potential com petition is unclear . 14 The standard view appears to be that the larger the number of potential entrants, the greater the perceived threat of entry and the lower the incumbent prices and profits. Some writers, however, have suggested that when more than one potential entrant exists, each potential entrant will recognize that entry by others could occur and could impact its number ■ 12 Since it is not clear that the size distribution of potential competitors influences their performance impact, and since construction of a measure of expected profit. 15 Researchers have demon strated that mutual awareness among potential entrants could cause the relationship between the number of potential entrants and the overall likelihood of entry to be non monotonic, per haps even negative. This type of relationship implies that the negative marginal impact of additional potential competitors on incumbent profitability could decline as the number of potential entrants increases. Because of this possibility, a quadratic potential competition specification is also explored. Several researchers have also suggested that the impact of potential competition could vary with the intensity of actual market competition, and possibly with the two measures of market structure employed here to proxy this force. 1 6 In particular, a given number of potential competi tors could impose a larger impact on incumbent profitability if actual competition in the market were less intense. To investigate this possibility, actual competition/potential competition inter action variables are included in several versions of the performance equations estimated. Our study uses two summary measures of incumbent market structure: the three-firm deposit concentration ratio and the number of actual competitors. Two variants of each of these measures are employed. One is calculated using data for commercial banks only. The other is calculated using data for both banks and sav ings and loans, in recognition of the typically considerable thrift share of deposits in counties throughout Ohio and their expanding ability to compete with commercial banks. The number of holding company organiza tions legally permitted to branch de novo into each market is the measure of potential compe tition employed in this analysis. Available data revealed that holding company affiliates were responsible for most of the de novo branching activity in Ohio from 1979 to 1981. We exclude smaller banks that are unlikely to branch de novo in order to produce a more precise meas ure of potential competition . 1 7 ■ 15 See Kalish, Hartzog, and Cassidy (1978). Empirical evidence support ing this view appears in Hannan (1981) and Morrison and Winston (1987). potential competitor concentration would be extremely tedious, only the number of potential competitors is employed, ■ 16 Possible interactions between measures of actual and potential com petition are discussed in Hannan (1979), pp. 442-43, and in Morrison and Win ■ 1 3 The possibility of a nonlinear relationship between measures of ston (1987), p. 63. market structure and performance is noted in Heggestad (1979), pp. 468-69. ■ B 14 For a discussion of the expected relationship between concentration, 17 Examination of data on branching in Ohio over the 1979 to 1981 period revealed that holding company affiliates established 61 percent of the potential competition, and incumbent profitability, see Call and Keeler (1986), total number of de novo branches over this interval. Further, they established p. 224; Schwartz (1986), pp. 47-48; and Morrison and Winston (1987). 64 percent of those opened in contiguous counties. See Whalen (1981). Following the approach taken with the con centration variable, market share for each bank is defined in two different ways: by its share of commercial bank deposits in the market and by its share of bank and savings and loan deposits in the market. An insignificant coefficient on the incumbent market structure variable, in con junction with a positive, significant coefficient on the related market share term, is evidence supporting the efficient structure hypothesis. The risk measure used in this study is the same one used by a number of previous researchers: the standard deviation of return on equity over the period examined (1979 to 1981). There is some disagreement about the nature of the rela tionship between this variable and profitability. Heggestad (1979) and Clark ( 1986b) have argued that the relationship should be positive; Liang (1987) has suggested that it should be negative. 18 There is empirical evidence in support of both positions. Because of the uncertainty and because the precise nature of the relationship between these two variables is not the primary focus of this paper, the anticipated sign of the coefficient on the risk measure is left ambiguous. The other explanatory variables in the esti mated profitability equations are elements of the vector, These are presumably exogenous variables that reflect differences in the character istics of an individual bank, or economic condi tions in its market or its regulatory environment that could influence its profitability. Three bank characteristic variables are employed: a bank size measure, a dummy vari able measure of the number of branches oper ated, and a dummy variable indicating whether the bank was a subsidiary of a bank holding company. Economic conditions in each bank’s local market are represented by two variables: average per capita personal income and per cap ita personal income growth. Finally, we use a Federal Reserve System membership dummy to control for regulation-related cost differentials. To determine if the estimated relationship between actual competition, potential competi tion, and profitability is materially influenced by the neglect of possible simultaneity, the profita bility equation is also viewed as a structural equation in a multi-equation simultaneous sys tem. Specifically, a two-equation system similar to that used in Liang (1987) is employed. In this Z. ■ 18 In Liang’s model, greater profit variability implies greater expected costs and associated penalties to the bank, resulting in a negative relationship between profit variability and expected profit margins. system, bank risk is the other endogenous vari able. The main difference between her specifi cation and the one employed here is the addi tion of the potential competition term. Liang’s structural equation for risk contains five predetermined variables that do not appear in the profitability equation discussed above. These variables are designed to proxy market uncertainty. They are the standard deviation of market per capita personal income, unexplained market deposit supply, unexplained variation in bank loan demand, unexplained variation in bank deposit supply, and the covariance of bank unexplained loan demand and deposit supply. The precise definition of each of these variables and the reduced-form equations for this model are detailed in the appendix. Vs Vs Vs II. Sample and Methodology Our sample consists of the 159 single-market banks headquartered in non-MSA Ohio counties at the end of 1981. Single-market banks are those with all offices located within their home office county. This criterion allows their performance to be related to the characteristics of their particular local markets. The presumption is that non-MSA counties approximate local rural banking markets. The profitability equations are estimated using two different statistical techniques. Ordinary least squares regression (OLS) is used to estimate ver sions in which risk is viewed as exogenous. Two-stage least squares (2SLS) is the technique used to estimate the profitability equation when it is viewed as part of a simultaneous system. III. Results Regression results are presented in tables 1 and 2. Only the equations containing measures of actual market structure and market share calcu lated using commercial bank data are included in the tables. The results were essentially the same when savings and loans were considered in the calculation of these variables and there fore are not reported. Table 1 contains versions of the profitability equation estimated using OLS; table 2 contains abbreviated results obtained by estimating ver sions of the equations in table 1 viewed as part of a two-equation simultaneous model. The esti mation technique is 2SLS. Only the coefficients and t-statistics for the actual competition, poten tial competition, market share, and risk variables are reported. In general, the overall explanatory B U T A B L E 1 O L S Versions of Profitability Equations Dependent Variable: AROE Variables CBO (l) Coefficient (2 ) Coefficient -0.003253 (-0.17) -0 . 0 1 2 1 2 1 (-0.62) NCBO MSBO PCPIGR PCPI OD FRM MBHC SIZE SDROE HCPE (3) (4) Coefficient Coefficient 0.018071 ( 0 .1 2 ) -0.520431 (-1.80) 0.036970 ( 1-58) 0.043162 ( 1.85) 0.036723 ( 1-48) 0.035676 ( 1.45) 0.096151 ( 1 -2 6 ) 0.100644 0.094267 ( 1 .2 0 ) 0.090787 ( 1-17) 0.000213 ( 0.80) 0.000204 ( 0.77) -0.495045 (-0.76) ( 1.33) ( 0.78) 0.000175 ( 0 .6 6 ) -0.634474 (-0.98) -0.496847 (-0.76) -0.467602 (-0.72) -0.003253 (-0.17) -0.149174 (-0.28) -0 . 0 3 1 2 2 1 (- 0 .0 6 ) -0.095712 (-0.18) 2.183394 ( 3.14) 2.113173 ( 3 .0 6 ) 2.186579 ( 3.10) 2.271347 ( 3.26) -0.741219 (-1.49) -0.791260 (- 1 .6 1 ) -0.734965 (-1.46) -0.834644 (- 1 .6 8 ) -0.757202 (-8.07) -0.737778 (-7.91) -0.757808 (- 8 .0 6 ) -0.750641 (-8.08) -0.158573 (-1.29) -1.219721 (-2.34) -0.156887 (-1.28) -0.806247 (-2.46) HCPESQ 0 .0 0 0 2 10 0.114109 ( 2 .1 0 ) HCNCBO INT 0.112784 ( 2 .1 6 ) 14.110421 13.787394 ( 4.97) 17.385094 ( 5.42) (4.31) 16.894396 (4.83) F 7.46 7.34 7.41 7.33 RSQ 0.34 0.35 0.34 0 .3 6 NOTE: T-statistics are in parentheses. SOURCE: Author. power of the estimated equations is good, given the size and cross-sectional nature of the sample. The coefficients on the actual and potential competition and market share variables are of primary interest. The signs and statistical signifi cance of the other variables in the estimated equations are of secondary importance here and will not be discussed. The coefficient on the concentration variable is never even marginally significant in any ver sion of the equation estimated. 19 The results were invariant to specification and estimation techniques. Including savings and loans in the calculation of this variable and excluding the market share term did not alter this finding. When the number of actual competitors is used as the actual competition proxy, the results obtained do vary with the specification employed. The coefficient on the number of actual competi tors term is insignificant when a linear specifica tion is employed and when an actual competition/ potential competition interaction term is not included in the estimated equation. However, when an interaction term is included, the coeffi cient on the number of actual competitors varia ble becomes negative and significant. This result holds when savings and loans are included in this measure and when a simultaneous-equations estimation technique is employed. The coeffi cients are not significant when a quadratic ver sion is examined. The estimated coefficient on the number of potential competitors variable is negative, but only marginally significant (that is, 1 0 percent level, one-tail test) when a linear specification is employed and when an actual competition/po tential competition interaction term is not included. However, when this variable is used in an estimated equation in conjunction with the number of actual competitors and an interaction term, the coefficient is negative and significant. In these equations, the actual competition/ potential competition interaction term, con structed by multiplying the number of actual and potential competitors, exhibits a positive signifi cant coefficient. This finding supports the view that the negative marginal impact of additional actual competitors declines as the number of potential competitors increases. Similarly, the larger the number of actual competitors in a market, the smaller the negative marginal impact of additional potential competitors. When a quadratic potential competition speci fication is employed, the estimated coefficients on the number of potential competitors term and the square of this variable are both signifi cant. The pattern of signs (negative and positive, respectively) could reflect mutual awareness among potential entrants. This result suggests that the marginal impact of additional potential competitors is initially negative. ■ 19 A Herfindahl-Hirschmann Index of market concentration w as also employed in place of the three-firm concentration ratio. The change in the definition of the concentration ratio did not materially impact the results. However, the size of the negative impact declines as the number of potential competitors increases and finally turns positive. The magni tudes of the coefficients imply that incumbent firm profitability is constrained in markets with five or fewer potential entrants. This finding supports the notion of a nonlinear relationship between the number of potential entrants and the overall probability of entry. Changing the definition of the market struc ture and market share variables to include sav ings and loans did not alter either the size or the statistical significance of the coefficients on the potential competition variables in any of the specifications examined. Further, a comparison of each equation in table 1 with its counterpart in table 2 also demonstrates that the sign and statistical significance of the coefficients on the variables of interest in the estimated equations are not sensitive to the estimation technique used. 20 This was true for the other exogenous control variables as well. ■ T A B L E 2 Summary Results 2 S LS Versions of Profitability Equations Dependent Variable: AROE Variables CBO (l) Coefficient (2 ) (3) (4) Coefficient Coefficient Coefficient -0.001529 (-0.08) -0.010870 (-0.53) 0.007936 ( 0.05) -0.520134 (-1.79) NCBO MSBO SDROE A HCPE ( 1-52) 0.042391 ( 1-79) 0.035831 ( 1-43) 0.035265 ( 1-43) -0.857773 (-2.95) -0.810309 (-2.78) -0.870202 (- 3 -0 0 ) -0.803704 (-2.80) -0.159169 (-1.28) -1.186875 (-2.17) -0.158830 (-1.30) -0.801043 (-2.46) 0 .0 3 6 0 0 2 HCPESQ 0.110422 ( 1.93) HCNCBO 0.111721 ( 2.13) F RSQ 1.80 2.33 1.81 2 .1 1 0 .1 1 0.15 0 .11 0.14 NOTE: T-statistics are in parentheses. In general, the coefficient on the market share variable is positive and at least marginally signifi cant (at the 1 0 percent level, one-tail test) in every variant of the profitability equation esti mated. As with the concentration measure, somewhat stronger results are obtained when savings and loan deposits are considered in the construction of this variable. IV. Summary and Conclusions The results support the notion that non-MSA bank ing markets are contestable. That is, we found bank performance to be systematically related to proxies designed to measure the intensity of actual and potential competition. The threat of entry by potential competitors does appear to limit incumbent firm profitability, although the threat of entry and the number of potential competitors may not be monotonically related. Incorporating risk into the analysis and consider ing possible simultaneity between risk and prof itability did not materially alter the results. Both proxies for actual competition were not found to be consistently related to bank perfor mance, however. The concentration measure was not found to be significantly related to the profitability of banks operating in rural markets in Ohio in any specification investigated. Only the number of competitors proxy was found to be significantly related to bank profitability in the expected way. The finding that potential competition has a significant impact on incumbent performance is somewhat surprising for several reasons. First, potential competition is generally expected to be a weak force in rural banking markets. Second, researchers have argued that potential entrants may not significantly impact incumbent prices and profits in periods immediately after a change in regulations that affects entry conditions. The interval analyzed was just such a period. In addi tion, the potential entrant variable used in this study does not include potential nonbank com petitors, particularly savings and loans. Thus, the variable is obviously not a perfect proxy for the threat of entry in the markets examined. Further research on the impact of potential competition in banking markets appears war ranted to determine if the observed relationships SOURCE: Author. ■ 20 In addition, to further examine the sensitivity of the results to changes in specification, versions of the profitability equation similar to the one appearing in the four-equation model developed in Clark (1986b) were also estimated. The only change in Clark’s specification was the addition of the potential competition measures used in this study. Again, this change in speci fication did not materially alter the results reported above. are evident for other samples of banks and in other time periods. However, the results of this study suggest that it is unclear whether the con solidation taking place in banking in recent years has substantially lessened competition, given the simultaneous reductions in barriers to market entry that have occurred. For bank regulatory agencies, the results also imply that the competitive impacts of bank A P P E N D mergers cannot be reliably determined solely from a mechanical analysis of changes in actual market structure. Entry conditions and the exis tence of potential competition should also be considered and used to temper conclusions drawn from an analysis of merger-related changes in concentration or in the number of actual competitors. X Variable Definitions AROE: i Bank ’s annual after-tax return on equity, averaged over the 1979-1981 period. CBO: PCPI: Per capita personal income in the market averaged over the 1979-1981 interval. PCPIGR: Three-firm market concentration ratio, banks only, June 1980. Per capita personal income growth in the market over the 1979-1981 interval. NCBO: Number of banks operating in the market i, June 1980. The standard deviation of market per cap ita personal income over the 1979-1981 interval. of bank HCPE. Number of holding company organizations legally permitted to branch de novo into the market. HCPESQ. The square of HC HCPE. MSBO: SDROE: i ’s deposit market share, banks only, i Bank ’s standard deviation of annual after tax return on equity over the 1979-1981 period. SIZE: Log of total assets of bank i. OD: i has FRM: i Dummy variable equal to one if bank at least one branch, otherwise equal to zero. MDU: Market deposit uncertainty variable equal to proportion of unexplained variation in market depos its derived from the regression of market deposits on market income over the 1979-1981 interval. LRISK: HCPE --- Interaction term. times various alternative measures of market structure. Bank June 1980. SDPCPI: Dummy variable equal to one if bank was a member of the Federal Reserve System, otherwise equal to zero. DRISK: Deposit uncertainty variable for bank i equal to proportion of unexplained variation in total transactions deposits derived from the regres sion of total transactions deposits on market income over the 1979-1981 interval. COVLD: Covariance of unexplained loans and de posits for bank over the 1979-1981 period. SDROE: A MBHC: i i SDROE Predicted value for derived from the following first-stage regression with the relevant actual and potential competition variable(s) added: SDROE = f (MSBO, SIZE, OD, FRM, MBHC, PCPI, PCPIGR, SDPCPI, MDU, LRISK, DRISK, COVLD). A Dummy variable equal to one if bank is a holding company subsidiary, otherwise equal to zero. i Loan uncertainty variable for bank equal to proportion of unexplained variation in total loans derived from the regression of total loans on market income over the 1979-1981 interval. References Bain, Joe S. “A Note On Pricing in Monopoly and Oligopoly.” 39(March 1949): 448-64. American Economic Review. Baumol, W illiam J., John C. Panzar, and Robert D. Willig. New York: Harcoint, Brace and Jovanovich, 1982. Contestable Markets and the The ory of Industry’ Structure. Berger, Allen N., and Timothy Hannan. “The Price-Concentration Relationship in Bank ing.” In Merging Commercial and Investment Banking: Proceedings from a Conference on Bank Structure and Competition. Federal Reserve Bank of Chicago (May 1987): 538-9. Berger, Allen N., Gerald A. Hanweck, and David B. Humphrey. “Competitive Viability in Bank ing: Scale, Scope, and Product Mix Econo mies.” Board of Governors of the Federal Reserve System. February 1986. Research Papers in Banking and Financial Economics, Boyle, Stanley E., Richard K Ford, and Charles G. Martin. “The Application of Industrial Organization Theory to Commercial Banking: A Review and Analysis.” University of Arkansas-Little Rock, Department of Econom ics and Finance. 1987. Working Paper, Brozen, Yale. Concentration, Mergers, and Pub lic Policy. New York: Macmillan Publishing Company, Inc., 1982. Burke, J., and S. Rhoades. “Profits and ‘Contest ability’ in Highly Concentrated Banking Markets.” (Fall 1987). Review of Industrial Organization. Call, C., and T. Keeler. “Airline Deregulation, Fares, and Market Behavior: Some Empirical Evidence.” In A. Daughety, ed., Cambridge, Mass.: Cambridge University Press, 1986. Analytical Studies in Transport Economics. Clark, Jeffrey A. “Single-Equation, MultipleRegression Methodology: Is It an Appropriate Methodology for the Estimation of the Structure-Performance Relationship in Bank ing?” 18, no. 3(November 1986a): 295-312. Journal of Monetary Economics. _________ _ “Market Structure, Risk, and Profita bility: The Quiet Life Hypothesis Revisited.” Quarterly Review of Economics and Busi ness. 26, no. l(Spring 1986b): 45-56. Evanoff, Douglas D., and Diana L. Fortier. “Re evaluation of the Structure-Conduct-Performance Paradigm in Banking.” 1, no. 3(June 1988): 277-94. Financial Services Research. Journal of Gilbert, R. “Bank Market Structure and Competi tion.” 16 , no. 4(November 1984). ing. Journal oj Money, Credit, and Bank Graddy, Duane B., and Reuben Kyle III. “The Simultaneity of Bank Decision-Making, Market Structure, and Bank Performance.” 34, no. l(March 1979): 1-18. Journal of Finance. Graham, David R., Daniel P. Kaplan, and David S. Sibley. “Efficiency and Competition in the Airline Industry.” 14, no. l(Spring 1983): 118-38. BellJournal of Econom ics. Hannan, Timothy H. “Mutual Awareness Among Potential Entrants: An Empirical Examina tion.” 47, no. 3(January 1981): 805-8. Southern Economic Journal. __________ “Limit Pricing and the Banking Industry7.” 11, no. 4(November 1979): 438-46. Banking. Journal oj Money, Credit, and Heggestad, Arnold A. “Market Structure, Compe tition, and Performance in Financial Indus tries: A Survey of Banking Studies.” In Frank lin R. Edwards, ed., New York: McGraw Hill Book Company, 1979. Regulation. Issues in Financial _________ _ “Market Structure, Risk, and Profita bility in Commercial Banking.” 32, no. 4(September 1977): 1207-16. Finance. Journal of Kalish, Lionel, Jerry Hartzog, and Henry Cas sidy. “The Threat of Entry With Mutually Aware Potential Entrants: Comment 8 6 , no. 1( February 1978): 147-50. oj Political Economy. ."Journal Liang, Nellie. “Bank Profitability and Risk.” Unpublished paper, Board of Governors of the Federal Reserve System, November 1987. Morrison, Steven A., and Clifford Winston. “Empirical Implications and Tests of the Con testability Hypothesis.” 30(April 1987): 53-66. Economics. Journal of Law and Osborne, Dale K , and Jeanne Wendel. “Research on Structure, Conduct, and Per formance in Banking, 1964-1979 ” Oklahoma State University, College of Business Adminis tration. 83-8, July 1983- Working Paper Rhoades, Stephen A. “Structure Performance Studies in Banking: An Updated Summary and Evaluation.” Board of Governors of the Federal Reserve System, Staff Study No. 119, August 1982. --------------, and Roger D. Rutz. “Market Power and Firm Risk: A Test of the ‘Quiet Life’ Hypothesis.” 9(Januarv 1982): 73-86. ics. Journal of Monetary Econom Schwartz, Marius. “The Nature and Scope of Contestability Theory.” 38, Supplement (November 1986): 37-57. Papers. Oxford Economic Smirlock, Michael. “Evidence on the (Non)Relationship Between Concentration and Profit ability in Banking.” 17, no. l(February 1985). and Banking. Journal of Money’, Credit, Whalen, Gary. “Concentration and Profitability in Non-MSA Banking Markets.” Federal Reserve Bank of Cleveland. (Quarter 1, 1987): 2-14. Review Economic ----- - “Bank Expansion in Ohio.” Federal Reserve Bank of Cleveland. April 6 , 1981. mentary, Economic Com G e ttin g th e N o is e O u t: F ilte r in g E a rly G N P E s tim a te s by John Scadding John Scadding was a visiting scholar at the Federal Reserve Bank of Cleveland when he wrote this paper. Currently, he is an economist with the California Public Utilities Com mission. The author would like to thank Stephen M cNees for helpful comments. Introduction Real, or inflation-adjusted, gross national product (GNP) is the most inclusive measure of the nation’s economic activity. As such, it is probably the most closely monitored economic barometer for the information it contains about the eco nomic well-being of the economy and about the economy’s prospects. It is the central focus of most macroeconomic models and their forecasts, and it plays a decisive role in shaping monetary and fiscal policy decisions. Given the critical role that GNP plays, it is not surprising that the accuracy of GNP estimates is crucial if informed decisions are to be made by both private agents and government policymak ers. There is a trade-off, however, between the estimates’ accuracy and their Delays in reporting and revising data as more inclusive information becomes available means later esti mates will typically be more accurate than earlier ones; but waiting longer entails forgoing the opportunity to take action sooner, when that may be a critical factor. In the United States, the first offical estimate for a particular quarter’s GNP is released by the U. S. Department of Commerce approximately three weeks that quarter has ended. Much of the data needed to construct GNP are still not available at that point, even though the quarter timeliness. after has ended. The missing data therefore must be estimated by the U.S. Department of Commerce’s Bureau of Economic Analysis (BEA), which is responsible for compiling the official estimate of GNP. This first estimate is followed in relatively rapid succession by two additional estimates, one and two months after the initial number is released. Thereafter, the delays in revisions become much longer. Estimates are usually sub ject to three further annual revisions. After that, an estimate is usually subject to further so-called benchmark revisions every five years as data from the Bureau of Census’ quinquennial eco nomic census are incorporated. At each stage, source data are incorporated that had not been available previously, and revisions to previous data are incorporated as well . 1 It is clear from this description that there is never a final estimate of GNP that could be equated with the “truth.” Nevertheless, the three estimates are early preliminary or obviously distinct from the later ones in terms of their timeliness. Although based on incomplete and preliminary information, the provisional esti mates have the advantage that they are available provisional ■ 1 Carson (1987) provides a comprehensive overview of the source data and estimation methods for constructing the different G N P estimates. See also Young (1987). m much sooner than the later, more comprehensive, and presumably more accurate numbers. It is rele vant, therefore, to examine their accuracy in pre dicting the later numbers. As Allan Young, direc tor of the Bureau of Economic Analysis, noted in a recent comprehensive survey of the properties of GNP estimates: “Much of the concern with the reliability of GNP comes down to whether the early . . . quarterly estimates . . . provide a useful indicator of the estimates . . . When complete and final source data are available.” (Young [1987], p. 18) T A B L E 1 that subsequent revisions to the GNP numbers would be correlated with the preliminary esti mates. This alternative view implies that the early GNP numbers are estimates of the final number, but estimates that are contaminated with error. If this alternative view is correct, then it is pos sible in principle to make estimates of the error in the preliminary numbers and to adjust the lat ter to remove the error— in other words, to filter out the “noise.” This paper investigates one method of doing this. The results suggest there is scope for adjusting the provisional GNP growth rate numbers to make them better predictors of what the final numbers will turn out to be. Final Revisions to Real G N P Growth and Components, 19 74 :110 - 19 8 4 :10 I. The Data final revisions Final Revisions Estimated Observation Error Residual Forecast Error Mean Variance Mean Variance Mean Varianct 0 .6 3 0 4.087 -0 . 6 3 0 0.764 0 .0 0 3.323 Final minus 45-day 0.413 2.876 -0.413 0.694 0 .0 0 2.182 Final minus 75-day 0.205 2.742 -0.205 0 .8 9 0 0 .0 0 1.852 Final minus 15-day SOURCE: Author. One important strand of the literature examin ing this question has concluded that the early numbers can be viewed as forecasts of the actual numbers. The term is used in the sense that the differences between a final GNP growth number and its corresponding pre liminary estimates are uncorrelated with the pre liminary numbers themselves (Mankiw and Sha piro [1986]; Walsh [1985]). On the face of it, this is a surprising result. It denies the intuitively appealing, and perhaps prevalent, view that if a preliminary estimate showed large positive growth for real GNP in a quarter, for example, it would be more likely than not that later esti mates would be revised down— in other words, that the final GNP number would be smaller than its preliminary estimate. And, similarly, a large (in absolute value) negative preliminary7 estimate would be revised upward subsequently. In a preliminary analysis reexamining this ques tion, Scadding (1987) concluded that the statisti cal test used in the analyses mentioned above could not discriminate very well between the ra tional forecasts hypothesis and the alternative view rational rational Table 1 has estimates of the for real GNP growth— that is, the difference between the final estimate and the three provisional esti mates. There are three final revisions, corres ponding to the difference between the final numbers and each of the three provisional numbers. For the sample period used in this paper (1974-1984), the early estimates came out 15 days, 45 days, and 75 days after the quarter ended, and the usual nomenclature is to refer to them as the 15-day estimate, and so on. Corres pondingly, there is the 15-day final revision, which is the difference between the final number and the 15-day estimate, and so on. I follow the usual practice and define the “final” number as the currently available final number as of the quarter in question. Thus, final estimates in the earlier part of the sample will have been through more revisions than those later in the sample . 2 For the 15-day estimate of GNP, many of the source data are not complete and are subject to revision. The data available for this estimate are monthly data, like retail sales, manufacturers’ shipments of machinery and equipment, and merchandise trade figures. Some of these data, like retail sales, are based on surveys, and typi cally are revised substantially. In addition, some of the monthly source data are not available for all three months of the quarter. For example, only one to two months of data are available for estimating consumer spending on services, which is about one-half of total consumer spend ing. And there are no monthly data at all for ■ 2 The data are from a study prepared by the Bureau of Economic Analy sis and are the data used by M ankiw and Shapiro (1986), Mork (1987), and Walsh (1985). The data were adjusted to abstract from the effects of defini tional changes and the change in the base year for calculating constant-dollar G N P . See Young (1987), p. 2 5 .1 am indebted to Professor Mork for providing me with a copy of these data. about 40 percent of spending on services. This component, therefore, is estimated by the Department of Commerce, either by extrapolat ing by related series or by judgmental projection. The succeeding 45- and 75-day estimates incor porate new monthly data unavailable for the 15day estimate, and as well incorporate revisions to the monthly data that were included in the 15-day number. As well, these two estimates include new information available only on a quarterly basis— domestic corporate profits, balance of payments figures, and data on financial assets from the Federal Reserve Board’s flow of funds accounts. The latter two sources are incorporated in the 75-day estimate only (Carson [1987], p. 107). As table 1 shows, the final revisions are not trivial. On average for the sample they are posi tive, suggesting a systematic tendency for the preliminary numbers to understate the final estimates, a phenomenon that has been noted elsewhere (Mork [1987] ). The deviations implied by the sample variance estimates reported in table 1 are large when measured against the the mean growth of real GNP for the period, which was 2.9 percent. Thus, plus or minus one standard error about a preliminary’ estimate equal to this trend growth translates into an economy that, with equal probability, could be enjoying near boom-like conditions or behaving as if it was close to recession. II. The Nature of the Provisional GN P Estimates As discussed briefly in the introduction, one pos sible way of thinking of the early GNP growth of what the final esti numbers is as mate will turn out to be. Thus, suppose * is the final estimate of GNP growth for quarter that estimate of course will not be made until some time after quarter In the meantime, however, a provisional estimate (in fact three), will be available soon after quarter call it has ended. This provisional estimate can be thought of as a forecast of what will be. From that perspective, it is natural to ask whether is a forecast in the sense that, at a m ini mum, it is unbiased and is uncorrelated with the forecast error, which is equal to the final revi sion, If this description fits then forecasts X t; t. Xt, X* Xt Xt t good X* X* (1) zt Xt . = Xt, Xt + z t, where is a zero-mean, serially uncorrelated forecast error (white noise) that is uncorrelated with t . X Walsh (1985) defines these to be the proper ties of a rational forecast. The competing charac terization of is that it is an early observation or “reading” of what will be, but an obser vation measured with error. Thus, Xt (2) X* X,= X* + u, ut X* where is also white noise, and uncorrelated with in this case. Note that this characteriza tion implies that the final revision correlated with the provisional estimate; in other words: (3) is E(X* - X ,) Xt = -E u) = -°z> where ctJ is the variance of the observation error The evidence on which characterization better describes the nature of the provisional estimates is decidedly mixed. Mankiw and Shapiro (1986) adduce evidence in favor of the position that preliminary numbers are rational forecasts, on the criteria just described. However, I have argued in a technical companion piece to this paper (Scadding [1987]) that their test is likely to have little power. They themselves raise this possibil ity because of the apparent contradiction of their conclusion with evidence elsewhere that two important data sources for the GNP estimates— retail sales and inventories— have significant measurement errors in them (Howrey [1984] and Conrad and Corrado [1979]). Walsh, using a slightly different sample from Mankiw and Shapiro, finds corroborating evi dence for their result, but this conclusion is compromised by his additional finding that the provisional estimates are inefficient forecasts. In addition, Mork, using different estimation tech niques from the other studies, found evidence that the provisional estimates were biased downwards, and that the final revisions were correlated with previous-quarter GNP growth and a forecast of GNP growth from a publicly available survey of private forecasters. u. III. Filtering the Early Data I have argued elsewhere (Scadding [1987]) that Walsh’s evidence of inefficient forecasting is equally compatible with the view that provi sional GNP numbers are observations rather than forecasts, with the observation errors in the three provisional numbers being sequentially corre lated. Howrey (1984) found this to be a useful characterization of the inventory investment component of GNP. In my earlier paper, I m devised a test for discriminating between an inef ficient forecasts model and a serially correlated measurement error model based on restrictions on the variance-covariance matrix of the final revisions. The results of that test suggest that the provisional estimates of real GNP growth contain measurement error. The purpose of this paper is to estimate the amount of observation (measurement) error in the provisional GNP growth numbers and sub tract that error to obtain modified, or filtered, provisional GNP estimates that have the proper ties of a rational forecast. Let X be the filtered estimate; then the estimated measurement and forecast errors are defined by * (4a) ut - (4b) z t = X* - X*. Xt - X* and interim The definitions (4a and 4b) implicitly define the decomposition of a final revision, X* - X t , into its measurement and forecast error component: (5) X* - X t = z t - ut . Nonrecursive Kalman filtering, described below, is used to specify equations for estimat ing X * . Least-scjuares estimation of these equa tions yields an X series with the desired fore casting properties: * ( 6 a) E(X* ( 6 b) E(X* - - X*) = 0 and X*)X* = 0. As well, the estimated measurement and fore cast errors are orthogonal to each other: ( 6 c) E(m, z t) about the size of the final revision in the current quarter. The time-series forecast presumably is picking up this information. In addition, filtering improves the precision of forecasting by exploit ing the fact that part of the final revision is meas urement error and therefore can be forecast from the provisional estimates. Note the uniformly negative means of the estimated observation errors, indicating a syste matic tendency of the provisional GNP estimates to underpredict the final numbers. This tendency has been noted by Mork, who ascribes it to con cern by the Department of Commerce that the provisional estimates not be seen as being too optimistic and therefore serving some political agenda. The presence of serially correlated measure ment errors makes it relatively easy to predict revisions— in other words, from the 15day to the 45-day, and so on— compared to final revisions. As we shall see, the standard errors of the regression predicting the provisional esti mates are about 50 percent lower than the standard errors of the equations predicting the final GNP estimates. Thus, the methodology out lined here provides forecasters with a relatively accurate way of forecasting subsequent prelimi nary estimates. More generally, this result sug gests that the provisional estimates are more like each other than they are like the later estimates, a point that has been made by McNees (1986). Many economists presumably would be offended by the notion that any attention should be paid to forecasting the provisional estimates themselves when what obviously matters is get ting a good estimate o f the final or “true” num ber. However, that is “obvious” only to the extent the Federal Reserve or private agents, in reacting to new provisional estimates, discount the measurement error in them, an assumption that is not obvious on its face at least. It is cus tomary to test market forecasts o f GNP by their ability to predict final GNP; it would be interest ing to inquire whether they do a better job o f forecasting provisional GNP estimates. A final observation suggested by this paper’s result is that the frequent practice by forecasters of discarding their GNP growth forecast for a quarter when the first provisional estimate for that quarter becomes available probably is not efficient. The filtering technique used in this paper combines the provisional estimates of GNP growth with a forecast from a simple timeseries model. The results suggest that the fore cast still has information about final GNP growth even after the preliminary estimates become available. As McNees has noted: “...the distinc tion between forecasts and ‘actual’ data is often = 0. Summary statistics for the final revisions and the estimated measurement and forecast errors are shown in table 1. Clearly, the filtering im proves the forecasting precision of provisional numbers. The sample variance of the forecasting error after filtering is on the order of 25 to 30 percent lower than the variance of the unfiltered final revision. Nevertheless, the residual forecast variance is still quite large. The improvement in forecasting precision would appear to be based on two factors. First, the filtered estimates are derived by combining the provisional estimates with a simple timeseries forecast of GNP growth. Mork has noted that the prior quarter’s GNP has information exaggerated. Both are estimates based on partial, incomplete information.” (McNees [1986], p. 3) (12a) X tn = X * + (12b) X* = <t>X *_, + w„ , where 0 is a fixed parameter and w is a ran dom, serially uncorrelated term with zero mean and constant variance (white noise). We cannot observe X* directly but have mea surements of it, X, that involve error (here X would be a provisional estimate of GNP growth). X t = hX * + u t, where h is a fixed parameter and u is also white noise. The Kalman filter optimally weights the fore cast of X* from equation (7) with the observa tion to form the best linear unbiased predictor of X * , called the filtered value: (9) X* = X t + Kt( X , ____ A (10a) X , +1 = <I>X* (10b )x ; +1 = a Xt/2, Xt/ 2 - (13c) ; /2 X't/l - x% - X ,/0), -^ 2 3 (1 ^ X t /i ], and - a n )X*t/2 ). X t/0 The initial forecast is taken from a simple time-series model for real GNP growth, Given the forecast and estimates of the and one could then calculate the filtered esti mates directly. The approach taken here, how ever, is to estimate the using ordinary least squares to produce a set of estimated measure ment errors and residual forecast errors that are uncorrelated. 3 Thus, the estimation equations corresponding to ( 1 3 ) are X*. K's as, K's X* = (14b) ^ 1(^ /1 ~ * ,/ o ) + z if *t/2 ~ a 2i Xf/\ + a 23^X*/l + x r = A■y*//i + K2v 2t + *2 /. ( 1 A (14c) (14d) and X t/5 = a 12 X,/2 + ~a \2^X*/2 + ( 1 - for n 1,2, 3, where n 1 refers to the 15-day estimate, n 2 to the 45-day estimate, and n 3 to the 75-day estimate. Similarly, - t// x - ( 1 X*/2 + - a n X t/2) - = X t/o + = - a = K 3 [(X , / 3 (14a) Two modifications are necessary to apply this algorithm to the program at hand. First, the three provisional estimates of GNP growth are repeated observations on the same final estimate. Thus, within the quarter, the law of motion is K,(X,n x]n + K-zKXf/i - a 2^X (i3b )i? X X,/\, (13a) X * 7, = X , / 0 + X t+1 + A^+|(X ?+1 - h X /+1). x : /n + « 3„ where the are the provisional estimates and the u n,'t s are the corresponding measure ment errors. Thus, within the quarter, the </> in equation (7) is unity, as is h in ( 8 ), while the intraquarter w ’ s are uniformly zero. The other modification follows from the fact that preliminary estimation suggests that the u ’ s (12) are sequentially correlated. This serial correlation structure is shown in table 2 . The filtering framework is easily adapted to this circumstance by expressing the observation variables in quasi-difference form, / / 3 12 « 2 3 where the ’s are the respective serial correlation coefficients of the errors from table 2 (see Bryson and Ho [1969], pp. 400-405). The modified set of filtering equa tions becomes - h X t), where X is the forecast and X* is the filtered value. The weighting coefficient, K, is called the Kalman gain, and is a function of the variances of w, u, and of h. The filtered value is used to update the forecast. Using (7), this new forecast is combined with the next observation to calcu late the next filtered value. d o + u2t, and X t/n's The general idea of filtering data is easily sketched out. Suppose the variable we are inter ested in, X * (which in our case is the final esti mate of GNP growth), evolves over time accord ing to the law of motion (8 ) X* (12c) X t/5 = IV. The Filtering Framework (7) X t/2 = X* u \t> ■ 3 Conrad and Conrado (1979) and Howrey (1984) have used the Kalman framework for analyzing retail sales and inventory investment data, respectively. (I4 e ) X * = X */ 2 + K} v3, + z 5t. + = * 1 (\4b’)X*t/2 o (1 4 a ')* ;/! H To complete (14) we append the set of defini tions of the filtered estimates of GNP growth: x*n (14c ') X ? /3 = K n — T A B L + + K ,(Xt/x - X t/0), K2v2n K 5v 5t E and ■ 2 on the 15-day estimate and the first filtered esti mate of GNP growth. The innovation in the 45-day number then is used to update the filtered estimate of final GNP growth by regressing the final GNP number on the first filtered estimate and the measurement innovation in the 45-day number (equation 14c). The residual z 2t provides an estimate of the forecast error in the 45-day number. The same sequence of estimations is performed to calcu late the new filtered estimate of final GNP condi tional on having the 75-day provisional GNP estimate, and its corresponding forecast error. Correlation Structure of Final1Revisions V. Estimation Results Standard Error of Estimate X* - Xt/} = X* - Xt/2 = (X* - Xt/2) + v3t 0.503 (X* - Xt/1) + v2t 0.503 0.230 + 0.932 (2.79) (19.61) 0.077 + 0.784 (0.76) (16.36) X* - X t/ 1 =-0.698 + vu (-2.18) 2.02 SOURCE: Author. The estimation of (14) proceeds sequentially. First ( 14a) is estimated, by regressing final GNP growth on the time-series forecast, and the 15-day provisional estimate, . The residual, , is the forecast error for the filtered 15-day estimate. The first filtered estimate of final GNP growth, is calculated using (14a')- The cor responding measurement error in the 15-day provisional GNP growth rate is estimated as Xf/1 Zj Xt/0, X*/v (15) u u = X ,n - X*n = ( 1 - * , ) ( X t/, - X , /0), X t/2 which by construction is uncorrelated with the forecast error. The next step is to calculate the innovation in the measurement error in the 45-day provisional GNP number. The correlation structure between the measurement errors in the 15-day and 45-day number is (16) X t/2 - X* = « 23 ( X t/] - X * ) + v2t, where v2t is the innovation in the measurement error. Rearranging (16) and substituting for yields (14b), which is then estimated by regressing the 45-day provisional GNP number X* The results of estimating equations (I4a)-(l4e) are shown in table 3- Almost uniformly, with one important exception discussed later, the esti mated coefficients in table 3 are statistically dif ferent from zero at the 95 percent confidence level. Perhaps more importantly, again with the same exception just noted, the restrictions implied by equations (14) are all met. Thus, for example, equation (14a) implies that the sum of coefficients on the time-series forecast and the 15-day provisional estimate sum to one. In other words, the 15-day filtered estimate of real GNP growth is a simple weighted average of the fore cast and 15-day GNP number. The last column reports the F-test statistic, and it clearly cannot reject the hypothesis of the 9 5 percent confi dence level that the coefficients sum to unity. Similarly, the restrictions in equations (14b) and (I4 d ) that the coefficients sum to unity and that the coefficients on the lagged dependent variables equal the estimated correlation coeffi cients from table 2 are also met. In>Athe case of , equation (14d) the coefficient on X* 2 was not itself statistically significant, even though the joint hypothesis could not be rejected. When the coefficient on was constrained to be 0 .9 3 2 2 — its a priori value as indicated by table 2— the coefficient on X* became significant, which is the result reported in (I4d). Only equation ( I4e) gave any significant troub le. In this case, the estimated K} was not signifi cantly different from zero, indicating that the 75day estimate did not have any additional informa tion about the final GNP number that was not already contained in the two preceding provi sional estimates and the time-series forecast. This last result stands in sharp contrast to the information provided by the first two provisional GNP numbers about final GNP. The estimated Kalman gain Kx and K2 in (14a) and (14c) are X*/x 2 Estimated Filtering Equations 19 74 :IIQ - 19 84:10 (T-statistics in parentheses) (14a) = X t/0 6 X t/1 (14b) (14c) 0.336 + 0.291 (0.89) ( 2 .6 1 ) -0.047 + 0.81 (-.29) (7.33) = X tn -0.077 + 1.020 (-0.28) (19.61) + zu 1.823 1.4793 X*n + v2t 0.714 2.247b 1.477 2.378c 0.560 0.578d 1 .3 6 1 0.339e + 0.236 (1.99) + 1.483 (4.30) 0.235 + 0.932X / / 2 + 0.098 (2.54) (5.95) (I4 e ) 0.083 + 0.974 (0.04) (21.86) v2t + z 2t X*/2 X t/2 + 0.784 v)t X t_j + wt - + + v3t z )t (1.77) Addendum: time-series forecasting model X*= F-Statistic 0.774Xt/1 + (8.45) (I4 d ) * * = = Standard Error o f Estimate 0.511 + 0.828 0.415 (2.19) (7.87) (2.40) Standard error of estimate = 3-323 wt_x a. Test that sum o f coefficients is unity. b. Test that coefficient on X = 0.0784 and that sum o f coefficients is unity. c. Test that coefficient on X t^2 is unity and that K-, = 1.234. d. Test that coefficient on X t^2 is 0.932 and that coefficients sum to unity. The equation was reestimated with coefficient on X ^ 2 re stricted to 0.932; the results are the ones reported in (l4 d ). e. Test that coefficient on X * ,2 is unity: the restriction that the coefficient was unity and that the coefficient on v$t was n on zero was rejected at the 5 percent confid ence level. SOURCE: Author. 0.774 and 1.483, respectively, and both are statis tically different from zero at the 95-percent con fidence level. The latter number may seem too high—presumably it should be between zero and one. However, with serial correlation in the measurement errors the constraint is that must be less than one. This constraint is satisfied by the calculated theoretical Kalman gains shown in table 4 . 4 Clearly, the estimated of 1.483 is not statistically different from the theo retical value of 1.234, given the size of its stan dard error, a conclusion that is substantiated by the results of the F-test in table 3. The nonzero coefficient on the time-series fore cast variable, t/0, suggests that the provisional estimates do not fully incorporate information about final GNP contained in the previous quar- a 25 K2 K2 X ■ 4 The similarity of the "theoretical" and estimated Kalman gains suggest ter’s estimates. 5 This suggests perhaps a tendency on the part of the BEA to be conservative in extrapolating trends in the GNP data. And it also suggests that the typical practice in forecasting and policy analysis to discard forecasts for the immediately prior quarter once the provisional estimates become available may be inefficient. The fact that the 75-day estimate does not appear to add any additional information about final GNP is interesting given that it is the first estimate to incorporate quarterly data. The high degree of serial correlation between the 45-day and 75-day provisional estimates shown in (I4 d ), with relatively low variance in the residual, indi cates, however, that the two estimates are not very different from each other despite the addition of the quarterly information. Indeed, to an impor tant extent this is true of all three provisional esti mates: they provide more information about each other than they do about the final GNP number. there would be no advantage from calculating the filtered estimator using the theoretical numbers. There does not appear to be any clear consensus whether regression-based weighting of forecasts is preferable to sample- ■ 5 estimated optimal-weighting. See, for example, Lupoletti and Webb (1986), pp. new provisional estimator first became available, not past values of the final 279-281. G N P growth estimator. The time-series forecasts used only past data available or the time the References Calculated and Estimated Kalman Gains Optimal Control. Calculated Estimated Standard Error of Estimated Gains 0.719 0.774 (0.092) 1.234 1.483 (0.345) 0.679 0.784 (0.443) Kx K2 K} Applied Bryson, Arthur Earl, and Yu-Chi Ho. Waltham, Mass.: Blaisdell Publishing Co., 1969. SOURCE: Author. VI. Conclusion A recent and interesting analysis of the early GNP estimates has concluded that “they behave neither as efficient forecasts nor as observations measured with error” (Mork [1987], p. 173)- The purpose of this paper has been to filter the early GNP numbers, to remove the measurement error, and to produce more accurate predictions of the final GNP growth estimates. In a related paper (Scadding [1987]), I have shown that these fil tered estimates do not exhibit the unconditional bias and inefficiency that Mork found for the raw estimates. Another interesting sidelight of the results of this paper is that the Mankiw-Shapiro test for discriminating between observation and forecast errors does a poor job when applied to the estimated observation and forecast errors calculated in this paper, corroborating other indications of the poor power of the test. For the forecaster, the filtering approach out lined in this paper provides an easy and syste matic way of adjusting the provisional numbers to make them better estimates of “actual” GNP growth. It would be intriguing to inquire whether forecasters do in fact adjust the early numbers in a way that is consistent with the approach taken here. The estimation results reported are model specific in the sense that they depend, to an unknown extent, on the specific forecasting model used to initialize the filtering procedure. Again, it would be interesting to see the extent to which the filtering results were sensitive to the forecasting model by using forecasts from alternative models. One offshoot of such an exercise would be that a particular model’s per formance could be evaluated in terms of the extent to which its forecasts contributed to improving the forecasting ability of the prelimi nary GNP numbers. Carson, Carol S. “GNP: An Overview of Source Data and Estimating Methods.” 67, no. 7 (July 1987): 103-26. Survey of Cur rent Business. Conrad, William, and Carol Corrado. “Applica tion of the Kalman Filter to Revisions in Monthly Retail Sales Estimates.” 1, no. 2 (May 1979): 177-98. Journal of Economic Dynamics and Control. Howrey, E. Philip. “Data Revision, Reconstruc tion and Prediction: An Application to Inven tory Investment.” 6 6 , no. 3 (August 1984): 386-93- Statistics. Review oj Economics and Lupoletti, William M., and Roy H. W ebb. “Defining and Improving the Accuracy of Macroeco nomic Forecasts: Contributions from a VAR Model.” 59 (April 1986) part 1 : 263-85. Journal of Business. Mankiw, N. Gregory, and Matthew D. Shapiro. “News or Noise? An Analysis of GNP Revi sions.” 6 6 , no. 5 (May 1986): 20-25. Survey of Current Business. McNees, Stephen K. “Estimating GNP: The Trade-off between Timeliness and Accuracy.” Jan/Feb. 1986: 3 -1 0 . New England Economic Review. Mork, Knut A. “‘Ain’t Behaving’: Forecast Errors and Measurement Errors in Early GNP Esti mates.” 5, no. 2 (April 1987): 165-75. Journal of Business and Economics Statistics. Scadding, John L. “The Nature of GNP Revi sions.” Federal Reserve Bank of Cleveland. No. 8718, December 1987. Working Paper Walsh, C.B. “Revisions in the ‘Flash’ Estimate of GNP Growth: Measurement Error or Forecast Error?” Federal Reserve Bank of San Francisco. ( Fall 1985): 5-13- Economic Review Young, Allan H. “Evaluation of the GNP Esti mates.” 67, no. (August 1987): 18-42. Survey of Current Business. 8 C o m m e n t— In te r v e n tio n a n d th e D o lla r ’s D e c lin e by Owen F. Humpage Owen F . Humpage is an economic advisor at the Federal Reserve Bank of Cleveland. This comment con cerns an article he wrote for the preceding issue of Review After publication of “Intervention and the Dol lar’s Decline” in the preceding issue of some confusion arose regarding exactly when the exchange-rate quotes in that article were taken and from what market they were derived. This comment will explain the dif ferences and respecify some of the equations to dispel any misinterpretation. The daily data for the article were taken from DRI-FACS in August 1987. We understood from reading the DRI-FACS manual that the data series from August 7, 1984 to August 28, 1987 were morning opening exchange-rate quotes from the New York market. The recently revised DRI-FACS manual (now called DRIFACS PLUS) indicates that after October 8 , 1986, the data refer to closing quotes in the London market. 1 We therefore reesti mated the equations in tables 3 and 4 of the arti cle to determine if this change had any signifi cant effect on the results. While some of the point estimates are slightly different under these new estimations, the over all conclusion of the article remains the same: nomic Review, ■ 1 Eco D R IF A C S P L U S , the Dictionary of M oney Markets and Fixed Income Data, Data Resources, Inc., February 1988. Data prior to October 8, 1986 are as originally reported. Economic (Quarter 2 1988), pp. 2-16. Between August 1984 and August 1987, day-today U.S. intervention did not systematically affect day-to-day exchange-rate movements. However, on some occasions, intervention did have a tem porary effect on mark-dollar and/or yen-dollar exchange rates. Statistical tests in the article included U.S. intervention with a one-day lag to avoid prob lems with bidirectional causality between exchange rates and intervention. Generally, the results are interpreted on the assumption that the effects of U.S. intervention on day occurred between the opening quote on day and the opening quote on day After October 8 , 1986, however, the data are closing quotes from the London market. Since the New York market opened before the London market closed, U.S. intervention on day could have affected the London closing exchange-rate quote on day and on day To allow for this possibility, we reestimated the relevant equations, including a contempo raneous intervention term. Tables 3A and 4A, which correspond to tables 3 and 4 of the origi nal article, present the results. t-l t. t-l t. t-l t-l T A B L E 6 7 Intervention I. Estimation Period: February 23, 1987 to July 2, 1987 A. Dependent Variable: mark-dollar exchange rate Independent Variables Coefficient T-statistic Intervention dummies Initial purchases no lag lagged Subsequent purchases no lag lagged Initial sales no lag lagged Subsequent sales no lag lagged Lagged dependent (1 ) (1 ) (0 ) (0 ) (3) (3) (2 ) (2 ) 0.009 -0.007 — — -0.007 -0 . 0 0 6 -0 . 0 0 6 -0.008 1.00 1.73a -1.35b — — -2 .3 8 c -2 .0 6 c -1.14 -1.56 994.8d Sum of Squared Residuals = 0.002 R2 = 0.824 n = 90 B. Dependent Variable: yen-dollar exchange rate Independent Variables Intervention dummies Initial purchases no lag lagged Subsequent purchases no lag lagged Initial sales no lag lagged Subsequent sales no lag lagged gged dependent Coefficient T-statistic — (0 ) (0 ) — (0 ) — — (0 ) -0 . 0 1 1 (2 ) -0 . 0 0 1 (2 ) ( 1 6 ) -0.007 (1 6 ) 0.0005 1.0 0 0 — — — — -1.893 -0 . 2 1 -3.08d 0 .2 1 70l6.4d Sum of Squared Residuals = 0.003 R2 = 0.969 n = 90 NOTE: Intervention refers to U.S. purchases or sales o f foreign currencies. Numbers in parentheses indicate the number o f times the dum m y equals 1. a. Significant at the 10% con fid en ce level. b. Significant at the 10% con fid en ce level (one-tailed). c. Significant at the 5% confid ence level. d. Significant at the 1% confid ence level. SOURCE: Author’s calculations. Table 3A lists the results for the period February 23, 1987 to July 2, 1987. For the West German mark, the coefficient for initial purchases of marks is positive and significant. One cannot interpret this coefficient unambiguously, because causality is bidirectional without the lag; never theless, the positive cofficient is not consistent with the view that intervention purchases of marks produced a dollar depreciation. The lagged value on initial intervention is marginally significant and correctly signed. The United States bought a small amount of marks on March 11, as the dollar rose above 1.85 marks. The dollar depreciated on the following day. The coefficients on the sales of marks are incorrectly signed and/or insignificant. For the Japanese yen, all of the coefficients are either incorrectly signed or insignificant. Table 4A presents the results for the period July 5, 1987 to August 28, 1987. For the West German mark, the coefficient for initial pur chases of marks is positive and significant. As before, this coefficient cannot be unambiguously interpreted, but the sign is not consistent with the view that intervention purchases of marks produced a dollar depreciation. The remaining intervention variables are not significant. For the yen, the coefficients are either incorrectly signed or are not significant. 6 7 Intsrvsntion I. Estimation Period: July 5, 1987 to August 28, 1987 A. Dependent Variable: mark-dollar exchange rate Independent Variables Intervention dummies Initial purchases no lag lagged Subsequent purchases no lag lagged Initial sales no lag lagged Subsequent sales no lag lagged Lagged dependent Coefficient (1 ) (1 ) (3) (3) (0 ) (0 ) (0 ) (0 ) 0 .0 12 -0 . 0 0 1 0.003 0.0 02 — — — — 0.999 T-statistic 2.53a -0.27 0.75 0.47 — — — — 758.5b Sum of Squared Residuals = 0 . 0 0 1 R2 = 0.849 n = 38 B. Dependent Variable: yen-dollar exchange rate Independent Variables Intervention dummies Initial purchases no lag lagged Subsequent purchases no lag lagged Initial sales no lag lagged Subsequent sales no lag lagged Lagged dependent Coefficient T-statistic — — — — ( 1 ) -0.018 0.009 (1 ) (0 ) — (0 ) — (0 (0 (0 (0 ) ) ) ) 1.0 0 0 — — — — -2.513 1.2 0 — — 4l66.2b Sum of Squared Residuals = 0 . 0 0 2 R2 = 0 . 8 3 0 n = 38 NOTE: Intervention refers to U.S. purchases or sales o f foreign currencies. Numbers in parentheses indicate the number o f times the dum m y equals 1. a. Significant at the 5% confid ence level. b. Significant at the \% con fid en ce level. SOURCE: Author’s calculations. Economic Review ■ Quarter III 1987 Can Services Be a Source o f Export-led Growth? Evidence from the Fourth District by Erica L Groshen Identifying Amenity and Productivity Cities Using Wage and Rent Differentials by Patricia E. Beeson and Randall W. Eberts FSLIC Forbearances to Stockholders and the Value o f Savings and Loan Shares by James B. Thomson ■ Quarter IV 1987 Learning, Rationality, the Stability o f Equilibrium and Macroeconomics by John B. Carlson Airline Hubs: A Study o f Determining Factors and Effects by Paul W. Bauer A Comparison o f Risk-Based Capital and Risk-Based Deposit Insurance by Robert B. Avery and Terrence M. Belton ■ Quarter I 1988 Can Competition Among Local Governments Constrain Government Spending? by Randall W. Eberts and Timothy J. Gronberg Exit Barriers in the Steel Industry by Mary E. Deily Why Do Wages Vary Among Employers? by Erica L Groshen ■ Quarter II 1988 Intervention and the Dollar’s Decline by Owen F. Humpage Using Financial Data to Identify Changes in Bank Condition by Gary Whalen and James B. Thomson Developing Country Lending and Current Banking Conditions by Walker F. Todd a Third Quarter Working Papers Working Paper Notice Current Working Papers of the Cleve land Federal Reserve Bank are listed in Economic Single copies of individual papers will Institutional subscribers, such as librar be sent free of charge to those who ies and other organizations, will be request them. A mailing list service for placed on a mailing list upon request personal subscribers, however, is not and will automatically receive available. Papers ■ 8805 ■ 8807 L e s s o n s o f th e P a s t and P ro s p e c ts fo r the Fu tu re in Le n d e r T e s tin g fo r ■ 8809 In te ru rb a n C o m p a ris o n s o f the Q u a lity o f Life each quarterly issue of the Review. Copies of specific papers may be requested by completing and mail Working as they are published. ing the attached form below. S p e c u la tiv e B u bble s in S to c k P ric e s o f L a s t R e s o rt T h e o ry by Asli Demirguc-Kunt by Patricia E. Beeson by W alker F. Todd and Hashem Dezhbaksh and Randall W. Eberts ■ 8806 ■ 8808 F in a n c ia l S tru c tu re and th e A d ju s tm e n t ■ 8 8 10 U s in g S M V A M a s a C a p ita l R e q u ire m e n ts and O p tim a l B a n k P o rtfo lio s : A R e e x a m in a tio n o f C a p ita l S to c k L in e a r A p p r o x im a tio n to a N o n lin e a r F u n c tio n : by William P. Osterberg A N o te by William P. Osterberg by Asli Demirguc-Kunt and James B. Thomson and James B. Thomson Please com plete and detach the form below and mail to: Federal Reserve B ank of Cleveland Research D ep a rtm en t R O . Box 6387 C leve lan d, O h io 44101 Check item(s) Send to : Please send the fo llo w in g W o rkin g Paper(s): □ 8805 □ 8807 □ 8809 □ 8806 □ 8808 □ 8 8 10 ________________________________________________________ Nam e Please p rin t A ddress C ity State Zip