Experimental indexes of leading and coincident economic indicators.

Stock, James H.

Annual Research Conference--I:

Experimental Indexes of Leading and Coincident Economic Indicators

The index of coincident economic indicators is a weighted average of several broad monthly indicators of current economic conditions. The index of leading economic indicators is a weighted average of a set of series that signal future changes in overall economic activity. My NBER research project with Mark W. Watson of Northwestern University takes a fresh look at these two indexes and develops new alternatives to the present Department of Commerce (DOC) indicators.

Although coincident and leading indicators currently are produced by the DOC, this is a fitting research project for the NBER. Indeed, the genesis of these indexes was a report written 52 years ago by Wesley Clair Mitchell and his research Associate Arthur F. Burns. That report developed the system of coincident, leading, and lagging indicators that has led to the indexes currently produced by the DOC.

Watson and I started this project with two broad questions: first, how should we construct indexes that provide a timely and interpretable forecast of the state of the economy over the next six months? This question is motivated by the practical problem of using the DOC index of leading indicators to forecast a recession. For example, the rule of thumb that forecasts a recession if there are three consecutive declines in the DOC's monthly leading index is neither timely nor precise. Thus we focused on producing direct forecasts of short-term overall economic growth and of the probability of a recession.

Second, we asked which series to include in constructing these indexes. At a deeper level, how should we decide which series to include and which to exclude? The traditional approach to selecting series for the leading index basically has been bivariate; that is, comparing series one at a time to movements in the coincident index. In contrast, our approach to variable selection is multivariate. It focuses on picking series that have important predictive content on the margin: that is, that have predictive content given the other series in the index.

Three New Experimental Indexes

Our project has resulted in three new indexes. To distinguish them from the indexes produced by the DOC, and to emphasize that this is an ongoing research project, we refer to these as "experimental" indexes. The first of these indexes is the experimental index of coincident indicators (XCI). Like the coincident index produced by the DOC, the XCI is designed to measure --on a monthly basis--the current level of overall economic activity. The second index, the experimental leading index (XLI), forecasts the growth of the XCI over the next six months, scaled to provide annual rates. This index is computed using a revised set of leading variables. The third index, the experimental recession index (XRI), represents a new concept in the context of coincident and leading economic indicators. This index estimates the probability that the economy will be in a recession in six months. Because it is a probability, the index can range from 0 to 100 percent.

Experimental Index of Coincident Indicators (XCI)

The XCI is plotted in Figure 1. Cyclical peaks and troughs, as determined by the NBER's Business Cycle Dating Committee, are indicated by vertical lines. Our XCI is quantitatively similar to the coincident index produced by the DOC. Like the DOC series, ours is a weighted average of four broad measures of economic activity: industrial production; real personal income (less transfers); real manufacturing and trade sales; and employee-hours at nonagricultural establishments. The index is scaled to equal 100 in 1967.

The two main differences between our XCI and the DOC coincident index are, first, that we use employee-hours rather than the number of employees and, second, that we put some weight on lagged values of these series. These lagged weights arise naturally from the statistical model--a so-called dynamic factor model--that we use to construct this index. In any case, these weights are small. Overall, the correlation between the monthly growth in the XCI and the growth of the DOC coincident index is 95 percent.

In practice, the XCI can be thought of as a monthly measure of GNP, although it is somewhat more volatile than GNP itself because of differences in coverage. For example, if you average three months to construct a quarterly XCI and then compute the correlation between the two-quarter growth in GNP and the two-quarter growth in this quarterly XCI, this correlation is almost 90 percent. Because the XCI focuses on cyclically sensitive series such as manufacturing and trade sales, it is more volatile than GNP. The two series have approximately the same average growth rates since 1960, but a 1 percent deviation from the mean growth in the XCI roughly corresponds to a 0.6 percent deviation from the mean growth in GNP, at annual rates.

Experimental Index of Leading Indicators (XLI)

Figure 2 plots the XLI. This is a forecast of the growth of the XCI over the next six months. For example, the value of the XLI for January is a forecast of the growth of the XCI from January to July, at annual rates.

The XLI is a weighted average of current and lagged values of the four coincident series and seven leading series. The seven leading series were selected from an original list of 280 series. On a conceptual level, there were two main criteria for selecting the series from this longer list. First, each of the series chosen must make a useful forecasting contribution, given that the other series already were included in the index. This is the focus on multivariate, rather than bivariate, predictive content that I mentioned earlier. Second, the role of each series had to be stable over time. It was not enough that a series helped forecast the XCI only during the 1970s; for example, we also required the forecasting relationship to be stable, to the extent that this can be determined by econometric analysis. This series selection procedure started from scratch, not with a single predetermined base list of series.

Our seven leading series are: 1) new private housing authorizations; 2) manufacturers' unfilled orders in durable goods industries; 3) a trade-weighted index of exchange rates between the United States and five other nations (Japan, the United Kingdom, West Germany, France, and Italy); 4) part-time work in nonagricultural industries because of slack work; 5) the change in the ten-year Treasury bond rate; 6) a measure of the risk premium on high-grade, short-term private paper (specifically, the spread between the six-month commercial paper rate and the six-month Treasury bill rate); and 7) a measure of the slope of the yield curve (specifically, the spread between the yields on ten-year and one-year Treasury bonds). See Table 1 for an example of the use of these series.

Two of these series--manufacturers' unfilled orders and housing authorizations--are in the current DOC leading index. Our series on part-time work is related closely to the DOC series on new claims for unemployment insurance. However, the remaining four series represent major departures from the traditional list of series.

Experimental Recession Index (XRI)

The XRI is plotted in Figure 3. It is a direct estimate of the probability that the economy will be in a recession in six months and, as such, is a new concept. This probability is computed within the context of the econometric model used to construct the coincident and leading indexes. Thus the XRI has the same components as the XLI, but the various series are combined so as to predict recessions directly.

How Useful Are the Experimental Indexes?

The ultimate usefulness of these new indexes can be determined only by their future ability to forecast recessions and expansions. We have been producing these indexes for less than a year--too short a time to be able to evaluate their out-of-sample performance.

However, we can examine the simulated predictive performance of the series within the historical sample. By this standard, their performance is very good. For example, the XRI ideally would have a value of one exactly six months before a cyclical trough and would shift to zero exactly six months before a cyclical peak.

As can be seen in Figure 3, the performance during the 1970, 1974, and 1979 recessions was good, although the performance during the 1982 recession was less satisfactory. Also, the XRI would have signaled a recession in 1967, although in fact there was no recession then. This is a much better track record than one based on the "three consecutive declines" rule of thumb applied to the DOC leading index. Thus we are optimistic about the potential of this index.

Comparison of Series in DOC and Experimental Indexes

Since its most recent revision in January 1989, the DOC leading index has been based on 11 leading indicators. However, only two of these series appear on our list. Therefore, it makes sense to ask whether the DOC series really do belong on our list and, if not, why not. When we looked into this, we concluded that, given the other series in the index, none of these nine series made any important additional forecasting contributions. In contrast, given various sets of series from the DOC list, when we added series from our list, these new series in fact did result in important improvements.

Perhaps the two most noteworthy examples of series that have been identified traditionally as important leading indicators but that are not on our list are the money supply (M2) and stock prices. Some people have found one conclusion of our analysis surprising: that including the money supply or stock prices in our indexes did not help. In fact, these series had considerable potential to hurt the performance of these indexes. Concerning stock prices, the clearest example of this was October 1987, although three months earlier we already had reached our conclusion that stock prices did not belong in the index.

Unconventional Series That Bear Watching

Our research has resulted in identifying indicators that deserve close attention. One of these is a measure of the slope of the far end of the yield curve as measured by the spread between ten-year and one-year Treasury bonds. This work provided statistical support for the observation that an inverted yield curve signals a future slowdown. A natural interpretation of this finding is that high interest rates today, relative to the future, could reflect tight monetary policy today and reduced future inflation associated with an overall economic slowdown.

A second unconventional series in our index is a measure of the risk premium on high-grade, short-term paper (that is, the spread between six-month high-grade commercial paper and six-month Treasury bills). This risk premium also has a natural interpretation. It provides a measure of the likelihood that on average these firms will have the future cash flow and credit stature to be able to meet these relatively short-term obligations.

Acknowledgments

This project has benefited greatly from the advice and counsel of Geoffrey H. Moore, Victor Zarnowitz, and other members of the NBER Business Cycle Dating Committee. We hope that these experimental indexes prove useful in developing improved short-term forecasts of economic conditions. [Figure 1 to 3 Omitted] [Tabular Data Omitted]