The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
The News in Financial Asset Returns GERALD P. DWYER JR. AND CESARE ROBOTTI Dwyer is the vice president in charge of the financial section and Robotti is a financial economist and assistant policy adviser, both in the Atlanta Fed’s research department. The authors thank Thomas Cunningham, Mark Fisher, Paula Tkac, and Daniel Waggoner for helpful comments on earlier drafts. They also thank Budina Naydenova and Shalini Patel for outstanding research assistance. re returns on financial markets useful for predicting the future course of the economy? It is widely thought that financial markets’ movements reflect the economy’s future and that finding the message in financial asset returns is one way to discern this future. The message is not always clear, though. For example, on November 3, 2003, a Wall Street Journal story attempted to reconcile apparently conflicting signals from stock and bond prices about whether economic growth would continue to be high in the future (Browning and Lucchetti 2003). The widespread notion that financial asset returns are related to future economic activity is plausible. An improvement in a company’s prospects is likely to result in a rise in its stock prices (Kamstra 2003). If a widespread increase in stock prices occurs, it is possible that many companies’ prospects have improved and the economy will grow faster. Those brighter prospects can be associated with faster output growth simply because increases in asset prices reflect good news about future economic conditions. Rising stock prices also can be associated with faster output growth because higher stock prices increase households’ wealth, thus boosting consumption and output and thereby improving firms’ prospects. Either way, rising stock prices can be associated with future increases in output. A Returns on bonds also can reflect the economy’s future although the relationship is slightly more complex. The actual return on a bond in any period includes both an expected and an unexpected return. By definition, the expected return is not a surprise and is not news. The unexpected return, on the other hand, reflects news about the future. Higher growth lowers bond prices and therefore returns; lower growth raises bond prices and therefore returns.1 Detailed analyses provide surprisingly little support for financial markets’ ability to reveal future economic activity. For example, Stock et al. (1989) find that aggregate stock returns have little value for predicting economic activity, given other variables. They emphasized new leading indicators such as term and default spreads, which promptly failed to forecast a recession (Stock and Watson 2003). The evidence on alternative indicators based on asset prices is mixed (Smith 1999). In part, the problem is that a variety of indicators based on different, seemingly plausible lines of argument have been proposed. Typically, researchers who propose an indicator find the data consistent with its importance, and then other researchers who test the indicator find the evidence lacking (Stock and Watson 2003). Despite the less than sterling record associated with such indicators, there is a strong temptation to use movements in stock indexes and more general returns on financial markets to help discern the Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 1 FIGURE Stock Prices and Recessions, January 1947 through December 2003 S&P 500 1,000 100 10 1945 1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 Source: S&P 500 from CRSP future of the economy. After all, other forecasts are not particularly helpful either. Stock prices reflect investors’ expectations of the future, and it is hard to imagine that they do not contain useful information.2 Even if stock returns do help predict future economic activity, the level of stock prices measured by an index number is not necessarily the best way to use asset prices to forecast the economy. Lamont (2001) provides evidence that the market stock portfolio is not the best portfolio for predicting future economic activity. Creating a market portfolio based on firms’ market values can bury industryspecific data that might be informative about future economic activity such as output growth and inflation. As a consequence, Lamont investigates predictions of economic activity from alternative combinations of the information contained in asset prices, which he calls economic tracking portfolios— portfolios of stocks and bonds with returns that have the best-fitting joint linear relationship with economic activity.3 Lamont (2001) finds that tracking portfolios are related to future economic activity and presents evidence that they are useful for forecasting economic activity over the next year. Curiously, there is more solid evidence that financial markets reflect news about economic activity than that markets’ reflection of news helps predict 2 economic activity. Two early papers on the effect on financial markets of news about economic activity are Chen, Roll, and Ross (1986) and Dwyer and Hafer (1989), and two recent papers are Flannery and Protopapadakis (2002) and Balduzzi, Elton, and Green (2001). Overall, these studies do find evidence of relationships between financial asset returns and economic activity although the relationship is more evident for bonds than for stocks. This article examines and answers two questions: First, what is a good way of extracting information about future economic activity from asset prices? Second, do financial asset returns help predict economic activity over horizons from one month to five years? While the questions are similar, in part, to Lamont’s (2001), the methods used here are different in some regards, and more recent data allow us to examine the late 1990s and the recession in 2001. Stock Returns and Recessions efore delving into evidence from a more technical analysis, we present some simple evidence on a basic question: How well do stock prices predict recessions? The figure above shows the S&P 500 stock index from January 1947 to December 2003; the shading represents periods of recession as defined by the National Bureau of Economic Research (NBER). B Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 TABLE 1 Recessions and Stock Returns Recessions (years) 1948–49 1953–54 1957–58 1960–61 1969–70 1973–75 1980 1981–82 1990–91 2001 Average Peak of stock index (1) (month) Peak of business cycle (2) (month) 06/48 12/52 07/57 07/59 11/68 12/72 01/80 11/80 05/90 08/00 11/48 07/53 08/57 04/60 12/69 11/73 01/80 07/81 07/90 03/01 Lead Stock index time decline between from peaks (1) to (2) (months) (percent) 5 7 1 9 13 11 0 8 2 7 6.3 –11.95 –6.83 –5.62 –10.15 –15.05 –17.37 0 –6.83 –1.40 –23.55 –9.88 The figure makes it easy to see why it is tempting to use a stock market index to predict recessions. Every recession is associated with a fall in the S&P 500, with all but the drop in 1980 preceding the recession. Table 1 summarizes details from the figure. The analysis is similar to Siegel’s (1998, chap. 12) analysis of decreases in the S&P 500 index before recessions.4 Table 1 shows whether the stock market falls before the beginning of a recession and starts to rise before the end of a recession. Stock prices can rise and fall on successive days, so it would be meaningless to simply examine whether stock prices fall before a recession. On some days they do; on some days they don’t. A more informative definition is Siegel’s; he defines a fall before a recession as a decline of 8 percent or more and a peak as the highest level from which prices fall 8 percent. The low point of the index—a trough—is the lowest level Trough of the stock index (month) Maximum decline of stock index (percent) 06/49 08/53 12/57 10/60 06/70 09/74 03/80 07/82 10/90 09/01 –15.47 –12.23 –16.53 –11.77 –32.90 –46.18 –10.57 –23.79 –15.84 –31.41 –21.67 Lead time from trough to the end Length End of the of the of the recession recession recession (month) (months) (months) 10/49 05/54 04/58 02/61 11/70 03/75 07/80 11/82 03/91 11/01 5 10 5 5 6 7 5 5 6 3 5.7 12 11 9 11 12 17 7 17 9 9 11.4 before stock prices rise 8 percent. Table 1 also examines whether a stock market increase signals the end of the recession; we use a related definition of such a signal as being a rise of 8 percent or more that started during the recession. Table 1 shows that stock prices peak anywhere from zero to thirteen months before the start of a recession. The average lead time between the peak of the S&P 500 and the start of a recession is 6.3 months, and the average stock market decline before a recession is 9.9 percent. The average decline in the stock market before and during a recession is 21.7 percent, with a wide range from a 10.6 percent fall in 1953–54 to the collapse of stock prices by 46.2 percent in 1973–75. On average, these declines took place over twelve months. Typically, the stock market falls less before a recession than during it, which limits the market’s value 1. In general, higher growth is associated with a higher expected return. News of higher growth in the future, though, is associated with a lower unexpected return today because the price of a fixed-income security must fall to provide higher expected returns in the future. 2. See, for example, Del Negro (2001), Smith (1999), and Stock et al. (1989). The basic idea is related to Hayek (1948). 3. Economic tracking portfolios are similar to “maximum correlation portfolios,” introduced by Breeden, Gibbons, and Litzenberger (1989), and “mimicking portfolios,” which have been used for a variety of purposes, including tests of asset pricing models. For example, Breeden, Gibbons, and Litzenberger (1989) construct tracking portfolios for current consumption to test the consumption capital asset pricing model, and Balduzzi and Robotti (2001) use tracking portfolios to test the intertemporal capital asset pricing model. Returns on such portfolios also can be used to calculate the risk premia received by holders of various types of risk. In fact, the economic risk premia are the excess cash flows on the mimicking portfolios (Robotti 2002). Such portfolios can be used as hedging devices by individuals who wish to insure themselves against a particular economic risk; for example, to insure themselves against inflation, individuals can take a position in the mimicking portfolio for inflation to offset predictable inflation. 4. The NBER recession dates—other than the 2001 recession, which occurred after Siegel’s analysis was published—are identical. The dates and stock market returns are similar but differ at least partly because Siegel appears to have used monthly averages of the S&P 500 index and this study uses the value at the end of the last trading day of the month. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 3 for forecasting recessions. Still, decreases in the stock market appear to be useful for identifying whether the economy is currently in a recession. Given the typical ups and downs of the economy, there have been many times in the last fifty years when it has been difficult to determine that the economy is in a recession even while one is under way—a difficulty even in the two most recent recessions. The stock market does not necessarily decline substantially before a recession, but the onset of a recession is invariably associated with a substantial decline in stock prices. In every case, prices have begun to increase before the end of a recession. The figure shows this It is widely thought that financial markets’ movements reflect the economy’s future and that finding the message in financial asset returns is one way to discern this future. pattern, with stock prices starting to increase 5.7 months, on average, before the end of the recession. This analysis answers part—but not all—of the question about stock prices’ ability to predict recessions. The analysis shows that if the NBER identifies a recession, then a fall in stock prices has occurred about the same time in every recession since World War II. These results do not imply that substantial decreases in stock prices indicate that there is a recession. Falling stock prices are not a certain indicator of a recession, as the patterns for 2001 to early 2003 show. The S&P 500 index fell 29.0 percent from December 2001 to September 2002, rose 14.9 percent from September to November 2002, and then fell 10.2 percent from November 2002 to February 2003. There was no recession within twelve months of the start of the 2002–03 decreases in stock prices. In fact, the revised estimate of GDP growth for the third quarter of 2003 is growth at more than an 8 percent annual rate—more than a little distant from a recession. Table 2 presents the other episodes since World War II in which stock prices fell 8 percent or more and no recession began within twelve months after the fall began. Even though they are associated with recessions, falling stock prices do not necessarily mean that a recession is coming or is under way. 4 In the post–World War II period, an 8 percent drop in stock prices has signaled nineteen recessions—nearly twice as many as the ten that have in fact occurred.5 Recessions have occurred only 53 percent of the time that falling stock prices would suggest a recession, indicating that falling stock prices are roughly a fifty-fifty predictor of recessions. This statement is not the same as saying that falling stock prices would predict a recession 50 percent of the time. Fortunately, an 8 percent drop in stock prices is not that common. If stock prices drop by 8 percent or more, there is about a 50 percent chance of a recession. Given that falling stock prices do appear to be a signal about the economy’s prospects, is there a way to extract more general information about the economy from stock prices and other financial asset returns? Asset Returns and News about Economic Activity his section outlines a way to extract the news about future economic activity from returns on financial assets. The unexpected part—by definition, the part that is news—of an asset’s excess return can reveal information about unexpected economic activity.6 Let εt+1 be the unexpected part of economic activity from t to t +1 and ηt be the unexpected part of a financial asset’s excess return from t –1 to t. The linear relationship between the unexpected part of economic activity and the news in the asset’s return is given by T (1) εt+1 = bηt + et+1, where bη t is the part of the next period’s economic activity predicted by the current news in the asset’s excess return, ηt, and et+1 is the part of economic activity not predicted by the news in the asset’s return. There are no data series called “unexpected economic activity” and “news in financial returns.” To determine εt+1 and η t, we must estimate expected economic activity and assets’ expected excess returns because the unexpected parts of economic activity and returns are the differences between actual values and expected values. Let yt+1 denote a measure of economic activity such as the growth rate of industrial production from period t to t +1, and let rt denote the excess return on an asset from the end of period t –1 to the end of period t. The unexpected part of the variation of economic activity and of the excess return can then be written as (2) ε t+1 = yt+1 − E[ yt+1 | Ωt−1 ] ηt = rt − E[ rt | Ωt−1], Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 TABLE 2 Years of False Alarm False alarms (years) Peak of stock index (month) Trough of stock index (month) Maximum decline of stock index (percent) 1956–57 1962 1966 1978 1984 1987 1998 2001–02 2002–03 07/56 12/61 01/66 08/78 11/83 08/87 06/98 12/01 11/02 02/57 06/62 09/66 10/78 05/84 11/87 08/98 09/02 02/03 –12.42 –23.48 –17.57 –9.82 –9.53 –30.17 –15.57 –28.99 –10.16 where Ωt–1 is the set of information available at the end of period t –1, E[yt+1 Ωt–1] is the expected level of economic activity conditional on information available at the end of period t –1, and E[rt Ωt–1] is the expected return in period t conditional on information available at the end of period t –1. Only the unexpected part of economic activity is related to news in the asset’s return because the information already known about economic activity is reflected in the expected part of economic activity and the asset’s expected excess return. These unexpected parts of economic activity and the asset’s return can be related to actual economic activity and the actual return on an asset by linear regressions.7 The news in the asset’s return about future economic activity is uncorrelated with the part of economic activity not predicted by such news by construction. This simple relationship easily can be extended to include more assets and cover more periods. If the unexpected return on asset i is ηi,t and there are N assets, then the relationship between unexpected growth of economic activity and the unexpected returns on the N assets is (3) εt+1 = b1η1,t + b2η2,t + … + bNηN,t + et+1. It is useful to extend economic activity in equation (1) to cover several periods instead of one period because the unexpected return on an asset in any given month generally reflects information about more than one month. For an asset such as stock or a bond with a life longer than one period, the unexpected return on the asset in period t reflects changes in expectations not just for this period but for all future periods reflected in the asset’s price. The unexpected return is part of the capital gain portion of the asset’s total dollar return— the change in the asset’s price—and the unexpected part of the change in this price reflects changes in the payoffs to investors in any or all of the periods over the entire life of the asset. This version of equation (1) over a longer horizon for economic activity is (4) ε kt+1 = bηt + etk+1 , where e kt+1 is the part of the growth rate of economic activity from t to t + k that is not predicted by the news in asset prices in period t and the superscript indicates the number of periods for which growth rates are computed. The error term in equation (4) is serially correlated in general if the data are sampled every period (that is, at t+1, t+ 2,. . .) and unexpected economic activity overlaps.8 This serial 5. This finding is similar to the findings of Samuelson (1966) and Siegel (1998, chap. 12). 6. The excess return on an asset is defined as the return on that asset minus the return on a riskless security. 7. We use the notation of mathematical expectations and call the measures “expected,” but linear projections are sufficient for our purposes. 8. This overlap induces a moving-average error term with k nonzero autocorrelations. This autocorrelation is consistent with the definition of news and unexpected economic activity. News (ηt, ηt+1,…) is serially uncorrelated. Unexpected economic activity for one period (ε1t+1, ε1t+2,…) has one moving-average term because expected economic activity from t to t+1 is conditioned on information in t–1. Unexpected economic activity for two periods (ε 2t+1, ε 2t+2,…) has two moving-average error terms, and so on. If the underlying relationship between news in asset returns and economic activity is exactly linear, then equation (1) with standard forecast updating and equation (4) yield identical forecasts with minimum mean-squared error. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 5 correlation complicates estimation of equation (4) but does not preclude the usefulness of the oneperiod unexpected return’s information about economic activity over several future periods. Identifying unexpected economic activity and the unexpected excess return is more problematic because neither is directly observable. To identify ε kt+1 and η j, t( j = 1,..., N ), equation (2) shows that it is necessary to estimate expected economic activity and expected excess returns on the assets because the unexpected parts of economic activity and returns are the differences between actual and expected values. With multiple periods and assets, equation (2) becomes k k k (5) ε t +1 = yt +1 − E[ yt +1 | Ωt−1] ηi,t = ri,t − E [ ri,t | Ωt−1], where y kt+1 is the growth rate of economic activity from t to t + k and ri,t is the return on asset i in period t. If there is a linear relationship between expected economic activity, the expected excess returns on assets, and other variables known to investors (zj,t–1, j = 1,…, M ), expected economic activity and expected excess returns are given by (6) E[ ytk+1 | Ωt−1 ] = α y + β1, y z1, t −1 + β2, y z2, t −1 + ... + βM, y z M, t −1. (7) E[ri, t | Ωt −1 ] = α r + β1,r z1, t −1 + β2,r z2, t −1 + ... + β M, r zM, t −1. Including the same variables in both equations may seem restrictive, but it is not because some coefficients in equations (6) and (7) can be zero. It is more restrictive to limit the analysis to a linear relationship. This limitation can be justified by assuming that variables are jointly normally distributed, but this assumption is implausible. Even without the assumption of normal distributions, the econometric analysis and conclusions are correct if they are limited to linear relationships—that is, if the analysis is limited to the information in the linear relationships among variables. Equations (6) and (7) raise another important issue, though. There is no reason to think that the set of variables used in equations (6) and (7) is the complete set of information known to investors. What are the consequences of not knowing all the information available to investors? The implications can be illustrated with one asset. Using equations (6) and (7), we can rewrite equation (4) as 6 (8) ytk+1 − (α y+ β1, y z1, t −1 + β2, y z2, t −1 + ... + β M, y zM, t −1 ) = b[rt − (α r + β1,r z1, t −1 + β2, r z2, t −1 + ... + β M, r zM, t −1)] + etk+1 , which can be simplified to k (9) yt +1 = ( α y − bα r ) + brt + (β1, y − β1, rb)z1, t −1 + (β2, y − β2, r b)z2, t −1 + ... +(β M, y − β M, rb)z M, t−1+ etk+1. The coefficients in equation (9) can be summarized for convenience by k (10) yt +1 = γ + brt + δ1 z1,t −1 + δ 2 z2, t −1 + ...δ M z M, t −1 + etk+1. If only a subset of the information available to investors—say, z1,t–1—is included in (10), the estimated equation is (11) ytk+1 = c0 + c1rt + dz1,t −1 + vtk+1 instead of equation (10). The estimated relationship between unexpected economic activity and the unexpected excess return, measured by the coefficient c1, will be the same as b if the excess return is uncorrelated with the variables left out of the estimated equation. If the variables included in equation (10) but left out of (11) are correlated with the excess return, then the estimated coefficient c1will not be the same as b. In general, this source of bias is not likely to be empirically important in our analysis because we use monthly data on excess returns, and excess returns on stocks and bonds are not very predictable at this frequency. It might seem that we could lessen the likelihood of this bias by including numerous, possibly superfluous, variables, but obtaining an excellent fit in sample and a worse fit out of sample is a likely and often serious consequence of this strategy. Given that the purpose of the analysis is to forecast economic activity, we limit the analysis to a relatively small set of variables to lessen problems of overfitting. Even if variables left out of the estimated equation do not help predict the excess return, they might help predict economic activity, in which case the estimated error term, v kt+1, in equation (11) is bigger than the underlying error term, e kt+1. If so, we are more likely to find that the relationship between economic activity and the excess return is statisti- Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 cally insignificant and therefore conclude that the excess return provides little information about economic activity even if there is such a relationship. The remainder of this article focuses on estimates of the following equation: k (12) yt +1 = c0 + c1 r1,t + c2r2,t + ... + cN rN, t + d1 z1, t−1 + d2 z2, t−1 + ... + dM zM, t−1 + ς tk+1 . For each measure of economic activity, y, with N financial assets and M additional variables, we study the properties of the “economic tracking portfolio,” c1r1+c2r2…+cNrN. Equation (12) can be estimated by ordinary least squares (OLS), and we do so. The standard errors and test statistics reported in the article are based on the Newey and West (1987) correction of estimated standard errors with twelve lags; this method corrects for the serial correlation in ζ kt+1 caused by economic activity being measured over overlapping periods. by risk factors for returns (Veronesi and Santos 2001)—and the unemployment rate. The inflation rate is based on the consumer price index. Measures of future financial economic activity included are the excess return on the Center for Research in Security Prices (CRSP) value-weighted aggregate portfolio over the horizon, the excess return on a portfolio of long-term government bonds (with a term of approximately twenty years), and the return on onemonth Treasury bills. The analysis uses growth rates for all but two of the variables; the analysis uses the change in the unemployment rate and financial assets’ excess returns themselves. Growth rates better represent Rising stock prices can be associated with faster output growth because higher stock prices increase households’ wealth, thus boosting consumption and output and thereby improving firms’ prospects. The Data or asset returns and economic activity, the period covered by the data in this article generally is February 1947 to August 2002, and for other variables, January 1947 to July 2002.9 These starting and ending dates are dictated by data availability and the end of World War II. All series are monthly. Economic activity. The measures of aggregate economic activity examined include industrial production, consumption, labor market activity, inflation, and future returns on financial markets. Industrial production is measured by total industrial production and broad production classes: manufacturing, consumer durable goods, consumer nondurable goods, mining, and utilities. Production of durable goods has more cyclical variation than the other classes of production, so it is worthwhile to examine durable goods separately from nondurable goods and manufacturing. Consumption is measured by total consumption and two components: consumption of durable goods and consumption of nondurable goods and services. Labor market activity is measured by real labor income—a variable suggested F the short-run variation in the series, which we are trying to predict, instead of long-term trends. The change in the unemployment rate filters out longterm trends in the level, and the excess returns themselves do not have long-term trends, so it is unnecessary, and undesirable, to use changes in returns. Returns. In the base set of regressions, returns are measured by one aggregate stock index, eight industry stock portfolios, and four bond returns. All excess returns are one-month returns in excess of the one-month Treasury bill return. The aggregate stock index is the NYSE-AMEX-Nasdaq valueweighted stock market portfolio (from CRSP). The industry indexes are for basic industries, capital goods, construction, consumer goods, energy, finance, transportation, and utilities.10 These industries are partly related to the component industries in industrial production but are far from a one-toone mapping. The four bond returns are returns on 9. The change in the unemployment rate starts in January 1948 and ends in August 2002. The growth rate of real consumption starts in February 1959 and ends in August 2002. 10. The data appendix provides a more detailed description of the industries. We also conducted the empirical analysis with the five Fama-French industries, which did not affect the conclusions. The five industry returns are for manufacturing, utilities, shops, finance, and a catch-all category called “other industries.” These other industries include agriculture, mining, oil, construction, telecommunications, health services, and legal services. Again, further details on the definitions of the industries are presented in the data appendix. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 7 a long-term government bond, an intermediate-term government bond, a one-year government bond, and a high-grade corporate bond. The long-term, intermediate-term, and high-grade bond returns are from Ibbotson Associates. The one-year bond return and the one-month Treasury bill rate are from CRSP. Additional variables. The estimates of the expected values of economic activity and returns are based on regressions that include a constant term and past values of eight variables that have been used in tests of asset pricing models and studies of stock and bond return predictability (including, among others, Chen, Roll, and Ross 1986; Burmeister and McElroy 1988; Ferson and Harvey Creating a market portfolio based on firms’ market values can bury industry-specific data that might be informative about future economic activity such as output growth and inflation. 1991, 1999; Downs and Snow 1994; Kirby 1998; Balduzzi and Robotti 2001; and Lamont 2001). All of these additional variables are the same in every estimated equation.11 Three of the variables are measured over the prior twelve months, and the rest are measured over the prior month. The difference in the time frames is suggested by prior evidence of differences in the apparent persistence of variables’ effects. The variables measured over the prior twelve months, which are assumed to have more persistent effects, are industrial production growth, the inflation rate, and the aggregate excess return on the CRSP value-weighted stock index. The variables measured over the prior month, which are assumed to have less persistent effects, are the dividend yield, term premia for one-year Treasury securities and long-term government bonds relative to the one-month Treasury bill yield, default premia measured by the commercial paper yield minus the Treasury bill yield and the BAA bond yield minus the yield on AAA corporate bonds, and the return on a one-month Treasury bill itself. The Evidence n this section, we examine the importance of news in financial market returns for future economic activity. We investigate whether the estimated tracking portfolios are related to otherwise unpre- I 8 dicted economic activity, and we estimate rolling regressions to assess the out-of-sample performance of the tracking portfolios in forecasting future economic activity. The statistical significance of news in excess returns. Is the news in asset returns related to unexpected changes in economic activity? Table 3 summarizes the evidence by presenting p-values of tests whether all of the thirteen excess returns are related to the measures of unexpected economic activity at horizons from one month to twelve months.12 These p-values are the probability of a test statistic as large as the one observed if the coefficients in equation (12) are zero. A large p-value means that the test statistic is quite likely if the restrictions are correct, and a small p-value means that the test statistic is unlikely if the restrictions are correct. A small p-value provides more support for a relationship between the excess return and the measure of economic activity, with a p-value of 0.05 or less being fairly unlikely if the variables do not belong in the regression. Hence, we use the conventional pvalue of 0.05, or 5 percent, for deciding whether news in financial asset returns is statistically related to a measure of economic activity.13 These tests are not independent, most obviously for components of an aggregate; hence, we examine the results for broad patterns and ignore occasional inexplicable “statistically significant” results. For each measure of economic activity and for each time horizon, Table 3 presents p-values for excluding all returns. The tests show that news in financial assets’ excess returns is related to unexpected economic activity. The p-values in Table 3 indicate that the news in financial returns is related to total industrial production, production of manufacturing goods, mining, and production of consumer durable goods—all cyclically sensitive—up to a six-month horizon. Production of consumer nondurable goods and of utilities are not related to the excess returns. This result is consistent with the permanent income theory of consumption, which implies that nondurable goods will be little affected by temporary changes in income. The general relationship between the news in financial asset returns and consumption of durables goods, and the general lack of such a relationship for consumption of nondurables and services, can be explained in a similar way. The only p-value of 5 percent or less for real labor income is at a horizon of twelve months. News in financial returns is related to changes in the unemployment rate and the inflation rate at all horizons. At all horizons, future excess returns are foreshadowed by news in financial returns. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 TABLE 3 Economic Activity and All Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months P-values for Excluding All Returns from Basic Equation Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.02 0.03 0.57 0.48 0.13 0.07 0.00 0.00 0.06 0.03 0.72 0.56 0.00 0.00 0.03 0.03 0.65 0.38 0.22 0.20 0.08 0.25 0.65 0.88 Consumption Total Durable goods Nondurable goods and services 0.17 0.24 0.25 0.00 0.00 0.05 0.20 0.00 0.31 0.06 0.00 0.14 Labor market Unemployment rate Real labor income 0.02 0.19 0.00 0.66 0.00 0.30 0.00 0.04 Inflation rate 0.00 0.00 0.00 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.01 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 Both stock and bond returns are related to economic activity. Table 4 shows the p-values for deleting the stock returns and for separately deleting the bond returns.14 The low p-values show that news in the market stock return and the eight industry portfolio returns are related to economic activity, with the strongest relationship at the three-month and six-month horizons. There is little relationship between returns and industrial production and consumption one month in the future but more of a relationship in the next three to six months. This result is not necessarily surprising: It is plausible that news about longer-term developments has larger effects on these securities’ returns.15 Stock returns are represented in the regressions by the return on the aggregate market port- folio and by returns on industry portfolios. Taken together, the p-values in Table 4 indicate that the combination of these returns is related to both total and manufacturing industrial production over the next three to six months. News in stock returns is related to unexpected changes in both unemployment and inflation at all horizons. Interestingly, little evidence exists that stock returns are related to unexpected changes in consumption, a surprising result given all the emphasis put on the “wealth effect”—a relationship between wealth in corporate stock and consumption. News in stock returns appears to be related to unexpected changes in the future financial returns on bonds at all horizons and on stocks at horizons of six months and more. 11. This strategy can be contrasted with a strategy of estimating autoregressions for the expected returns and measures of economic activity, which would include different variables in every equation and would run a risk of overfitting in sample and quite possibly fitting worse out of sample. 12. In addition to one to twelve months, we also examined the ability of returns to forecast economic activity five years ahead. There is little evidence that returns help to predict activity over this longer horizon. 13. This method is not always a good way to proceed, but it is informative here because of the large number of tests and the underlying concern that an apparent relationship for one period will not persist in later data. 14. The excess returns on fixed-income securities are termed “bonds” for brevity in the table and the text. 15. Six months is, of course, a short horizon in other contexts. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 9 TABLE 4 Economic Activity and Stock and Bond Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months P-values for Excluding Market and Industry Returns from Basic Equation Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.13 0.16 0.57 0.87 0.18 0.04 0.00 0.00 0.08 0.38 0.42 0.37 0.00 0.00 0.05 0.05 0.50 0.23 0.13 0.19 0.10 0.60 0.87 0.78 Consumption Total Durable goods Nondurable goods and services 0.40 0.34 0.37 0.08 0.10 0.08 0.24 0.05 0.53 0.55 0.47 0.46 Labor market Unemployment rate Real labor income 0.01 0.12 0.00 0.68 0.00 0.19 0.00 0.04 Inflation rate 0.00 0.00 0.00 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.15 0.00 0.49 0.13 0.00 0.48 0.02 0.00 0.09 0.00 0.00 0.00 P-values for Excluding Bond Returns from Basic Equation Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.04 0.06 0.43 0.19 0.12 0.46 0.01 0.02 0.23 0.30 0.99 0.71 0.23 0.24 0.16 0.26 0.91 0.85 0.84 0.63 0.19 0.21 0.26 0.75 Consumption Total Durable goods Nondurable goods and services 0.13 0.31 0.14 0.03 0.03 0.20 0.04 0.02 0.29 0.05 0.02 0.07 Labor market Unemployment rate Real labor income 0.40 0.68 0.06 0.70 0.20 0.82 0.62 0.75 Inflation rate 0.07 0.00 0.00 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.01 0.00 0.00 0.06 0.08 0.00 0.00 0.54 0.00 0.07 0.15 0.00 10 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 At the one-month and three-month horizons, news in bond returns generally is related to industrial production—total and manufacturing. At horizons beyond one month, bond returns are related to total consumption and consumption of durable goods in particular. At all horizons beyond one month, bond returns also are related to inflation. Excess bond returns appear to be more closely related to excess stock returns and the Treasury bill rate than to the excess bond return although there is some relationship between news in excess bond returns and excess bond returns in the next few months. Is the news in stock returns due to news reflected in the aggregate market return, or is there substantial information in returns by industry? Table 5 presents evidence on this issue. The first two parts of the table show p-values for excluding aggregate stock returns and for excluding industry returns from the basic equation with both aggregate and industry returns included in the regressions. The second two parts of the table show p-values for excluding market stock returns with industry returns excluded and for excluding industry returns with the market return excluded. Even though Table 4 shows that news in the combination of aggregate stock market and industry returns is important, a pattern is evident in the first two parts of Table 5—news in neither the market return nor the industry returns seems to be generally important. A glaring exception is the informativeness of the industry returns for inflation—a surprising result.16 These p-values, though, are for tests that drop the market return with industry returns included in the regression and for tests that drop the industry returns with the market return included. Correlation of the aggregate market return and the industry returns is a plausible explanation of these results. It may not matter whether the market return or the set of industry returns is included as long as one of the two is included. In fact, it would not be entirely surprising if the regressions might include either the aggregate return or the set of industry returns because the aggregate market return is a weighted average of the industry returns with time-varying weights. This relationship between the market return and the industry returns suggests that correlation of the aggregate return and the industry returns may well explain why either can be deleted with the other left in the regressions.17 The importance of the correlation of the industry and market returns is supported by comparing the p-values in the last two parts of Table 5 with the pvalues in the first two parts. Both the market return and the set of industry returns have low p-values if the other is excluded. Overall, the p-values in Table 5 indicate that it is important to include either the market return or the industry returns, but once one is included, the other generally is uninformative. Is there some reason to prefer the market return or the industry returns? There is little evidence in Table 5 to support a choice of one over the other.18 The market return appears to be more closely related to industrial production, especially at a horizon of one The stock market does not necessarily decline substantially before a recession, but the onset of a recession is invariably associated with a substantial decline in stock prices. year. The industry returns appear to be more closely related to inflation. On the other hand, the market return has one estimated coefficient instead of the eight coefficients for industry returns. Fitting well in sample and predicting poorly out of sample is likely to be less of a problem with one estimated coefficient than with eight. In the rest of the article, we report results based on estimates with the market return but not the industry returns included in regressions. Appendix tables show statistics for evaluating the informativeness of forecasting with both the market return and industry returns as well as with industry returns alone. Forecasts with the industry returns and market return are roughly as accurate as forecasts with the market return or industry returns. The economic significance of news in excess returns. While p-values are measures of statistical significance, they do not provide a measure of the economic significance of the news in returns for unexpected economic activity. This section reports statistics that summarize the economic significance of the news. 16. The inflation rate reflects price changes in the entire economy, and it is not obvious why financial news by industry should be informative about this aggregate variable. 17. Even though the market portfolio is an aggregate of the industry portfolios, the industry returns can be less informative than the market return in a linear regression because the market return is not a constant linear function of the industry returns. 18. This result is in contrast with Lamont’s (2001) for a different period. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 11 TABLE 5 Stocks versus Industry Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months P-values for Excluding Market Return from Basic Equation When Industry Returns Included Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.57 0.63 0.68 0.78 0.49 0.78 0.87 0.98 0.67 0.47 0.93 0.40 0.77 0.72 0.82 0.15 0.30 0.23 0.79 0.74 0.81 0.46 0.74 0.88 Consumption Total Durable goods Nondurable goods and services 0.64 0.68 0.91 0.45 0.95 0.12 0.95 0.67 0.62 0.74 0.70 0.37 Labor market Unemployment rate Real labor income 0.29 0.09 0.98 0.76 0.65 0.50 0.64 0.56 Inflation rate 0.29 0.14 0.93 0.32 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.93 0.02 0.74 0.22 0.04 0.48 0.29 0.07 0.05 0.25 0.04 0.37 P-values for Excluding Industry Returns from Basic Equation When Market Return Included Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.31 0.27 0.58 0.87 0.20 0.06 0.31 0.29 0.20 0.29 0.60 0.71 0.79 0.75 0.10 0.44 0.79 0.76 0.87 0.88 0.44 0.52 0.97 0.95 Consumption Total Durable goods Nondurable goods and services 0.67 0.71 0.36 0.11 0.08 0.15 0.36 0.26 0.43 0.48 0.41 0.37 Labor market Unemployment rate Real labor income 0.04 0.08 0.14 0.69 0.48 0.63 0.44 0.30 Inflation rate 0.00 0.00 0.00 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.11 0.00 0.50 0.43 0.03 0.70 0.13 0.03 0.28 0.28 0.00 0.50 12 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 T A B L E 5 (continued) Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months P-values for Excluding Market Return from Basic Equation When Industry Returns Not Included Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.04 0.08 0.30 0.41 0.22 0.11 0.00 0.00 0.03 0.00 0.10 0.04 0.00 0.00 0.07 0.00 0.06 0.00 0.00 0.00 0.01 0.65 0.13 0.09 Consumption Total Durable goods Nondurable goods and services 0.05 0.03 0.32 0.13 0.56 0.07 0.10 0.01 0.88 0.62 0.51 0.86 Labor market Unemployment rate Real labor income 0.02 0.99 0.00 0.09 0.00 0.01 0.00 0.00 Inflation rate 0.12 0.01 0.03 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.83 0.00 0.28 0.02 0.00 0.08 0.00 0.00 0.00 0.00 0.00 0.00 P-values for Excluding Industry Returns from Basic Equation When Market Return Not Included Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 0.10 0.12 0.48 0.81 0.15 0.02 0.00 0.00 0.05 0.02 0.33 0.34 0.00 0.00 0.04 0.06 0.52 0.24 0.09 0.14 0.06 0.55 0.82 0.69 Consumption Total Durable goods Nondurable goods and services 0.32 0.26 0.28 0.06 0.07 0.12 0.17 0.03 0.45 0.47 0.38 0.44 Labor market Unemployment rate Real labor income 0.01 0.19 0.00 0.40 0.00 0.15 0.00 0.03 Inflation rate 0.00 0.00 0.00 0.00 Financial market returns Excess stock return Excess bond return Treasury bill rate 0.10 0.00 0.40 0.14 0.00 0.42 0.01 0.00 0.00 0.04 0.00 0.00 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 13 TABLE 6 The Percentage of Variation in Unexpected Economic Activity Predicted by News in Financial Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months Industrial production Total Manufacturing Mining Consumer durable goods Consumer nondurable goods Utilities 2.60 2.20 0.80 1.30 1.20 1.00 4.60 4.10 1.60 2.10 0.50 0.90 4.00 3.90 1.70 2.50 0.90 1.40 2.00 2.10 2.00 1.40 1.30 0.70 Consumption Total Durable goods Nondurable goods and services 2.30 2.10 1.40 4.00 4.30 1.90 4.40 6.70 1.40 2.80 4.20 2.00 Labor market Unemployment rate Real labor income 1.60 0.50 3.50 0.70 4.00 1.40 3.20 2.00 Inflation rate 2.70 5.40 4.50 4.40 2.20 5.40 19.80 2.40 4.40 27.30 4.20 3.20 20.40 2.70 3.00 19.20 Financial market returns Excess stock return Excess bond return Treasury bill rate Analysis. Table 6 presents estimates of the closeness of the relationship between unexpected economic activity and unexpected excess returns, measured by the percentage of the variation of otherwise unexpected economic activity associated with unexpected excess returns.19 The estimates are noisy because unexpected economic activity and unexpected returns are not directly observable and must be estimated, and the estimates of expected returns and expected economic activity are themselves noisy. As a result, these estimates are likely to understate the value of the news in unexpected returns, especially if part of what we estimate to be unexpected activity really was expected and unexpected returns are well estimated.20 Even so, the estimates in Table 6 are not exactly overwhelming. For example, the highest percentage of variation for a variable other than a financial asset return is 5.4 percent for inflation at the three-month horizon. This percentage is not high enough to inspire confidence in using the estimated relationship for hedging. The percentages of variation predicted by news in excess returns are positive, but they are far from the maximum value of 100. These estimates provide a partial measure of unexpected news’ importance: News in financial returns has some information 14 about future unexpected economic activity, but it is far from perfect. An alternative estimate of the relative importance of news is its importance for forecasting, which is examined next. Forecasting using excess returns. Tables 3 through 5 indicate that financial assets’ excess returns do contain news about future economic activity. While informative, this evidence is not really sufficient to ensure that the excess returns are useful for forecasting because the estimated regressions are based on the data that the excess returns supposedly are forecasting. How well do these returns help to predict the future when the relationship is estimated based only on the past? Tables 7 and 8 summarize the results of using rolling regressions estimated using data for successive twenty-year periods to evaluate the out-ofsample performance of the financial returns. For each month, we estimate the basic equation (12) using data for the most recent twenty years and make a forecast for a horizon from one to twelve months. The forecasts are “out of sample” because the forecast is made for a period that is not included in the estimated regression. Running these regressions for every possible period generates a set of forecasts for every possible month. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 Not all the measures of economic activity included in Tables 3 through 6 are included in Tables 7 and 8. The evidence in Tables 3 through 6 suggests financial news generally is not informative for some of the measures of economic activity in these tables, and there is little reason to buttress that evidence by showing that financial news does not help to predict these measures.21 Tables 7 and 8 include industrial production—total, manufacturing, and consumer durable goods—consumption of durable goods, the unemployment rate, the inflation rate, the excess returns on stocks and bonds, and the Treasury bill return. Table 7 summarizes the forecasting ability of the regressions at various horizons by their meansquared errors (MSEs) and R2s. The table presents the MSEs of forecasts based on the rolling regressions and, for comparison, the MSEs of forecasts based on the estimated regressions for the whole period. The estimated regressions for the whole period can be considered in-sample regressions that can be contrasted with the rolling regressions used to forecast out of sample. The rolling regressions’ MSEs in Table 7 are higher than the in-sample MSEs, a result that is not surprising. The regressions for the whole period are the minimum MSEs from constant regressions for the period; the rolling regressions have extra flexibility because the estimated coefficients can change over time, but the rolling regressions cannot fit idiosyncratic changes in the data and then forecast those same idiosyncratic changes. Apparently, too good a fit is a more serious issue than changing coefficients. Table 7 also shows out-of-sample R2-like measures (1 – rolling regression MSE/variance of the measure of economic activity) for the rolling regressions and, for comparison, the R2s for the regressions estimated for the whole period. The out-of-sample R2s are lower. The lower fit out of sample indicates that deterioration of forecast accuracy compared to in-sample fit has to be considered when using tracking portfolios for forecasting or for hedging risks. The rolling regressions have the highest R2s for the Treasury bill return, the inflation rate, and the unemployment rate. The R2s for the Treasury bill return are quite high, but enthusiasm over the rolling regressions’ R2s—from 82 percent one month ahead to 65 percent twelve months ahead—must be tempered by the likelihood that the prior month’s Treasury bill return included in the rolling regressions plays a large role in these relatively high R2s. The R2s for the inflation rate are not as high, ranging from 44 percent at the nine-month horizon to 16 percent at the twelve-month horizon. But enthusiasm over these results must be tempered somewhat again by the realization that the prior year’s inflation rate is included in the rolling regressions. While not particularly helpful for explaining the unemployment rate a month ahead, the forecasts from the rolling regressions predict 27 percent of changes in the unemployment rate over the next year, with no Expected returns on stocks and bonds are affected by developments in the economy, and it is impossible for those developments to affect future expected returns without affecting prices and current returns. lagged unemployment rate in the regressions; this result is the most promising one in the table. The rolling regressions forecast total and manufacturing industrial production much worse than the in-sample fit suggests, but the R2 of 20 percent for industrial production over the next six months is not entirely trivial.22 The rolling regressions uniformly have negative R2s for the excess returns on stocks and bonds, which indicates that a forecast of a recent average might have been better than these forecasts conditional on past financial returns and economic activity. Interestingly, given its cyclical sensitivity, the rolling regressions also are particularly poor at forecasting growth of consumption of durable goods. The rolling regressions are estimated regressions for each month based on data for the most recent twenty years; the forecasts over the horizons can be broken into two parts—one part due to the excess returns in the economic tracking portfolio and the 19. In other words, Table 6 shows the R2s for the estimates of equation (11) times 100. 20. Variables that are left out and uncorrelated with the excess returns raise the residual variance of the regression including the financial returns. Because they are uncorrelated with the excess return, the variables left out do not affect the estimated increase in the residual variance associated with deleting the financial returns. Hence, the marginal R2, which is the change in the residual variance divided by the residual variance with all variables, is lower than it would be otherwise. Table 6 reports this marginal R2 times 100. 21. The statistics in Tables 7 and 8 for measures of economic activity not included in the table show that financial returns are not useful for forecasting variables that are unrelated to financial returns. 22. Recall, though, that the prior year’s growth of total industrial production is included in the regressions. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 15 TABLE 7 Forecast Accuracy Stock Market Returns and Bond Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months Industrial production total Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.54 0.42 0.11 0.31 2.57 1.59 0.18 0.49 7.45 3.63 0.15 0.58 21.67 7.25 0.04 0.68 Industrial production manufacturing Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.67 0.52 0.11 0.30 3.11 1.94 0.19 0.49 8.69 4.27 0.20 0.61 25.73 8.96 0.08 0.68 Industrial production of consumer durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 6.30 5.33 –0.03 0.13 21.77 14.88 –0.02 0.30 40.60 23.65 0.07 0.46 76.08 34.71 0.10 0.59 Consumption of durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 10.14 8.65 –0.07 0.09 17.14 12.83 –0.09 0.18 24.89 15.57 –0.07 0.33 53.71 25.24 –0.43 0.33 Unemployment rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.03 0.02 0.08 0.23 0.10 0.07 0.25 0.50 0.31 0.15 0.24 0.62 0.87 0.34 0.27 0.71 Inflation rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.07 0.06 0.35 0.47 0.39 0.24 0.43 0.65 1.27 0.61 0.44 0.73 6.86 1.76 0.16 0.78 Excess stock return Rolling MSE In-sample MSE Rolling R2 In-sample R2 22.71 19.76 –0.07 0.08 73.83 59.06 –0.10 0.12 161.22 102.97 –0.22 0.22 351.71 183.77 –0.32 0.31 Excess bond return Rolling MSE In-sample MSE Rolling R2 In-sample R2 8.83 7.54 –0.03 0.12 29.49 23.27 –0.05 0.17 57.45 40.16 –0.08 0.24 148.35 65.46 –0.26 0.44 Treasury bill return Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.01 0.01 0.82 0.86 0.07 0.05 0.82 0.88 0.38 0.22 0.75 0.86 2.01 0.79 0.65 0.86 16 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 TABLE 8 Contribution of Financial News to the Forecasts Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months Industrial production total Constant Financial news Other variables 0.037 0.090 0.653 0.168 0.332 0.628 0.285 0.367 0.605 0.826 0.644 0.526 Industrial production manufacturing Constant Financial news Other variables 0.038 0.064 0.665 0.193 0.331 0.633 0.294 0.430 0.627 0.863 0.672 0.546 Industrial production of consumer durable goods Constant Financial news Other variables 0.028 –0.061 0.527 0.227 0.024 0.519 0.160 0.567 0.556 0.227 0.898 0.562 Consumption of durable goods Constant Financial news Other variables 0.520 0.456 –0.141 0.896 0.830 0.254 1.379 0.730 0.394 3.762 0.617 0.331 Unemployment rate Constant Financial news Other variables 0.008 –0.094 0.660 0.020 0.164 0.695 0.031 0.434 0.650 0.048 0.891 0.650 Inflation rate Constant Financial news Other variables 0.082 0.345 0.802 0.353 0.815 0.734 0.776 0.499 0.712 2.278 0.762 0.564 Excess stock return Constant Financial news Other variables 0.282 0.218 0.289 0.923 0.349 0.280 1.952 0.477 0.207 3.828 0.152 0.177 Excess bond return Constant Financial news Other variables 0.094 0.543 0.301 0.283 0.522 0.333 0.524 0.497 0.400 0.620 0.401 0.310 Treasury bill return Constant Financial news Other variables 0.000 0.923 1.068 0.031 1.281 1.054 0.095 1.289 1.060 0.430 1.469 1.017 Note: For each variable and horizon, the three numbers listed are the estimated constant term, estimated coefficient for the improvement in the forecast due to including the estimated news in financial returns, and the estimated coefficient for the forecast using the variables other than the unpredictable part of financial returns. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 17 other part due to other variables that help predict expected economic activity. Table 8 summarizes the value of adding financial news to forecasts by comparing forecasts based on the basic equation (12) to forecasts excluding the estimated news in financial returns. Table 7 is based on rolling-regression estimates of the basic equation (12), and the forecast values from these rolling regressions can be denoted f r. z.23 The superscripts r and z reflect the financial returns, r, and the other variables, z, included in the regressions. Rolling regressions also can be estimated without the recent returns on financial assets, and these regressions can be used to generate forecast values that can be denoted f z. What is the additional value of using the returns to make forecasts? A standard way of combining forecasts is to regress the actual values of the series on the two forecasts of the series, with the two coefficients reflecting the relative value of the two forecasts.24 In our application, the two forecasts are correlated because they are based on the common base set of variables, z; the coefficient on the forecast including financial returns, f r.z, is the value of that forecast, not the marginal improvement in forecasts by adding financial returns. To estimate the more informative improvement in forecasts by adding financial market returns, we estimate the parts of the forecast f r.z from rolling regressions that are uncorrelated with f z and include them in the forecast regressions with f z.25 Table 8 presents the results of estimating these regressions for horizons of one to twelve months.26 If the forecasts were unbiased, the constant terms would be zero and the sums of the two coefficients would be one. Because the forecasts underlying Table 8 are based on rolling regressions, the constant terms in Table 8 need not be zero and the sums of coefficients need not add up to one; the regressions generally do not satisfy these restrictions. With variation by horizon, the news in financial returns is useful for all the series. For example, financial news is uninformative for changes in the unemployment rate over the next month and is more informative than other variables over the next year. The overall picture in Table 8 is one of financial news being informative about the future in addition to, and often more than, the other variables. Conclusions and Discussion he evidence in this article shows that movements in financial markets do presage developments in the economy. In one sense, this evidence is not surprising. Expected returns on stocks and bonds are affected by developments in the economy, and it is impossible for those developments to affect future expected returns without affecting prices and current returns. In another sense, the results are surprising. Evidence on the connection between movements in financial markets and the economy is mixed, with conclusions that typically do not survive scrutiny in a succeeding paper or the passage of time. This article provides evidence of exactly that pattern: Lamont’s (2001) evidence that industry returns are useful complements of the market return is not borne out by the experience of the 1990s. What is to say that the results of this study will hold up? Our conclusion is extremely general: Returns on financial markets are informative about future developments in the economy. We believe this conclusion is unlikely to be affected by variations in technique or the passage of time, but only future research and time will tell. As it stands, the evidence indicates that news revealed in financial markets helps to predict future economic activity. Whether the passage of time will be kind to other conclusions also remains to be seen. We find that asset returns are informative about both real developments, such as industrial production, and inflation. We find that returns on both stocks and bonds are informative about future economic activity and that industry returns are no more informative about future economic activity than is the overall market return. We also find that forecasts based on data actually available before a forecast is made are noticeably less informative than is suggested by computed forecasts based on subsequent data available later. The deterioration of forecasts with rolling regressions compared to forecasts based on all the data is not inevitable. The reasons for such deterioration, other than the trivial one of using fewer observations, are likely to be informative about the relationship between asset returns and economic activity, which can itself inform knowledge about asset returns and about the economy. T 23. The subscript t is suppressed for notational simplicity. 24. Diebold et al. (1996) provide a convenient summary of the literature. 25. Operationally, the additional information in the basic equation due to financial market returns is estimated by the residuals from a regression of f r. z on f z. 26. Reported standard errors are calculated using the Newey and West (1987) correction with a truncation lag of twelve months. 18 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 APPENDIX Data and Sources his appendix presents details about the data and the sources. All growth rates are continuously compounded. T Economic Activity • The growth rate of industrial production is the change in the logarithm of total production, seasonally adjusted. Industrial production by sectors is included for manufacturing production, consumer durables, consumer nondurables, mining, and utilities. These series are seasonally adjusted and are from Data Research Inc. (DRI). • The growth rate of consumption is the change in the logarithm of total real consumption. Consumption also is analyzed for the component parts, consumption of durable goods and consumption of nondurable goods and services. All series are seasonally adjusted and expressed in real terms using deflators from DRI for the corresponding part of total consumption. • The growth rate of real labor income is the change in the logarithm of personal income from wages and salaries, seasonally adjusted, minus the inflation rate measured by the consumer price index (CPI) for all consumers. Personal income from wages and salaries and the inflation rate are from DRI. • The unemployment rate is the total unemployment rate for all workers sixteen years and over, seasonally adjusted. This series is from DRI. • The inflation rate is the change in the logarithm of the CPI for all urban consumers, not seasonally adjusted. This series is from DRI. • The excess return on the CRSP value-weighted index is the continuously compounded return on the CRSP value-weighted index minus the continuously compounded return on Treasury bills. Both series are from CRSP. • The excess return on long-term government bonds is the continuously compounded return on a portfolio of long-term government bonds minus the continuously compounded return on Treasury bills. The return on long-term bonds is from Ibbotson Associates, and the return on Treasury bills is from CRSP. • The return on Treasury bills is the continuously compounded return on a Treasury bill from CRSP. Returns • The aggregate stock price index is the CRSP NYSE-AMEX-Nasdaq value-weighted stock market portfolio. • Industry portfolios are computed in two ways. The returns in the paper are based on industries calculated as in Lamont (2001) and Sharpe (1982). For each year, the industry indexes are based on every NYSE-AMEX-Nasdaq stock being assigned to an industry portfolio based on its four-digit standard industrial classification (SIC) code at the end of June. This classification then is used for returns computed until the following June. Returns are then computed from the end of the month to the end of the next month. The industry definitions are from Sharpe (1982). The eight industry stock portfolios (and their SIC codes) are finance (6000– 6999), utilities (4800–4829, 4900–4999), transportation (3720–3799, 4000–4799), energy (1300–1399, 2900–2999), basic industries (1000–1299, 1400–1499, 2600–2699, 2800– 2829, 2879–2899, 3300–3399), capital goods (3400–3419, 3440–3599, 3670–3699, 3800– 3849, 5080–5089, 5100–5129, 7300–7399), construction (1500–1999, 2400–2499, 3220– 3299, 3430–3439, 5160–5219), and consumer goods (0000–0999, 2000–2399, 2500–2599, 2700–2799, 2830–2869, 3000–3219, 3420–3429, 3600–3669, 3700–3719, 3850–3879, 3880–3999, 4830–4899, 5000–5079, 5090–5099, 5130–5159, 5220–5999, 7000–7299, 7400–9999). French’s classification of industries produces similar results. For each year, the industry indexes are based on every NYSE-AMEXNasdaq stock being assigned to an industry portfolio based on its four-digit SIC code at the end of June. This classification then is used for returns computed until the following June. Returns are then computed from the end of the month to the end of the next month. The five industry stock portfolios (and their SIC codes) are manufacturing (2000–3999), utilities (4900–4999), shops (wholesale, retail, and some services (5000–5999, 7000–7999), finance (6000–6999), and other. The five industry stock portfolios were downloaded from Kenneth French’s Web page <mba.tuck.dartmouth.edu/ pages/faculty/ken.french/data_library.html> (June 12, 2003). Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 19 A P P E N D I X (continued) TABLE A Forecast Accuracy Stock Market, Industry, and Bond Returns Horizon Measure of economic activity 20 1 Month 3 Months 6 Months Industrial production total Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.58 0.42 0.04 0.32 2.71 1.56 0.13 0.50 7.96 3.54 0.09 0.59 23.72 7.14 –0.06 0.68 Industrial production manufacturing Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.70 0.51 0.07 0.31 3.23 1.89 0.16 0.51 9.24 4.14 0.15 0.62 28.22 8.83 –0.01 0.68 Industrial production of consumer durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 6.75 5.24 –0.10 0.15 23.07 14.45 –0.08 0.32 43.26 22.83 0.01 0.48 84.97 33.63 –0.00 0.60 Consumption of durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 10.72 8.51 –0.13 0.10 17.49 12.32 –0.11 0.21 25.47 14.64 –0.09 0.37 57.79 24.48 –0.54 0.35 Unemployment rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.03 0.02 0.03 0.25 0.11 0.07 0.22 0.52 0.32 0.15 0.21 0.63 0.93 0.33 0.22 0.72 Inflation rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.07 0.05 0.35 0.52 0.38 0.22 0.44 0.68 1.26 0.55 0.45 0.76 7.02 1.59 0.14 0.80 Excess stock return Rolling MSE In-sample MSE Rolling R2 In-sample R2 23.66 19.18 –0.10 0.10 76.92 57.59 –0.14 0.14 167.20 99.19 –0.27 0.25 367.55 179.33 –0.38 0.35 Excess bond return Rolling MSE In-sample MSE Rolling R2 In-sample R2 8.89 6.96 –0.04 0.19 30.86 22.60 –0.10 0.19 60.62 38.77 –0.14 0.27 149.11 62.27 –0.26 0.47 Treasury bill return Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.01 0.01 0.82 0.87 0.07 0.05 0.82 0.88 0.39 0.21 0.75 0.86 2.11 0.78 0.64 0.86 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 12 Months • Four bond returns are also included in the regressions. The bond returns are for a longterm government bond, an intermediate-term government bond, a one-year government bond, and a high-grade corporate bond. The long-term, intermediate-term, and high-grade bond returns are from Ibbotson Associates (see the Stocks, Bonds, Bills, and Inflation 2003 Yearbook for further details). The one-year bond return is from CRSP. All rates of return in the regressions are excess returns relative to the one-month Treasury bill return from CRSP and are based on month-end bid-ask average values. Additional Variables Estimates of the expected value of economic activity and excess returns are based on a constant term and lagged values of variables. Eight lagged variables are included that have been used in tests of multiple-beta models and studies of stock-bond return predictability in, among others, Chen, Roll, and Ross (1986); Burmeister and McElroy (1988); Ferson and Harvey (1991, 1999); Downs and Snow (1994); Kirby (1998); Balduzzi and Robotti (2001); and Lamont (2001). Some of these lagged variables span the prior twelve months and some span only the prior month. Variables for the prior twelve months • The inflation rate for the prior twelve months is the change over the prior twelve months in the logarithm of the CPI, not seasonally adjusted, from DRI. • The excess aggregate stock return is the aggregate return (including dividends) on the NYSEAMEX-Nasdaq value-weighted stock market portfolio from CRSP for the prior twelve months minus the one-month Treasury bill return over the past twelve months from CRSP. • The growth rate of industrial production for the prior twelve months is the change in the logarithm of total production, seasonally adjusted, from DRI. Variables for the prior month • The prior month’s term premium on one-year Treasury securities is the yield on the one-year constant maturity note from Global Financial Data Inc. minus the thirty-day Treasury bill yield from CRSP. • The prior month’s long-term premium is the yield on long-term government bonds minus the one-month Treasury bill yield. The yield on long-term government bonds is from Ibbotson Associates, and the Treasury bill yield is from CRSP. • The return on a one-month Treasury bill is from CRSP. • The default premium on short-term debt is the yield on commercial paper minus the one-month Treasury bill yield. The commercial paper rate is from various issues of Banking and Monetary Statistics, Annual Statistical Digest, and Domestic Financial Statistics. The one-month Treasury bill rate is from CRSP. • The default premium on corporate securities is the BAA yield on corporate debt minus the AAA yield on corporate debt. Both series are from DRI. • The prior month’s dividend yield is the annualized dividend yield on the S&P 500 composite common stock. The series is from DRI. TABLE B Forecast Accuracy Industry and Bond Returns Horizon Measure of economic activity 1 Month 3 Months 6 Months 12 Months Industrial production total Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.57 0.42 0.06 0.31 2.66 1.57 0.15 0.50 7.85 3.55 0.10 0.59 23.35 7.15 –0.04 0.68 Industrial production manufacturing Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.69 0.52 0.08 0.31 3.19 1.90 0.17 0.50 9.10 4.16 0.16 0.62 27.72 8.83 0.01 0.68 (continued) Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 21 A P P E N D I X (continued) T A B L E B (continued) Horizon Measure of economic activity 22 1 Month 3 Months Industrial production of consumer durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 6.68 5.24 –0.09 0.15 23.03 14.51 –0.08 0.32 43.07 23.06 0.02 0.47 82.21 33.63 0.03 0.60 Consumption of durable goods Rolling MSE In-sample MSE Rolling R2 In-sample R2 10.57 8.51 –0.12 0.10 17.35 12.32 –0.10 0.21 25.19 14.65 –0.08 0.37 57.28 24.50 –0.53 0.35 Unemployment rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.03 0.02 0.05 0.25 0.10 0.07 0.24 0.52 0.32 0.15 0.21 0.63 0.92 0.33 0.22 0.72 Inflation rate Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.07 0.05 0.36 0.52 0.38 0.22 0.44 0.68 1.25 0.56 0.45 0.76 6.89 1.59 0.16 0.80 Excess stock return Rolling MSE In-sample MSE Rolling R2 In-sample R2 23.40 19.18 –0.09 0.10 76.71 57.97 –0.14 0.14 166.98 99.22 –0.27 0.25 364.82 179.38 –0.37 0.32 Excess bond return Rolling MSE In-sample MSE Rolling R2 In-sample R2 8.81 6.99 –0.03 0.18 30.79 22.69 –0.10 0.19 60.34 38.90 –0.13 0.27 150.11 62.83 –0.27 0.47 Treasury bill return Rolling MSE In-sample MSE Rolling R2 In-sample R2 0.01 0.01 0.82 0.87 0.07 0.05 0.82 0.88 0.39 0.22 0.75 0.86 2.08 0.78 0.64 0.86 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 6 Months 12 Months REFERENCES Balduzzi, Pierluigi, Edwin J. Elton, and Clifton T. Green. 2001. Economic news and bond prices: Evidence from the U.S. Treasury market. Journal of Financial and Quantitative Analysis 36 (December): 523–43. Hayek, Friedrich A. 1948. The use of knowledge in society. American Economic Review 35 (September): 519–30. Kamstra, Mark. 2003. Pricing firms on the basis of fundamentals. Federal Reserve Bank of Atlanta Economic Review 88, no. 1:49–70. Balduzzi, Pierluigi, and Cesare Robotti. 2001. Minimumvariance kernels, economic risk premia, and tests of multi-beta models. Federal Reserve Bank of Atlanta Working Paper 2001-24, November. Kirby, Chris. 1998. The restrictions on predictability implied by rational asset pricing models. Review of Financial Studies 11 (Summer): 343–82. Breeden, Douglas T., Michael R. Gibbons, and Robert H. Litzenberger. 1989. Empirical tests of the consumptionoriented CAPM. Journal of Finance 44 (June): 231–62. Lamont, Owen A. 2001. Economic tracking portfolios. Journal of Econometrics 105:161–84. Browning, E.S., and Aaron Lucchetti. 2003. Prestigious party-poopers persist. Wall Street Journal, November 3. Burmeister, Edwin, and Marjorie B. McElroy. 1988. Joint estimation of factor sensitivities and risk premia for the arbitrage price theory. Journal of Finance 43 (July): 721–33. Newey, Whitney K., and Kenneth D. West. 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica 55 (May): 703–8. Robotti, Cesare. 2002. Asset returns and economic risk. Federal Reserve Bank of Atlanta Economic Review 87, no. 2:13–25. Chen, Nai-Fu, Richard Roll, and Stephen A. Ross. 1986. Economic forces and the stock market. Journal of Business 59 (July): 383–403. Samuelson, Paul Anthony. 1966. Science and stocks. Newsweek, September 19. Sharpe, William F. 1982. Factors in NYSE security returns, 1931–1979. Journal of Portfolio Management 8, no. 2:5–19. Del Negro, Marco. 2001. Turn, turn, turn: Predicting turning points in economic activity. Federal Reserve Bank of Atlanta Economic Review 86, no. 2:1–12. Diebold, Francis X., Jose A. Lopez, G.S. Maddala, and Calyampudi R. Rao. 1996. Forecast evaluation and combination. In Handbook of Statistics Series, vol. 14, 241–68. Amsterdam, N.Y., and Oxford: Elsevier, North-Holland. Siegel, Jeremy J. 1998. Stocks for the long run: The definitive guide to financial market returns and long-term investment strategies. 2nd ed. New York: McGraw-Hill. Smith, Stephen D. 1999. What do asset prices tell us about the future? Federal Reserve Bank of Atlanta Economic Review 84, no. 3:4–13. Downs, David H., and Karl N. Snow. 1994. Sufficient conditioning information in dynamic asset pricing. University of North Carolina, unpublished paper. Stock, James H., and Mark W. Watson. 2003. Forecasting output and inflation: The role of asset prices. Journal of Economic Literature 41, no. 9:788–829. Dwyer, Gerald P., Jr., and R.W. Hafer. 1989. Interest rates and economic announcements. Federal Reserve Bank of St. Louis Review 71, no. 2:34–46. Ferson, Wayne E., and Campbell R. Harvey. 1991. The variation of economic risk premiums. Journal of Political Economy 99 (April): 385–415. Stock, James, Mark W. Watson, Oliver Jean Blanchard, and Stanley Fischer. 1989. New indexes of coincident and leading economic indicators. In NBER macroeconomics annual, 351–94. Cambridge, Mass.: MIT Press. ———. 1999. Conditioning variables and the cross section of stock returns. Journal of Finance 54 (August): 1325–60. Veronesi, Pietro, and Tano Santos. 2001. Labor income and predictable stock returns. NBER Working Paper 8309, May. Flannery, Mark J., and Aris A. Protopapadakis. 2002. Macroeconomic factors do influence aggregate stock returns. Review of Financial Studies 15, no. 3:751–82. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 23 Leading Indicators of Country Risk and Currency Crises: The Asian Experience MARCELLE CHAUVET AND FANG DONG Chauvet is a research economist at the Atlanta Fed. Dong is an assistant professor at Providence College in Rhode Island. n recent years, capital restrictions in emerging markets have been substantially reduced. As a result, international financial flows to these countries have risen. Most emerging markets have adopted a pegged exchange rate system in which central banks are committed to keeping their domestic currency in terms of the U.S. dollar within narrow bands. Under this system, a country can finance a current account deficit from its reserves or by borrowing from abroad. That is, the country can buy time in handling external deficits without decreasing the monetary base or reducing the public deficit. Such a regime relies on a delicate balance and makes a country vulnerable to shocks in mobile international capital markets, especially with respect to outflows in bank deposits. When international markets are relatively calm, lenders may be willing to finance countries with mildly weak fundamentals. As international conditions deteriorate, however, investors’ perception about a borrower’s creditworthiness may change. Economies that look sound one moment seem riskier the next— not necessarily because of new developments within their borders but perhaps because interconnected countries are in distress. As foreign investors become more risk averse, they may withdraw short-term investments and sell local currency. The country’s central bank must then increase interest rates sufficiently to dampen the outflow and avoid a collapse of the pegged exchange rate system. The result of I such reactive strategies may be a credit crunch that spreads from country to country, driving each into economic recessions with high inflation. In the last decade several developed and developing countries experienced currency crises. For example, the European Monetary System (EMS) was severely undermined by intense speculative pressure in 1992–93, which led to the exit of Britain and Italy in 1992. More recently, several emerging market economies underwent large devaluations of their currencies: Mexico in 1994, several Asian countries in 1997, Russia in 1998, and, subsequently, Brazil in 1999, among others. These events cast a bleak outlook for the global financial system and caused widespread economic distress. Even the U.S. economy experienced slowdowns associated with these international events, especially the Mexican and Asian crises. A “country risk” of currency crisis is not directly observable, but prior currency pressures can be detected in several sectors of the economy. In particular, financial variables reflecting investors’ expectations and banking distress are highly sensitive to changes in the economic environment. This article aims to construct an early warning system for international currency crises using such variables. The system uses a dynamic factor model with regime switching to construct leading indicators of country risk and currency crises. In this model, an unobservable factor switches regimes, representing periods Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 25 of relative calmness and periods prone to currency crises, using a two-state Markov process. The method is applied to evaluate the model’s in-sample and out-of-sample performance in anticipating currency crises in the last two decades in Thailand, Indonesia, and Korea. The dynamic factor index gives early distress signals of country risk and currency crisis, using several financial and banking variables. Leading indicators have been a successful forecasting tool adopted by the National Bureau of Economic Research (NBER) since the work of Burns and Mitchell (1946). New econometric models have now been used to explore more formally potential dynamic differences across cycle phases in several This article constructs an early warning system for international currency crises using financial variables that reflect investors’ expectations and banking distress. variables. The method used to construct economic indicators is distinct from econometric regression methods. In particular, the goal is not to form a forecast of exchange rates based on the information set. Instead, leading indicators are indexes composed of several variables, designed to give early signals of major cyclical changes in exchange rates, particularly the beginning and end of cyclical phases (that is, their turning points). Variables that exhibit low power in explaining the linear long-run variance of exchange rates may be highly important in specific situations. In fact, unusually large changes in some variables at particular historical episodes—as opposed to the linear average behavior of the series— can be important independent factors in determining large exchange rate devaluations. A large theoretical and empirical literature aims to characterize or forecast the recent experiences of currency crises.1 Few of these studies, however, focus on forecasting turning points representing episodes of speculative attacks. The method this study uses to construct indicators differs from the previous currency crisis literature in several ways. First, since currency crises are caused by different shocks over time, the inclusion of different variables increases the model’s ability to signal future crises. In addition, the combination of variables reduces measurement errors in the individual series 26 and smooths out noise inherent in monthly data. This smoothing reduces the likelihood of signaling false turning points, which can be a significant problem in the monthly frequency. Second, in contrast to composite indicators that are constructed as weighted averages of statistical transformations of their components, the dynamic factor model takes into account cross-correlations and potential long-term relationships among the variables. Finally, the method yields probabilities that can signal turning points in real time. This method contrasts with the rules of thumb used to build some composite indicators, which require the use of substantial ex post data. Because these rules are based on the unusual behavior of some variables compared to their frequency distribution, turning points can be identified and predicted only a couple of months after their occurrence, which undermines their usefulness for real-time forecasting. Thus, the advantage of the proposed approach in comparison with alternative models and rules of thumbs is that it treats foreign exchange market regimes as unobservable priors instead of observed ex post events, and no ad hoc criterion is adopted in determining the crisis state. Instead, the model generates regime probabilities from the leading indicators that can be used to signal increases in country risk and potential currency crises in real time. The approach in this article implements several linear and nonlinear methods to select the variables composing the indicators. For the Asian countries studied, the best candidates are monetary and banking series. The study shows that the leading indicators built from the nonlinear dynamic factor model unveil, both in sample and out of sample, early warning signals of an increase in the country risk and subsequent depreciation of nominal exchange rates experienced by Thailand, Indonesia, and Korea, especially before the 1997 crisis. In general, phases of the leading index exhibiting a higher mean and volatility precede currency crises, whereas the noncrisis state is associated with a lower mean and volatility. For all the countries studied, the regime probabilities give early signals of the 1997 crisis and reveal a contagion pattern. For Thailand, a crisis was signaled six months earlier than the actual one. For Indonesia, the probabilities indicated a crisis seven months before the actual one, which was minimized by preemptive government actions. However, once Thailand’s currency crisis hit, the probability of a crisis in Indonesia also increased substantially and thus increased the probability of a crisis in Korea. This finding suggests a contagion pattern that is being further examined in ongoing projects. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 The article first discusses the currency crises experienced by the Asian countries studied. The discussion then presents the data and statistical analysis used to select the leading variables, presents the dynamic factor model used to construct the leading indicators, and reports the in-sample and out-of-sample empirical results. Currency Crises in Asia adelet and Sachs (1998) study the broad features underlying the recent experiences of currency crises in Asian countries. One striking finding is that typical international and domestic problems were not present before the onset of the crises. In fact, for the most part conditions in international financial markets, commodity markets, and the trading system were favorable. These countries were not pursuing tight anti-inflationary policies, and their real exchange rates were only mildly overvalued because of the persistent inflow of capital. In addition, their overall debt-carrying capacities did not seem to present imminent risks of default. In particular, Radelet and Sachs find that instability in international lending and self-fulfilling speculative attacks are the most likely explanations for the Asian crisis in 1997. International loan markets may be subject to selffulfilling crises even when individual creditors act rationally. Changes in investors’ risk perception may result in sharp, costly, and fundamentally unnecessary panicked reversals in capital flows. In this situation, exchange rates may immediately depreciate under intense pressure. The unwillingness or inability of the capital market to provide new loans to the illiquid borrowers is a chief factor during crises. Another common feature of these countries prior to the crises was the growing weaknesses in East Asian financial systems resulting from incomplete markets and some market-oriented reforms, which made the countries vulnerable to capital flight. In this regard, the intensity and propagation of the crises were also the result of partial banking and financial reforms that exposed these economies more directly to the instability of international financial markets. Examples of bank weaknesses were the growth of short-term foreign debt, the rapid expansion of bank credit/lending, the inadequate regulation and supervision of financial institutions, and the sharp R increase in the number of financial institutions and private banks (including foreign and joint venture banks) that could borrow or lend in foreign currencies, both on- and offshore.2 These problems made the countries more vulnerable to a rapid reversal of capital flows that put downward pressure on their currencies. Whereas Radelet and Sachs (1998) find that the problems were centered in the private sector rather than in the government, this article finds that they were also present in the monetary system. Thailand. Three major currency devaluations in Thailand occurred during the 1981:05–1981:07, 1984:11–1985:03, and 1997:07–1998:01 periods.3 This study demonstrates that the leading indicators of currency crises can be informative tools for signaling future currency crises in real time and could thus allow preemptive counterpolicy measures by the central bank. These devaluations of the baht are illustrated in Figure 1, which plots Thailand’s nominal exchange rate in the form of logarithmic first differences (GW_N$BAHT). During the 1990s, capital inflows into Thailand averaged over 10 percent of gross domestic product (GDP) and reached a remarkable 13 percent of GDP in 1995 alone. These inflows consisted predominantly of borrowing by banks and financial institutions. Throughout the decade the government fixed the exchange rate within very narrow bands. In effect, the central bank absorbed the risks of exchange rate movements on behalf of investors and thus encouraged capital inflows, especially of short-maturity instruments. However, increasing capital inflows put upward pressure on the prices of nontradable goods and services. The real effective exchange rate appreciated by more than 25 percent between 1990 and early 1997. Indonesia. Three major currency devaluations in Indonesia occurred in April 1983, September to 1. See, for example, the list of more than 100 recent papers and books related to the NBER Project on Exchange Rate Crises in Emerging Market Countries at <www.nber.org/crisis> or the reference list at <www.stern.nyu.edu/globalmacro>. 2. State-owned banks in Indonesia and Korea were regularly allowed to break many prudential regulations without penalty. 3. During the 1984:11–1985:03 period, Thailand abandoned a fixed exchange rate vis-à-vis the dollar. The central bank abolished general credit restrictions but reimposed restrictions on bank lending rates and lowered the ceiling for loans to priority sectors (see Bekaert and Harvey 1999). Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 27 FIGURE 1 Thailand’s Nominal Exchange Rate 20 97:07–98:01 15 10 84:11–85:03 Percent change 81:05–81:07 5 0 –5 –10 –15 –20 1980 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database FIGURE 2 Indonesia's Nominal Exchange Rate 40 97:08–98:12 20 83:04 86:09–86:10 Percent change 0 –20 –40 –60 –80 1980 1983 1986 1989 Source: Datastream, International Financial Statistics database 28 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 1992 1995 1998 FIGURE 3 Korea's Nominal Exchange Rate 20 97:11–98:01 10 Percent change 0 –10 –20 –30 –40 1980 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database October 1986 (Sachs, Tornell, and Velasco 1996), and August 1997 to December 1998. The devaluations of the rupiah are shown in Figure 2, which plots Indonesia’s nominal exchange rate in the form of logarithmic first differences (GW_N$RUPIAH). Capital inflows to Indonesia in the 1990s averaged a more modest 4 percent of GDP and were mostly in the form of borrowing by private corporations. Indonesia’s government fixed the exchange rate subject to small and predictable changes. Here too the government absorbed the borrowing risks undertaken by the private sector, inducing higher inflows of capitals. As a result, the real effective exchange rate appreciated by more than 25 percent between 1990 and early 1997. Korea. The only major nominal devaluation of the Korean won was related to the Asian crisis, which hit the country in November 1997. Annual capital inflows averaged over 6 percent of GDP between 1990 and 1996. The government maintained the exchange rate with small and predictable changes and absorbed the loan risks. The real effective exchange rate appreciated by 12 percent between 1990 and early 1997. Figure 3 plots the logarithmic first differences of Korea’s nominal exchange rate (GW_N$WON). Data and Statistical Analysis S election of candidate leading variables. In the first triage, the variables were selected according to several criteria, such as their frequency, sample size, and how quickly new releases of the series were available. For these indicators to be useful for real-time forecasting of currency crises, the variables used should be available at least at the monthly frequency and be timely.4 We found approximately ten variables for each country as potential candidates to predict abrupt changes in nominal exchange rates. Several econometric procedures were then used to select and rank the potential variables. First, all series were transformed to achieve stationarity.5 The variables were then classified according to their cross-correlation with nominal exchange rates and 4. For example, although some series are available monthly, their release takes place two to three months later. 5. A variable is said to be (weakly) stationary if the mean and autocovariances of the series do not depend on time. Any series that is not stationary is said to be nonstationary. The augmented Dickey-Fuller (1979) and Phillips-Perron (1988) tests were used to test for stationarity. In addition, Perron’s (1989) test was also used to test for nonstationarity against the alternative of deterministic trend in the presence of sudden changes in the series. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 29 FIGURE 4 Thailand: Domestic Credit, Net Foreign Assets, Private Bank Credits, and the Consumer Price Index Domestic credit Net foreign assets 8 150 6 100 Ab s o lu t e c h a n g e Pe r c e n t c h a n g e 97:07–98:01 4 2 0 81:05–81:07 –2 84:11–85:03 50 0 81:05–81:07 –50 84:11–85:03 –100 –4 97:07–98:01 –6 1980 1983 1986 1989 1992 1995 1998 –150 1980 1983 Private bank credits 1989 1992 1998 1995 Consumer price index 3 97:07–98:01 97:07–98:01 81:05–81:07 50 Ab so lu te c h a n ge Ab so lu te c h a n ge 100 1986 0 81:05–81:07 –50 84:11–85:03 2 84:11–85:03 1 0 –100 –150 –1 1980 1983 1986 1989 1992 1995 1998 1980 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database their ability to Granger-cause exchange rates.6 Granger causality tests select variables that have a linear predictive content for exchange rates, but not necessarily those that perform well in anticipating peaks and troughs in exchange rate changes. Variables that are poor predictors of linear long-run exchange rate variances may be significant in particular situations. Large changes in such variables during specific historical episodes can be important in predicting large exchange rate devaluations. For this reason, we use probability methods to study the nonlinear relationship of each series to determine whether it anticipates peaks and troughs of exchange rate dynamics. In particular, different specifications of two-state first-order Markov switching models were fitted to each candidate leading variable (see Chauvet and Dong 2002). The following leading variables were selected from both linear and nonlinear procedures: (1) for Thailand, domestic credit, net foreign assets and private bank credits from the central bank in billions of baht, and the consumer price index (CPI) (1995 = 100) (see Figure 4); (2) for Indonesia, the 30 money supply (M1), net foreign assets and private bank foreign liabilities in billions of rupiah, and official foreign reserves minus gold in millions of U.S. dollars (Figure 5); and (3) for Korea, domestic credit, net foreign assets and private bank credits from monetary authorities in billions of won, and the CPI (Figure 6). These data were obtained from the International Financial Statistics database from Datastream. The sample available in monthly frequency covers the 1980:01–1999:06 period for Indonesia and Korea and the 1980:02–1999:06 period for Thailand. Nominal exchange rates are measured in U.S. dollars per unit of the national currency. The Dynamic Factor Model with Markov Regime Switching his analysis uses a dynamic factor model with Markov regime switching to construct the leading indicators of currency crises for Thailand, Indonesia, and Korea.7 This model is a combination of the linear Kalman filter and Hamilton’s (1989) Markov regime switching model and has been widely T Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 FIGURE 5 Indonesia: M1, Net Foreign Assets, Private Bank Foreign Liabilities, and Foreign Reserves M1 Net foreign assets 20 80 83:04 83:04 97:08–98:12 86:09–86:10 10 5 0 –5 40 20 0 –20 –10 –40 1980 –15 1980 1983 1986 1989 1992 1995 1998 1983 1986 Private bank foreign liabilities 100 83:04 86:09–86:10 1989 1992 1995 1998 Foreign reserves 30 97:08–98:12 80 83:04 97:08–98:12 86:09–86:10 20 P e rc e n t c h a n ge 60 P e rc e n t c h a n ge 97:08–98:12 86:09–86:10 60 Pe r c e n t c h a n g e Pe r c e n t c h a n g e 15 40 20 0 –20 –40 –60 10 0 –10 –20 –30 –40 –80 –100 –50 1980 1983 1986 1989 1992 1995 1998 1980 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database applied to business cycle studies (see, for example, Diebold and Rudebusch 1996; Chauvet 1998; Kim and Nelson 1998). In this framework, the latent factor for each country—the leading indicator—is constructed as the common correlation underlying the country’s leading financial variables. The motivation for this setup is to combine the leading variables and extract their common characteristics, which switch regimes representing foreign exchange market pressures. The mean and variance of the dynamic factor are subject to discrete regime shifts governed by a two-state Markov process. That is, the foreign exchange market can be either under high pressure to devaluate (state or regime 0) or under low speculative pressure (state or regime 1), with the alternation between states controlled by the outcome of the Markov process. Since the probabilistic infer- ence on crises is based on shocks to several leading variables used for each country, the model used here can give more accurate signals of crises (fewer false or missed signals) than univariate autoregressive models with Markov regime switching. (See Chauvet and Dong 2002 for further discussions.) In-Sample Results M aximum likelihood estimates. Table 1 reports the maximum likelihood estimates of the Markov-switching dynamic factor model for Thailand, Indonesia, and Korea. For each country, the analysis shows that regime 0 (high speculative pressure) is characterized by a large variance. For Thailand, estimation shows that the net foreign asset (NFA) variable is the most sensitive to changes in the country’s leading indicator. A oneunit increase in the factor is associated with a 6. A Granger causality test determines how much of a current time series can be explained by past values of itself and whether adding lagged values of another series can improve the explanation. 7. A Markov process is a simple stochastic process in which the distribution of future states depends only on the present state and not on how the present state was achieved. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 31 FIGURE 6 Korea: Domestic Credit, Net Foreign Assets, Private Bank Credits, and the Consumer Price Index Domestic credit 97:11–98:01 Pe r c e n t c h a n g e 6 4 2 0 –2 –4 1980 1983 1986 1989 1992 1995 Absolute change (in thousands) Net foreign assets 8 12 97:11–98:01 8 4 0 –4 –8 1998 1980 1983 Private bank credits 1986 1989 1992 1995 1998 Consumer price index 100 3.0 80 2.5 Ab so lu te c h a n ge P e rc e n t c h a n ge 97:11–98:01 60 40 20 0 2.0 1.5 1.0 0.5 0.0 –0.5 –20 97:11–98:01 1980 1983 1986 1989 1992 1995 –1.0 1998 1980 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database monthly decrease in NFA of about 5 billion baht, ceteris paribus. On the other hand, the CPI variable is the least sensitive to changes in the leading indicator. The leading indicator for Thailand is highly persistent, with an autoregressive coefficient equal to 0.91. In the crisis state, the volatility of the leading indicator is about 256 times larger than in the normal or noncrisis state. For Indonesia, the private bank foreign liabilities (PBFL) variable is the most sensitive to changes in the country’s leading indicator. A one-unit increase in the factor is associated with a monthly increase in PBFL of 2.72 percent. The reserves variable, with a factor coefficient of 0.47 percent, is not as sensitive as other variables. The leading indicator for Indonesia is somewhat persistent, with an autoregressive coefficient of –0.64. In the crisis state, the volatility of the leading indicator is about 31 times greater than in the noncrisis state. For Korea, the NFA variable is the most sensitive to changes in the factor; a one-unit increase in the factor is associated with a monthly increase in NFA of 212.39 billion won. As in Thailand, CPI is the 32 least sensitive series to the factor, with a factor coefficient of about 0.30 percent. The leading indicator for Korea is highly persistent, with an autoregressive coefficient of 0.92. The volatility of the leading indicator is about 364 times larger in the crisis state than in the noncrisis state. Table 2 shows that, for all three countries, the leading indicator of currency crisis is negatively correlated with exchange rates. That is, increases in the level of the leading indicator are associated with currency depreciation. The currency crises for all countries are anticipated by the dynamic factor behavior in state 0, that is, for the high-mean and high-volatility regime. Variables such as NFA, private bank credits from the central bank, and PBFL are the most useful in signaling speculative pressures and currency crises in these three countries. Crises would also be anticipated with a smaller lead by internal macroeconomic fundamentals such as domestic credits, the money supply, the CPI, or foreign reserves. This finding supports evidence that the currency crises across these three countries are likely to have originated in Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 TA B L E 1 Maximum Likelihood Estimates: Dynamic Factor Model with Regime Switching Thailand Indonesia Korea α̂0 –0.0756 (0.2190) α̂0 4.9982 (2.0785) α̂0 0.2075 (0.7200) α̂1 0.1243 (0.0550) α̂1 1.8668 (0.3120) α̂1 0.0950 (0.0340) φ̂1 0.9133 (0.0384) φ̂1 –0.6442 (0.1729) φ̂1 0.9181 (0.0257) σ̂ 2υ0 9.0228 (3.9717) σ̂ 2υ0 28.5281 (10.9832) σ̂ 2υ0 5.5735 (3.8378) σ̂ 2υ1 0.0352 (0.0140) σ̂ 2υ1 0.9139 (0.6835) σ̂ 2υ1 0.0153 (0.0085) P̂00 0.9277 (0.0680) P̂00 0.8641 (0.1083) P̂00 0.8835 (0.1352) P̂11 0.9933 (0.0069) P̂11 0.9810 (0.0188) P̂11 0.9903 (0.0069) 2 σ̂ GW_ DC 0.5136 (0.0689) 2 σ̂ CH_ NFA 622.7890 (59.0376) 2 σ̂ CH_ PB 242.5626 (22.6699) 2 σ̂ CH_ CPI 0.1661 (0.0166) 2 σ̂ GW_ M1 4.2514 (0.9551) 2 σ̂ GW_ NFA 91.1092 (8.7597) 2 σ̂ GW_ PBFL 276.6354 (27.8780) 2 σ̂ GW_ RESV 38.3251 (3.6027) 2 σ̂ GW_ DC 2 σ̂ CH_ NFA 1.2626 (0.1473) 2431783.9236 (226865.0648) 2 σ̂ GW_ PB 52.4856 (4.9980) 2 σ̂ CH_ CPI 0.0797 (0.0093) λ̂GW_ DC 1.0000 (Restricted) λ̂GW_ M1 1.0000 (Restricted) λ̂GW_ DC 1.0000 (Restricted) λ̂CH_ NFA –4.9412 (1.1487) λ̂GW_ NFA 0.7839 (0.3134) λ̂CH_ NFA 212.3871 (75.3244) λ̂CH_ PB 2.6914 (0.6788) λ̂GW_ PBFL 2.7161 (0.5460) λ̂GW_ PB 1.6953 (0.3660) λ̂CH_ CPI 0.2096 (0.0191) λ̂GW_ RESV 0.4683 (0.1793) λ̂CH_ CPI 0.2951 (0.0205) Note: The sample period is 1980:01–1999:06. Asymptotic standard errors (computed numerically) appear in parentheses. The factor mean for crisis state is µ̂0 = α̂ 0 /(1– φ̂1), and for off-crisis state it is µ̂1 =α̂1 /(1– φ̂1). TA B L E 2 Correlation of Factor with Exchange Rate and Leading Indicators Thailand N$BAHT GW_DC CH_NFA CH_PB CH_CPI Indonesia –0.6471 0.7845 –0.6530 0.5089 0.2304 N$RUPIAH N$BAHT GW_M1 GW_NFA GW_PBFL GW_RESV Korea –0.4762 –0.3022 0.8823 0.2171 0.4600 0.1911 N$WON N$BAHT GW_DC CH_NFA GW_PB CH_CPI –0.7083 –0.4076 0.4401 0.0184 0.4148 0.6318 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 33 FIGURE 7 Thailand: Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered probability of currency crises Filtered dynamic factor 8 1.0 97:07–98:01 97:07–98:01 6 0.8 84:11–85:03 81:05–81:07 Pr o b a b ilit y In d e x 4 2 0 81:05–81:07 0.6 0.4 –2 84:11–85:03 0.2 –4 –6 1980 1983 1986 1989 1992 1995 0.0 1980 1998 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database and model results their respective private financial sectors and monetary sectors as a result of unsustainable financial liberalization policies. For Thailand, in particular, acceleration in the growth rate of domestic credits and increases in the level of private bank credits from the central bank and in the level of the CPI led to increases in the leading indicator of currency crises. Hence, pressures to devalue Thailand’s baht are associated with increases in the dynamic factor and with decreases in the level of NFA. For Indonesia, acceleration in the growth rate of money, NFA, PBFL, and reserves are associated with increases in the factor and, therefore, with the devaluation of Indonesia’s rupiah. For Korea, acceleration in the growth rate of domestic credits and private bank credits from the central bank and increases in the level of NFA and the CPI are associated with increases in the factor and, hence, with the devaluation of Korea’s won. Probabilities of currency crises. Figure 7 plots the dynamic factor (the leading indicator) and the probability of currency crises for Thailand. The leading indicator is quite stable for most of the sample except for the periods prior to the currency crises in 1981:05 and 1997:07, when the factor moves up and down considerably. This pattern can also be observed in the probability of currency crises, which increases substantially in 1981:02 (three months before the 1981:05 currency crisis) and in 1997:01 (six months before the 1997:07 crisis). The factor is less sensitive to the depreciation in 1984:11, when Thailand’s authorities abandoned the fixed exchange rate vis-àvis the dollar. The economy displayed stronger fundamentals during this time and was less susceptible to external shocks. 34 Figure 8 plots the dynamic factor and the probability of currency crises for Indonesia. Both are quite stable, with values close to 0 for most of the sample except around the currency crises. In fact, they display abrupt oscillations in 1986–87, 1989–91, and 1997–98, anticipating the crises. In particular, the factor and probability of currency crises signal the currency crises in 1986:09 and in 1997:08 nine months in advance. On the other hand, the devaluation in 1983:04 was very small. This pattern is also reflected in the probability of currency crises, which indicates weak speculative pressure (around 2 percent in 1982:12). The small probability of currency crises at the end of 1982 reinforces the view that the 1983 devaluation did not originate from strong pressures from the financial sector and was mostly unanticipated. The devaluation in 1986 was much larger in comparison, and the probability of currency crises—ranging from about 11 percent in 1986:06 to 58 percent in 1986:09—gives clear signals of it, indicating stronger speculative pressure. The 1997 devaluation was the most severe one experienced by Indonesia (see Figure 2). The probability of currency crises ranged from 19 percent in 1996:11 to 60 percent in 1997:01—seven months prior to the crisis in 1997:08. After the onset of the crisis, the probability, ranging from 15 percent in 1997:10 to almost 100 percent in 1998:01, indicated continuous speculative pressure. One should note that the probability increased substantially between 1989:07 and 1991:04. During this period Indonesia underwent financial liberalization and experienced fluctuations in capital inflow (a deceleration in portfolio and other short-term Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 FIGURE 8 Indonesia: Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered dynamic factor Filtered probability of currency crises 20 1.0 97:08–98:12 0.8 86:09–86:10 Pr o b a b ilit y 10 In d e x 97:08–98:12 0.9 15 83:04 5 0 86:09–86:10 0.7 0.6 83:04 0.5 0.4 0.3 0.2 –5 0.1 0.0 1980 –10 1980 1983 1986 1989 1992 1995 1998 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database and model results FIGURE 9 Korea: Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered dynamic factor Filtered probability of currency crises 8 1.0 97:11–98:01 97:11–98:01 7 0.9 6 0.8 P ro b a b ility In d e x 5 4 3 2 0.7 0.6 0.5 0.4 1 0.3 0 0.2 –1 0.1 0.0 1980 –2 1980 1983 1986 1989 1992 1995 1998 1983 1986 1989 1992 1995 1998 Source: Datastream, International Financial Statistics database and model results flows and continued growth in foreign direct investment) while interest rates decreased significantly. However, the exchange rates did not succumb to the high speculative pressure in 1989:07–1991:04 because the government made a preemptive policy response to structural changes in capital inflows (see Radelet and Sachs 1998). Figure 9 plots the dynamic factor and probability of currency crises for Korea. Again, the dynamic factor series is quite stable except during the currency crisis in 1997–98. The probability of currency crisis reflects the speculative pressure and possible contagion from the crises in Thailand and Indonesia one month earlier, in October 1997. When the depreciation of the Korean won occurred in November 1997, the probability of currency crisis reached 100 percent. As the exchange rate fluctuation continued into early 1998, the speculative pressure measured by the probabilistic inference reached another peak of 100 percent in 1998:02. Out-of-Sample Results n this section we examine the performance of inferred probabilities in predicting currency crises in an out-of-sample exercise. We compare and evaluate the model performance of ex post forecasts with real-time ex ante forecasts using only data available at the time of the forecast. The parameters were estimated using data up to 1997:01. The in-sample estimates were then used to generate I Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 35 FIGURE 10 Thailand: In-Sample and Out-of-Sample Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered dynamic factor Filtered probability of currency crises 8 1.0 97:07–98:01 84:11–85:03 6 Pr o b a b ilit y 4 In-sample and out-of-sample 2 In d e x 97:07–98:01 0.8 Full 84:11–85:03 0 81:05–81:07 In-sample and out-of-sample 0.6 Full 0.4 –2 0.2 –4 81:05–81:07 0.0 1980 –6 1980 1983 1986 1989 1992 1995 1998 1983 1986 1989 1992 1995 1998 Note: In-sample data cover the 1980:01–1997:01 period; out-of sample data cover the 1997:02–1999:06 period. The full sample covers the 1980:01–1999:06 period. Source: Datastream, International Financial Statistics database and model results FIGURE 11 Indonesia: In-Sample and Out-of-Sample Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered dynamic factor Filtered probability of currency crises 20 1.0 15 P ro b a b ility In d e x 0.8 In-sample and out-of-sample 10 5 0 83:04 –5 83:04 Full 0.4 0.2 0.0 1980 –10 1983 86:09–86:10 0.6 97:08–98:12 86:09–86:10 1980 97:08–98:12 In-sample and out-of-sample Full 1986 1989 1992 1995 1998 1983 1986 1989 1992 1995 1998 Note: In-sample data cover the 1980:01–1997:01 period; out-of sample data cover the 1997:02–1999:06 period. The full sample covers the 1980:01–1999:06 period. Source: Datastream, International Financial Statistics database and model results FIGURE 12 Korea: In-Sample and Out-of-Sample Filtered Dynamic Factor and Filtered Probability of Currency Crises Filtered dynamic factor Filtered probability of currency crises 1.0 97:11–98:01 7 0.9 6 0.8 5 0.7 4 P r ob a bi lit y In d e x 8 In-sample and out-of-sample 3 2 1 In-sample and out-of-sample 0.6 0.5 Full 0.4 0.3 Full 0 97:11–98:01 0.2 0.1 –1 –2 1980 1983 1986 1989 1992 1995 1998 0.0 1980 1983 1986 1989 1992 1995 1998 Note: In-sample data cover the 1980:01–1997:01 period; out-of sample data cover the 1997:02–1999:06 period. The full sample covers the 1980:01–1999:06 period. Source: Datastream, International Financial Statistics database and model results 36 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 out-of-sample forecasts of the filtered probabilities and filtered dynamic factors. The out-of-sample performance is analyzed for 1997:02–1999:06, which is the period that includes the recent Asian currency crises. The dynamic factor model with regime switching successfully captures the crisis through the filtered factor and filtered probability (see Figures 10, 11, and 12). The out-of-sample filtered dynamic factors based on data up to 1997:01 closely mimic the factors based on full-sample data up to 1999:06. The filtered probability of currency crises for Thailand based on information up to 1997:01 signals the country’s currency crisis in 1997:02, that is, five months before the actual crisis occurred. For Indonesia, the probability signals the crisis in 1997:01, seven months before the actual crisis. For Korea, the probability signals a crisis in 1997:11, coinciding with the actual crisis. Conclusions his article uses a dynamic factor model with regime switching to construct leading indicators of currency crises for Thailand, Indonesia, and Korea. The analysis finds that most of the large currency depreciations in these countries during the sample periods can be attributed in great part to the deterioration of monetary and banking sector conditions, which was intensified by speculative pressures. The dynamic factor model successfully produces early probabilistic forecasts of the Asian currency crises, particularly the most severe one, which occurred in 1997. These results hold for both insample and recursive out-of-sample estimation. This study demonstrates that the leading indicators of currency crises can be informative tools for signaling future currency crises in real time and could thus allow preemptive counterpolicy measures by the central bank. T REFERENCES Bekaert, Geert, and Campbell R. Harvey. 1999. Chronology of economics, political and financial events in emerging markets. Columbia University, photocopy. Hamilton, James D. 1989. A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica 57 (March): 357–84. Burns, Arthur, and Wesley Mitchell. 1946. Measuring business cycles. New York: National Bureau of Economic Research. Kim, Chang-Jin, and Charles Nelson. 1998. Business cycle turning points, a new coincident index, and tests of duration dependence based on a dynamic factor model with regime switching. Review of Economics and Statistics 80 (May): 188–201. Chauvet, Marcelle. 1998. An econometric characterization of business cycle dynamics with factor structure and regime switching. International Economic Review 39 (November): 969–96. Chauvet, Marcelle, and Fang Dong. 2002. A framework for modeling country risk and financial crisis contagion. Unpublished paper. Perron, Pierre. 1989. The great crash, the oil price shock, and the unit root hypothesis. Econometrica 57 (November): 1361–1401. Phillips, Peter, and Pierre Perron. 1988. Testing for a unit root in time series regression. Biometrika 75 (June): 335–46. Dickey, David, and Wayne Fuller. 1979. Distribution of the estimators for autoregressive time series with a unit root. Journal of the American Statistical Society 74: 427–31. Diebold, Francis X., and Glenn D. Rudebusch. 1996. Measuring business cycles: A modern perspective. Review of Economics and Statistics 78 (February): 67–77. Radelet, Steven, and Jeffrey Sachs. 1998. The onset of the East Asian financial crisis. NBER Working Paper No. 6680, August. Sachs, Jeffrey, Aaron Tornell, and Andres Velasco. 1996. Financial crises in emerging markets: The lessons from 1995. NBER Working Paper No. 5576, May. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 37 Decomposing Inflation ANDREW BAUER, NICHOLAS HALTOM, AND WILLIAM PETERMAN Bauer and Haltom are senior economic analysts in the macropolicy section of the Atlanta Fed’s research department. Peterman is an economic analyst with The Brattle Group. The authors thank Robert Cage and Mary Lynn Schmidt for invaluable help with BLS methodologies and Juan Rubio-Ramírez and Ellis Tallman for helpful comments. ecent declines in U.S. core inflation measures have prompted a renewed effort to understand inflation dynamics. Since late 2001, core consumer inflation rates have declined to levels not seen since the early 1960s. Core inflation as measured by the consumer price index (CPI) declined to 1.1 percent (year-over-year) by the end of 2003 while the core personal consumption expenditures price index (PCEPI) moved below 1 percent. This decline in measured inflation rates, coupled with uncertainty about future demand conditions, generated concern and debate among analysts and policymakers about near-term inflation prospects. That concern was reflected in the May 2003 Federal Open Market Committee (FOMC) statement: “The probability of an unwelcome substantial fall in inflation, though minor, exceeds that of a pickup in inflation from its already low level.” As core consumer inflation rates have edged lower, an increasing and probably undue amount of attention is being placed on the most recent observation. An aggregate inflation rate is limited in the information it provides, especially with regard to the sources of its movements. It is generally difficult to know whether changes in aggregate inflation result from broad-based price changes or from price changes in only a few components. There may be instances in which significant but otherwise idio- R syncratic relative price changes among a few underlying components drive movements in the aggregate inflation rate for a sustained period of time. Analysts often attempt to confront this issue by looking at price changes of major components and making inferences about the impact of those changes on the aggregate inflation rate. However, these inferences are imprecise and do not provide a complete accounting of aggregate inflation. A more rigorous approach is to provide a precise decomposition of the inflation rate. In this article we take the latter, more rigorous approach.1 We calculate and plot the percentage point contributions of major consumer expenditure categories to core inflation measures over time. This technique provides a wealth of information concerning aggregate inflation behavior in a concise way, enabling us to describe the composition of inflation at any point in time. By highlighting the composition of aggregate inflation, we gain greater insight into the underlying trends in inflation and are able to make more informed inferences about the direction of inflation in the near term. A particularly important benefit of this method is that it allows us to distinguish broad-based changes in inflation from changes due to relative price movements of a few components. Using this approach to examine long-run trends in core inflation, our analysis finds that the primary contributor to core inflation over the last two Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 39 decades has been core services. As core services inflation has moved lower over this period, its composition has been relatively stable, with contributions of major components moderating uniformly. In contrast, core goods inflation experienced a distinct downward shift in the early 1990s, marked by a dramatic change in its composition. From examining the long-term trends in the composition of core services and core goods inflation, we believe low inflation will likely persist in the near term. Short-term movements in core services and core goods inflation largely reflect relative price changes of a few components. These relative price changes are generally not persistent enough to drive sus- From examining the long-term trends in the composition of core services and core goods inflation, we believe low inflation will likely persist in the near term. tained movements in the aggregate inflation rate. However, in 2002 and 2003, we conclude, movements in core inflation mostly resulted from significant relative price changes of two components that were persistent enough to alter the path of core inflation for a sustained period. Methodology his analysis examines the two most widely followed measures of consumer inflation, the consumer price index (CPI) and the personal consumption expenditures price index (PCEPI). More specifically, the analysis focuses on the core components of these two measures. The core measures are preferable because they strip out the more volatile food and energy components. While large, persistent movements in food and energy prices may represent important relative price changes, these movements could potentially mask other important price changes that we are more interested in identifying. Our approach follows the methodologies of the Bureau of Labor Statistics (BLS) and the Bureau of Economic Analysis (BEA) to calculate contributions for both CPI and PCEPI inflation, respectively.2 A contribution is the amount in percentage points of the aggregate inflation rate that is attributed to a particular component.3 We use the following general formula for the CPI:4 T 40 ( β Xi, t *βWi ) −( β Xi, t −1*βWi ) Ci, t = 100 * ( β XI, t *βWI ) −( β XI, t −1*βWI ) ( X − XI, t −1 ) * I, t , ( XI, t −1 ) where β Xi,t is the price index of the component i in period t based to the reference period β; β XI,t is the price index of the aggregate I in period t based to the reference period β; βWi is the relative importance of the component i at the reference (or base) period β; and βWI is the relative importance of the aggregate I at the reference period β.5 For the PCEPI we use the general formula [ q + (q / Q F )]*( p − p ) i, t t i, t i, t −1 Ci, t = 100 * i, t −1 , ∑ j [ qj, t −1 + ( qj, t / QtF )]* pj, t −1 ( ) where qi,t is the chained-dollar quantity of the component i in period t, pi,t is the chain-type price index of the component i in period t, Q F is the Fisher quantity index for the aggregate in period t relative to period t – 1, and the subscript j includes all the components of the aggregate. As the formulas suggest, the magnitude of the contribution of a particular component reflects its change in price and its relative share, or weight, in the aggregate. The sum of contributions of all components equals the aggregate inflation rate at any point in time. For the purposes of this article, contributions of individual goods and services are aggregated into major consumer expenditure categories, such as transportation goods, recreation services, and information processing equipment.6 We first decompose aggregate core inflation into its contributions of core services and core goods.7 Then, to obtain greater detail on the underlying trends and recent movements in core inflation, we analyze separately the contributions of major components to core services inflation and core goods inflation. We use the PCEPI to examine the composition and underlying trends of core inflation over the long term. For our short-term focus on the recent past, we use the CPI because it garners the most attention from analysts and markets.8 It is important to note that the BEA relies heavily on CPI series in the construction of its price indexes. Consequently, we believe our analysis of the longterm trends in the PCEPI applies to the CPI as well. Contributions to Core Inflation igures 1 and 2 show core inflation broken down into its contributions of core services and core F Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 FIGURE 1 Contributions to PCEPI Core Inflation 7 6 Core PCEPI (year/year) Percentage points 5 4 3 Core services 2 1 0 Core goods –1 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 Source: PCEPI core inflation rate, BEA; contributions calculated by the authors goods for the PCEPI and CPI, respectively. Both figures clearly indicate that core services is the principal contributor to overall core inflation. This prominence reflects both its larger weight and rate of price increase relative to core goods.9 The figures also indicate that the contribution of core goods eventually turned negative. But note in Figure 1 that the contri- bution of core goods in the PCEPI turned negative as early as 1995. Core goods in the CPI did not turn negative until 2001. Figure 1 brings to light the point that weakness in core goods prices is not as recent a phenomenon as commonly thought.10 Focusing on recent inflation movements in Figure 2 shows that CPI core inflation peaked in November 2001 at 2.8 percent and 1. For a more extensive treatment of the issues in this article, see Bauer, Haltom, and Peterman (forthcoming). 2. For the PCEPI, the methodology is derived from formulas in BEA (2001). For the CPI, the methodology is derived from information in BLS (1997) and from conversations with BLS staff members. 3. The BLS refers to contributions to percent change for the CPI as “effects” although these effects are not published. The BEA does not publish contributions to percent change for the PCEPI. However, the BEA does publish contributions to percent change for the gross domestic product (GDP) and gross domestic purchases price indexes. 4. We modify the general formulas for both the CPI and PCEPI to account for contributions to year-over-year price changes. All year-over-year price changes and contributions for the CPI are calculated using data that are not seasonally adjusted, consistent with BLS reporting procedures. 5. From 1998 to 2001, the BLS uses 1993–95 base period relative importances. From 2002 to 2003, the BLS uses 1999–2000 base period relative importances. 6. See Appendix 1 for a detailed description of how we constructed these categories. 7. The CPI splits items into commodities and services while the PCEPI splits items into categories of goods and services. For the sake of consistency, commodities in the CPI will be referred to as goods. 8. The PCEPI is a methodologically consistent index—that is, it revises historical data when there is a change in methodology. The CPI, however, does not revise history when new methodologies are introduced. This distinction was a primary factor considered in choosing to focus on the PCEPI for long-term trend analysis. In addition, the comprehensive change in the structure of the CPI in 1998 complicates calculating contributions before 1998. 9. The nominal expenditure share of core services to core PCE has increased from 65 percent in 1983 to 70 percent currently. The average rate of price increase for PCEPI core services from 1983 to 2003 is 3.8 percent, compared to 1.1 percent for PCEPI core goods. 10. The decline in core goods prices in the CPI in late 2001 garnered much attention. The decline was easily identifiable because core goods in the CPI is a published index. The decline in core goods prices that was exhibited much earlier in the PCEPI may have been less perceptible because core goods is not a published index in the PCEPI. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 41 FIGURE 2 Contributions to CPI Core Inflation 3.5 3.0 Core CPI (year/year) Core services 2.5 Percentage points 2.0 1.5 1.0 0.5 0.0 –0.5 Core goods –1.0 1999 2000 2001 2002 2003 2004 Source: CPI core inflation rate, BLS; contributions calculated by the authors FIGURE 3 Contributions to PCEPI Core Services Inflation 8 7 Rent Recreation Education Household ops Transportation Personal Other housing Comm & info Medical 6 Percentage points 5 Core services (year/year) 4 3 2 1 0 –1 1983 1985 1987 1989 1991 1993 Source: Calculated by the authors 42 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 1995 1997 1999 2001 2003 FIGURE 4 Contributions to CPI Core Services Inflation 5 Rent Recreation Other shelter Transportation Personal Medical Education Household ops Comm & info 4 Percentage points Core services (year/year) 3 2 1 0 –1 1999 2000 2001 2002 2003 Source: CPI core services inflation rate, BLS; contributions calculated by the authors fell to 1.1 percent in December 2003.11 The contribution of core services fell 0.9 percentage point while the contribution of core goods dropped 0.7 percentage point. The decline in the CPI core inflation rate garnered a great deal of attention among the media, analysts, and policymakers. Although there was some dispute about its significance, many interpreted the decline as an indication that the economy may be headed toward overall price deflation. Contributions to Core Services aving identified core services as the primary contributor to core inflation, the analysis now turns to the historical composition of core services inflation. Figure 3 plots the PCEPI core services inflation rate and the contributions of its major components from 1983 to 2003. Rent, medical care, and personal services are the primary contributors to core services inflation. The long-run movements in rent and medical care largely determine the long-run trend in core H services inflation.12 The contribution of personal services exhibits sharp fluctuations over time, resulting in short-run peaks and troughs in core services inflation. Outside of rent, medical care, and personal services, components’ contributions are relatively small and stable over time. Overall, core services inflation has slowed over the last two decades, with contributions of major components moderating uniformly. We now turn our attention to the behavior of core services inflation over the recent past. Recall from Figure 2 that the disinflation in the overall core CPI in 2002 and 2003 resulted in part from a sustained moderation in the contribution of core services. To better describe this movement, we plot the CPI core services inflation rate and its contributions from 1999 to 2003 in Figure 4.13 Most notably, the figure reveals that the movement in CPI core services inflation was almost entirely driven by rent during this period. The contributions of other components were relatively stable. Core services inflation fell from a peak of 11. The decline in the core CPI inflation rate over this period is 1.6 percentage points, rounded to one decimal place. 12. Over the 1983–2003 period, the correlation between the PCEPI core services inflation rate and the contribution of rent is 0.92, and the contribution of medical care services, 0.81. 13. Figure 4 displays some notable differences from Figure 3. In contrast to the PCEPI, rent is by far the largest contributor to core services inflation in the CPI, with relatively small contributions coming from medical care and personal services. In the CPI, rent has a much larger weight than in the PCEPI, while medical care and personal services have smaller weights. For a thorough examination of the differences in the CPI and PCEPI as well as a detailed discussion of weighting issues, see Clark (1999). Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 43 FIGURE 5 Contributions to PCEPI Core Goods Inflation 4 3 Core goods (year/year) Percentage points 2 1 0 –1 Alcohol Recreation Household furnishings Transportation Other Medical care –2 Education Apparel Info process –3 1983 1985 1987 1989 1991 1993 1995 1997 1999 Source: Calculated by the authors FIGURE 6 Contributions to CPI Core Goods Inflation 2.5 Alcohol Recreation 2.0 Apparel Info process Education Transportation Household furnishings Other Medical Care 1.5 Percentage points 1.0 0.5 0.0 –0.5 –1.0 Core goods (year/year) –1.5 –2.0 –2.5 –3.0 1999 2000 2001 2002 Source: CPI core goods inflation rate, BLS; contributions calculated by the authors 44 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 2003 2001 2003 4.0 percent in February 2002 to 2.6 percent in December 2003, a decline of 1.4 percentage points. The contribution of rent to core services inflation was 2.3 percentage points in February 2002 and fell to 1.1 percentage points in December 2003, a decline of 1.2 percentage points. Consequently, the decline in core services inflation in the 2002–03 period did not reflect broad-based disinflation. Instead, it reflected a significant and persistent relative price change in rent. Contributions to Core Goods he shift in core goods over the past twenty years from being a positive contributor to core inflation to being a negative contributor reflects a number of changes in major components that have dramatically altered its long-run trend. We plot the PCEPI core goods inflation rate and the contributions of its major components from 1983 to 2003 in Figure 5.14 From 1983 through 1991, most components contributed positively to core goods inflation. A number of relative price changes drove short-term fluctuations in the core goods inflation rate. However, these relative price changes were often offsetting and were not persistent enough to significantly drive the core goods inflation rate from its relatively flat trend. In 1992 there was a distinct drop in the core goods inflation rate, characterized by broad-based downward movement in its composition (see Appendix 2). Many components that contributed positively before this shift began to either contribute less positively or began to contribute negatively. In addition, the negative contribution of information processing equipment increased dramatically. From 1997 on, it is difficult to identify the trend in core goods inflation mainly because the core goods inflation rate dropped sharply from 2001 through 2003. However, the composition of core goods has been relatively stable since 1997, suggesting that the core goods inflation rate has settled to a lower, perhaps slightly negative long-run mean. We noted in our discussion of Figure 2 that the steepened decline in core goods prices in 2002 and 2003 was a significant factor in overall core disinflation. This movement is described in Figure 6, which plots the CPI core goods inflation rate and its contributions from 1999 to 2003.15 From November T 2001 to December 2003, the CPI core goods inflation rate fell from 0 percent to –2.5 percent. This drop resulted from a less positive contribution of other goods (largely tobacco) and increasingly negative contributions of household furnishings and transportation. The largest contributor to the decline was transportation. From November 2001 to December 2003 the contribution of transportation fell 1.4 percentage points, with used vehicles accounting for 0.9 percentage point. The collective contribution of other goods and household furnishings also fell 0.9 percentage point during this period. Thus, the drop in core goods inflation resulted from price declines in several components. Most notable The decline in core services inflation in the 2002–03 period did not reflect broad-based disinflation. Instead, it reflected a significant and persistent relative price change in rent. among these was the large price decline in used vehicles (see the box). Conclusion n this article, we determine the precise impact of major components on aggregate inflation measures. We calculate and plot the percentage point contributions of major consumer expenditure categories to core inflation measures over time. This technique provides an information-rich picture of inflation behavior, highlighting its composition and underlying trends. By analyzing the composition of aggregate inflation, we are able to make more informed inferences about the direction of inflation in the near term. We are also able to distinguish broad-based changes in inflation from changes in inflation due to relative price movements of a few components. We find that core services has been the primary contributor to core inflation over the last two decades. The composition of core services inflation is I 14. Alcoholic beverages are not included in the BEA’s core PCEPI. In Figure 1, we presented the BEA core PCEPI without alcoholic beverages. However, in Figure 5, we include alcoholic beverages in our PCEPI core goods in order to be consistent with CPI core goods, which does include alcoholic beverages. 15. Figure 6 shows the same pattern of contributions in CPI core goods as exhibited in the PCEPI during this period in Figure 5. However, there are some differences in the magnitude of the contributions, reflecting the different weighting in the two indexes. Most notably, the negative contribution of information processing equipment is considerably less for CPI core goods because the CPI uses fixed expenditure weights from historical base periods. CPI core goods inflation did not turn negative until late 2001 largely because of the smaller negative contributions of information processing equipment prior to 2002. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 45 BOX Significant and Persistent Relative Price Changes: The Case of Rent and Used Vehicles nalysts have widely discussed rent and used vehicle prices as important factors in CPI core disinflation during the 2002–03 period. But the precise degree to which these components were lowering core inflation was not clear. Our analysis shows that, from November 2001 (the date of the peak in CPI core inflation) to December 2003, the contribution of rent to CPI core inflation fell 0.8 percentage point while the contribution from used vehicles dropped 0.3 percentage point—a total of 1.1 percentage points, a considerable portion of the 1.6 percentage point decline in CPI core inflation. What would CPI core inflation have looked like without these recent movements in rent and used vehicles? As a counterfactual exercise, we construct a hypothetical CPI core inflation measure assuming alternative rates of price change for rent and used vehicles.1 We compare this alternative measure to the actual CPI core inflation rate in Figure A. In the counterfactual index, we hold the rates of inflation of rent and used vehicles constant from November 2001 through December 2003.2 In contrast to the steep decline in the actual CPI core inflation rate, our constructed measure of core inflation shows relatively moderate disinflation over the past two years. This exercise indicates that the decline in actual CPI core inflation reflected significant, persistent relative price changes of rent and used vehicles, not broadbased disinflation. We further argue that these price changes reflect not a fundamental weakening in housing and vehicle demand but, instead, the dynamic effects of interest rates on consumer demand for substitutes. A relatively stable over time and largely driven by movements in a few major components. The story is quite different for core goods inflation. The composition of core goods inflation has changed dramatically over time, resulting in a distinct downward shift in the core goods inflation rate in the early 1990s. Trends in the composition of core inflation lead us to believe that low inflation will likely continue to persist in the near term. The relative stability in the composition of core services inflation suggests little change, in either direction, in the aggregate core services inflation rate. The composition of core goods inflation suggests that core goods defla- 46 FIGURE A CPI Core Inflation Percent change (year/year) 3.5 Holding rent & used vehicles inflation constant after Nov. 2001 3.0 2.5 2.0 1.5 Actual 1.0 0.5 0.0 1999 2000 2001 2002 2003 Source: Actual CPI core inflation rate, BLS; counterfactual, authors Rent Downward pressure on rental prices mainly resulted from an increase in demand for homeownership, which was spurred by historically low mortgage interest rates (see Figure B). As housing starts and home sales surged in the recent recession and recovery, the national rental vacancy rate jumped from 7.8 percent in the fourth quarter of 2000 to 10.2 percent in the fourth quarter of 2003. This effect was compounded by the way owner-occupied housing prices are measured in the CPI. The CPI uses a rental-equivalence approach, measuring the value of the shelter services an owner receives from his or her home. Price movements in owners’ equivalent rent reflect changes in prices of tion will likely continue into the near term. There have been significant changes in market structure, trade patterns, productivity growth, and price measurement that suggest continued downward pressure on goods prices going forward. At the same time, it is not obvious to us that the decline in goods prices will accelerate. The general stability in the composition of core goods inflation since 1997 suggests that the core goods inflation rate will rather revert to a moderately negative rate of decline. With stable core services inflation and stable core goods deflation, we expect that overall core inflation will remain low. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 FIGURE B FIGURE C Owners' Equivalent Rent and Mortgage Interest Rates Used Vehicle Prices and New Auto Finance Rates 9.5 9.0 CPI owners' equivalent rent (left axis) 4.0 8.5 8.0 3.0 7.5 2.5 7.0 6.5 30-year fixed mortgage rate, lagged 20 months (right axis) 1.5 1996 2000 4 10 0 8 –4 6 –8 2002 Source: BLS; Federal Home Loan Mortgage Corporation 4 CPI used vehicles (left axis) –12 1996 6.0 1998 Auto finance company interest rates on new car loans, not seasonally adjusted (right axis) Percent 3.5 2.0 Percent change (year/year) 4.5 12 8 Percent Percent change (year/year) 5.0 2 1998 2000 2002 Source: BLS; Federal Reserve Board rental units that are comparable in characteristics to owner-occupied homes. Therefore, increased demand for homeownership put downward pressure not only on tenants’ rent but also on owners’ equivalent rent—the largest component in the CPI. Used Vehicles The decline in prices of used vehicles largely reflected an increase in demand for new vehicles in response to record-low financing and rebate offers (see Figure C). Used vehicle prices in the CPI are derived from wholesale auction prices. The surge in demand for new vehicles increased the supply of used autos in the wholesale market while also decreasing dealers’ demand for used autos. According to Manheim Auctions, a leader in the used vehicle wholesale auction market, its used vehicle value index (manheimvalueindex.com) fell 5.3 percent between November 2001 and December 2003. Manheim has cited new vehicle incentives as a primary contributor to this decline. 1. We do not exclude rent and used vehicles from our alternative index because doing so would significantly alter the consumer basket, in effect redistributing the weights of these two components to the remaining components. 2. The November 2001 inflation rate was 4.7 percent for tenants’ rent, 4.4 percent for owners’ equivalent rent, and –1.2 percent for used vehicles. We note that short-term movements in core services and core goods inflation largely reflect relative price change of a few components. These relative price changes are generally not persistent enough to cause the aggregate inflation rate to deviate considerably from its perceived trend. However, for 2002 and 2003, we conclude that movements in core inflation mostly resulted from two significant relative price changes—the moderation in the increase of rental prices and the decline in used vehicle prices—that were persistent enough to alter the path of core inflation for a sustained period. From November 2001 to December 2003, the contribution of rent to core CPI inflation fell 0.8 percentage point while the contribution of used vehicles dropped 0.3 percentage point—totaling 1.1 percentage points. The core CPI inflation rate over this period declined 1.6 percentage points. Absent the movements in these two components, core disinflation over the past two years has been very moderate. These results suggest that the concern and discussion regarding overall price deflation were perhaps overstated. Moreover, our results highlight the importance of gauging the impact of relative price changes in a low-inflation environment. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 47 APPENDIX 1 Constructing Major Components for PCEPI and CPI Core Goods and Core Services he BEA publishes only the aggregate core PCEPI, not indexes for core goods and core services. We create these indexes to obtain comparable measures to the core commodities and core services series of the CPI. Within core goods and core services, we aggregate the PCEPI series to create major components comparable to the breakdown of major components in the CPI. The BEA does not include alcoholic beverages in core PCEPI, but we include them in Figure 5 to be consistent with the BLS definition of core CPI. T PCEPI Core Goods • Alcoholic beverages • Household furnishings: semidurable furnishings; cleaning, light supplies, and miscellaneous paper products; flowers, seeds, and potted plants; furniture, mattresses and bedsprings, and kitchen and other household appliances; china, glassware, tableware, and utensils; and other durable house furnishings • Apparel: clothing and shoes and jewelry and watches • Transportation: motor vehicles and parts • Medical care: drug preparations and sundries and ophthalmic and orthopedic equipment • Recreation: toys and sports equipment; magazines and newspapers; audio, video, and musical instruments; sports, photographic equipment, and cycles; and boats and aircraft • Education: books and maps • Information processing: computers, peripherals, and software • Other: tobacco, toilet articles and preparations, stationery and writing supplies, and expenditures abroad by U.S. residents CPI Core Goods • Alcoholic beverages • Household furnishings: window and floor coverings and other linens; furniture and bedding; appliances; other household equipment and furnishings; tools, hardware, outdoor equipment, and supplies; and housekeeping supplies • Apparel • Transportation: new vehicles, used cars and trucks, and motor vehicles parts and equipment • Medical care: medical care commodities • Recreation: televisions; other video equipment, videocassettes, and discs; audio equipment, audio discs, tapes, and other media; pets and pet products; sporting goods; photographic 48 equipment and supplies; other recreational goods; and recreational reading materials • Education: educational books and supplies • Information processing: personal computers and peripheral equipment, computer software and accessories, and other information processing equipment • Other: tobacco and smoking products, personal care products, and miscellaneous personal goods PCEPI Core Services • Rent: owner-occupied nonfarm dwellings, space rent; tenant-occupied nonfarm dwellings; and rental value of farm dwellings • Other housing: household insurance premiums, household insurance benefits paid, and other housing services • Household operations: water and sanitary services, domestic services, moving and storage, rug and furniture cleaning, electrical repair, upholstery and furniture repair, and other household operations • Transportation: transportation services • Medical care: medical care services • Recreation: recreation services • Education: private education and research services • Communication and information: telephone and telegraph and postage • Personal: personal care services, personal business services, religious and welfare activities, and net foreign travel CPI Core Services • Rent: rent of primary residence and owners’ equivalent rent of primary residence • Other shelter: lodging away from home and tenants’ and household insurance • Household operations: water and sewer trash collection services and household operations • Transportation: transportation services • Medical care: medical care services • Recreation: cable television; rental of videotapes and discs; pet services, including veterinary; photographers and film processing; and recreation services • Education: tuition and other school fees and childcare • Communication and information: telephone services and computer information processing services • Personal: personal care services and miscellaneous personal services Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 APPENDIX 2 Examining the Shift in the Composition of Core Goods Inflation he composition of PCEPI core goods inflation has displayed a dramatic shift over the past twenty years. Prior to 1992, the contributions to core goods inflation were mostly positive. Since that time, many components have become consistently negative contributors to core goods inflation. Consequently, core goods inflation has fallen from an average rate of 2.6 percent during the 1983–91 period to an average rate of –0.1 percent post 1991. To understand this shift, we look at the magnitudes, signs, and volatility of the contributions by component both within and across the pre- and postshift periods. We then examine those components that have changed most dramatically. The table presents the average contribution, the standard deviation, and the high/low contribution for each major component within each period. The table also shows the difference in the average contribution for each component across periods. We see relatively large average contributions across periods from apparel (0.44 and –0.30) and transportation (0.54 and 0.19) and increasingly large negative contributions in the postshift period from information processing (–0.55) and recreation (–0.21).1 The most notable changes in sign across periods are within apparel and recreation. The most volatile components across periods, as mea- T sured by standard deviation of contributions, are transportation (0.27 and 0.46) and other goods (0.20 and 0.32). Apparel too is quite volatile in the preshift period (0.38) although the volatility in contributions decreases notably in the 1990s (0.24). There have been significant changes in market structure, trade patterns, productivity growth, and price measurement that have placed downward pressure on goods prices in many components. The components most affected have been apparel, information processing equipment, recreation goods, and transportation goods. The following sections explore the impact of these changes. Apparel The change in apparel from positive contributor to negative is not especially surprising. Significant changes have occurred in the apparel industry at both the manufacturing and retail levels. Most apparel manufacturing has shifted abroad to lowcost producers, increasing the volume of apparel imports. Since 1994 U.S. industrial production of apparel has fallen by nearly 40 percent. At the retail level, discount retailers have become more prominent in the industry. Both of these developments have put downward pressure on apparel prices. During the 1983–91 period, apparel prices grew TABLE Info rma tion goo ds Oth er goo ds goo ds Edu cat ion Rec rea tion goo ds Tra nsp or t atio ng ood s Me dic al g ood s App are l furn ish ing s Ho use hol d Alc oho l PCE P goo I core ds infl atio n A Breakdown of Contributions to PCEPI Core Goods Inflation 1983–91 Average growth/ contribution Standard deviation High Low 2.65 0.63 3.81 1.23 0.33 0.17 0.83 0.16 0.34 0.13 0.84 0.12 0.44 0.38 1.15 –0.42 0.54 0.27 1.21 –0.03 0.39 0.07 0.54 0.29 0.07 0.12 0.29 –0.17 0.06 0.03 0.14 –0.01 –0.10 0.06 –0.01 –0.27 0.55 0.20 1.22 0.28 0.28 0.13 0.55 0.05 –0.21 0.21 0.22 –0.55 0.03 0.03 0.08 –0.09 –0.55 0.23 –0.14 –0.95 0.34 0.32 1.39 –0.30 –0.04 –0.45 –0.21 1992–2003 Average growth/ contribution Standard deviation High Low –0.08 0.99 2.47 –2.23 0.13 0.06 0.42 0.02 –0.03 0.21 0.37 –0.61 –0.30 0.24 0.43 –0.75 0.19 0.46 1.11 –0.74 1992–2003 period less 1983–91 period Difference in average growth/contribution –2.73 –0.19 –0.37 –0.74 –0.35 –0.11 –0.28 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 49 A P P E N D I X 2 (continued) 1.8 percent on average, but during the 1992–2003 period, they fell 1.4 percent, subtracting 0.30 percentage point on average from core goods inflation. Information Processing Equipment The magnitude of the negative contributions of information processing equipment is remarkable. The rapid pace of computer innovation and the role of hedonic quality adjusting in contributing to price declines have been well documented.2 In addition, productivity gains have been especially strong among high-technology manufacturers, reducing production costs. On average, prices of computers, peripheral equipment, and software declined 14.5 percent in the 1983–91 period, and this decline accelerated to 23.4 percent in the 1992–2003 period. At the same time, the nominal expenditure share of computers, peripheral equipment, and software to core goods personal consumption expenditures rose from 0.4 percent in January 1983 to 2.5 percent in December 2003. Together, the steepened price decline and greater nominal expenditure share resulted in a dramatic increase in the magnitude of the average contribution over the two periods from –0.10 to –0.55. Recreation Goods The increasingly negative contribution from recreation goods reflects a variety of factors, including quality adjustment of price indexes, the introduction of new products, and import competition. Over the last several years, the BLS has introduced hedonic quality adjustment procedures for many consumer electronic goods, including televisions, VCRs, DVD players, and audio equipment.3 In addition, the BLS changed its sampling procedures in 1998 to facilitate the introduction of new goods on a more frequent basis (Cage 1996). Incorporating items early in their product cycle captures the dramatic reduction in price that is often associated with relatively new products. This change is particularly relevant for consumer electronic goods. Import competition has also put downward pressure on recreation goods prices. Import prices for most recreation goods began to fall considerably in the mid 1990s—averaging –2.7 percent for home entertainment equipment and –0.8 percent for toys and sporting goods. The increase in discount retailers has also placed greater downward pressure on recreation goods prices. Transportation Goods The contribution of transportation goods prices is noteworthy in that it is large in magnitude, exhibits considerable volatility, and turns negative in the mid-1990s. Expenditures on motor vehicles and parts are a relatively large share of total core goods consumption, averaging 23 percent of core goods personal consumption expenditures from 1983 to 2003. Prices fluctuate considerably, especially for used vehicles, resulting in large swings in contributions to core goods inflation. Since 1983, the average year-over-year price change for used autos was 3.4 percent, with a standard deviation of 7.0 percent. In addition, motor vehicle prices shifted downward in the mid-1990s. Before 1995, price changes for motor vehicles and parts averaged 2.8 percent. Since 1995, they have averaged just 0.3 percent. This shift in prices reflects the changing structure of the motor vehicle industry. Since 1996, the share of domestic light vehicle sales to total light vehicle sales has fallen by nearly 10 percentage points. Domestic vehicle manufacturers’ attempt to retain market share has placed downward pressure on new vehicle prices. Meanwhile, leasing has increased dramatically over the past decade, resulting in an influx of latemodel, low-mileage used cars into the vehicle market. In essence, leasing has produced a new category of used cars that is a closer substitute for new vehicles. Used car superstores have emerged, increasing competition at the retail level. 1. “Other” goods also has a large contribution across periods. The magnitude and volatility of this contribution mostly reflect large price swings in tobacco goods. 2. Beginning in 1991, the BEA used the BLS producer price index (PPI) series for electronic computers, which adjusts for quality changes in computers. Once the BLS began using hedonic quality adjustment for computers in the CPI in 1998, the BEA switched to the CPI series. For a good discussion and example of hedonic quality adjustment for computers, see Holdway (2000). 3. Hedonic quality adjustments were incorporated for televisions in January 1999 and for VCRs and DVD players in April 2000. The impact of hedonic pricing does not necessarily translate to a downward adjustment to the published (nonadjusted) index. Liegey (1994) and Liegey and Shepler (1999) show that the introduction of hedonic pricing for apparel and VCRs, respectively, did not greatly affect the price changes. However, Moulton, LaFleur, and Moses (1998) show that the introduction of hedonic pricing of televisions did result in a downward adjustment. 50 Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 REFERENCES Bauer, Andrew, Nicholas Haltom, and William Peterman. Forthcoming. Examining contributions to core consumer inflation measures. Federal Reserve Bank of Atlanta working paper. Liegey, Paul R., and Nicole Shepler. 1999. Adjusting VCR prices for quality change: A study using hedonic methods. Monthly Labor Review 122 (September): 22–37. Cage, Robert. 1996. New methodology for selecting CPI outlet samples. Monthly Labor Review 119 (December): 49–61. Moulton, Brent R., Timothy J. LaFleur, and Karin E. Moses. 1998. Research on improved quality adjustment in the CPI: The case of televisions. Presented at the Conference of the Ottawa Group, April. Clark, Todd E. 1999. A comparison of the CPI and the PCE price index. Federal Reserve Bank of Kansas City Economic Review 84 (Third Quarter): 15–29. U.S. Bureau of Economic Analysis (BEA). 2001. A guide to the NIPA’s methodology, National Income and Product Accounts, 1929–97. Washington, D.C.: BEA. Holdway, Michael. 2000. Quality-adjusting computer prices in the producer price index: An overview. Bureau of Labor Statistics. <www.bls.gov/ppi/ppicomqa.htm> (February 17, 2004). U.S. Bureau of Labor Statistics (BLS). 1997. The consumer price index. Chap. 17 in BLS Handbook of Methods. Washington, D.C.: BLS. Liegey, Paul R. 1994. Apparel price indexes: Effects of hedonic adjustment. Monthly Labor Review 117 (May): 38–45. Federal Reserve Bank of Atlanta E C O N O M I C R E V I E W First Quarter 2004 51