The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
SEPTEMBER/OCTOBER 1992 ECONOMIC PERSPECTIVES A review from the Federal Reserve Bank of Chicago Making sense of economic indicators: a consumer's guide to indicators of real economic activity Conference announcement: Shaping the Great Lakes Economy i I l f ' .C ■ • '■ Contents Making sense of economic indicators: a consumer's guide to indicators of real economic activity..........................................................................2 Francesca Eugeni, Charles Evans, and Steven Strongin Policymakers and business analysts face a daunting array of economic indicators; which indicators are useful in a given situation depends on the question and the time frame. The authors evaluate a number of different indicators and discuss their use in a variety of contexts. Conference announcement: Shaping the Great Lakes Economy....................................................... 32 ECONOMIC PERSPEC I IVES Karl A. Scheld, S e n io r V ic e P r e s id e n t a n d D ir e c to r o f R e s e a rc h Editorial direction Carolyn McMullen, e d ito r , David R. Allardice, r e g io n a l Herbert Baer, f in a n c ia l s tr u c tu r e a n d re g u la tio n , Steven Strongin, m o n e ta r y p o lic y , Anne Weaver, a d m in is tr a tio n s tu d ie s, Production Nancy Ahlstrom, ty p e s e ttin g c o o r d in a to r , Rita Molloy, Yvonne Peeples, ty p e s e tte r s , Kathleen Solotroff, g r a p h ic s c o o r d in a to r Roger Thryselius, Thomas O’Connell, Lynn Busby-Ward, John Dixon, g r a p h ic s Kathryn Moran, a s s is ta n t e d ito r Septem ber/O ctober 1992 V olum e XVI, Issue 5 ECONOMIC PERSPECTIVES is published by the Research Department of the Federal Reserve Bank of Chicago. The views expressed are the authors’ and do not necessarily reflect the views of the management of the Federal Reserve Bank. Single-copy subscriptions are available free of charge. Please send requests for single- and multiple-copy subscriptions, back issues, and address changes to Public Information Center, Federal Reserve Bank of Chicago, P.O. Box 834, Chicago, Illinois 60690-0834, or telephone (312) 322-5111. Articles may be reprinted provided source is credited and The Public Information Center is provided with a copy of the published material. ISSN 0164-0682 M aking sense of econom ic indicators: a consum er's guide to indicators of real econom ic activity Francesca Eugeni, Charles Evans, and Steven Strongin Economic data arc used prima rily in two ways. Academic | economists typically use data to build models of the economy in I order to understand how the economy works. Business analysts, on the other hand, use economic data to forecast future eco nomic activity. These two activities and groups of people are not truly distinct groups, neverthe less the two activities do involve some substan tive differences. The problem facing the busi ness analyst, and to a large extent the policy maker or businessman who has to make deci sions based on the economic outlook, is how each piece of new information should be as sessed. Does it portend higher growth or lower, a recession or a boom, slow growth or stasis? Such assessments are crucial to running a suc cessful business and to the proper ongoing evalu ation of economic policy. Yet economic analy sis rarely focuses on precisely these questions. In the current article, we develop an organized structure for evaluating economic indicators and apply that structure to a wide variety of financial indicators and a selected group of real indicators as well. This process is fundamentally more eclectic than the usual econometric analysis which looks for or constructs a “best” indicator, where “best” typically refers to winning some narrowly de fined contest of general purpose forecasting ability measured over some preselected time span.1 Unfortunately, experience tells us that such a search is likely to end in failure. Eco nomic history is full of examples of indicators, such as stock prices and various monetary aggre 2 gates, which work for a short period of time after their discovery and then fail dramatically just as they become widely used. There are many reasons for this, but one stands out. As the following analysis will show, indicators do well at different things and at different times. Without an understanding of the limitations this implies, these “best” indicators are often stretched well beyond their capabilities. What the business analyst really needs to know is the type of information that an indicator possesses and the types of purposes to which it can rea sonably be put. Indicators, like people, perform better or worse depending on the context in which they operate. Efficient usage requires matching indicators both with appropriate questions and with other complementary indicators. For instance, some indicators, such as the Purchas ing Managers’ Index of the National Associa tion of Purchasing Management (NAPM), do well at predicting short run changes in activity, but do not do very well at pinning down the level of activity over longer time spans. Other indicators, such as the growth in real M2, fore cast short run phenomena poorly, but do better at predicting average activity over a longer time Francesca Eugeni is an associate economist, Charles Evans is a senior economist, and Steven Strongin is vice president and econom ic advisor at the Federal Reserve Bank of Chicago. The authors w ould like to thank the participants of the Special Meeting on Operating Procedures held at the Federal Reserve Bank of St. Louis, June 18 and 19, 1992, and the participants of the Federal Reserve Bank of Chicago's Macro W orkshop held June 9, 1992, fo r their comments. ECONOMIC PERSPECTIVES span such as a year. Also, while some indicators are very close substitutes, such as the twenty or so interest rates sometimes used in econometric studies, each providing little additional informa tion beyond the first, other indicators possess substantial independent information, thus pro viding important confirming or contradicting information. The analyst needs to know how to match questions with indicators depending on current needs. A swiss army knife is a fine gen eral purpose tool, but it is hardly a substitute for a well equipped workshop. It is not enough just to produce a “best” model; rather, it is important to understand what type of information is con tained in a given indicator so that its message can be properly evaluated and also to determine how much weight to give that message given what else is also known. This article develops and implements a set of procedures for evaluating indicators of eco nomic activity that closely match the actual use of such indicators by policymakers and business men alike. We see that process as primarily involving the reassessment of short to medium term economic activity based on an indicator by indicator analysis, with the primary decision matrix being whether to revise the assessment of activity up or down. We do not address related issues of assessing long run growth, inflation, interest rates, or the value of the dollar. Evaluat ing indicators in this context has four primary parts: ranking candidate indicators; characteriz ing the nature of the information in those indica tors; assessing their usefulness in practice; and determining what relative weight should be given to each indicator. The idea is to develop the information that an analyst needs in order to interpret information as it comes in and to choose which indicators to watch depending on the questions being asked. All of our analyses will be carried out on a bivariate (two variable) basis. Multivariate regression models allow indicators to play off against one another making it impossible to determine exactly what information is in each indicator. This in no way reduces the generality of the methods developed in this study, in that the forecast of a given multivariate model can be treated as a single indicator, just like any other. In fact, the National Bureau of Economic Re search (NBER) Experimental Leading Index examined in Section 4 is just such an indicator. Once the indicators are assessed and charac terized, the last section of the article formally FEDERAL RESERVE BANK OF CHICAGO addresses the question of how to weight the information in one indicator relative to another. This is done through mixing models, which effectively produce a forecast based on the weighted average of the individual forecasts generated by the indicators. There are a num ber of advantages that are derived from using this mixing approach over the classical multiva riate forecasting techniques. First, when one of the indicators begins to fail, which they do, you can reweight or at least temporarily just ignore that indicator. Second, by using only the pri mary information in each indicator, these mod els are less subject to the type of overfitting arising from interactions between indicators that plagues large econometric models. Third and most important, the mixing approach al lows a much more precise assessment of exact ly the type and value of information that is contained in each indicator and thus allows analysts to reoptimize their choice of indicators based on the type of question being asked. Our investigation indicates that this type of analysis is crucial to the effective use of indica tors. First, we find that a number of commonly used indicators, such as the monetary base and M 1, actually contain negative information, in the sense that forecasts based purely on the past history of activity, ignoring these indicators, do better in practice than forecasts which include the information in these indicators. Second, we find that long term interest rate levels provide no additional information about future econom ic activity beyond that contained in short term interest rate levels, while the slope of the term structure contains substantial additional infor mation. This would seem to indicate that a rise in long term interest rates is associated with an improvement in the near term outlook of the economy. It is interesting to note that this is contrary to popular wisdom, according to which a scenario with declining short term interest rates and increasing long term rates is viewed as negative. Third, we find that some indica tors, such as the spread between the 3 month eurodollar rate and the 3 month Treasury bill rate, do a very good job of forecasting growth during expansions, but rarely signal recessions, while others, such as real M 1 and the mix be tween bank and nonbank financing do better at forecasting during recessions, even though they are poor forecasters in general. Fourth, we find that composite indicators, such as the Depart ment of Commerce Composite Index of Lead 3 ing Indicators and the NBER Experimental Leading Index, are very good predictors of economic activity over a two quarter horizon, while real M2 and the slope of the term struc ture are more useful over a one year horizon. This last finding illustrates a crucial point: the forecast horizon is fundamental to the choice of indicators. Short horizons favor inter est rate risk spreads, such as the difference between the 6 month commercial paper rate and the 6 month Treasury bill rate (risk spreads are yield differences between private and public debt instruments with the same maturity), and activity based indicators, such as the Purchas ing Managers’ Index and the Sensitive Materi als Price Index. Longer horizons, on the other hand, favor monetary indicators, such as real M2, and interest rate term spreads, such as the difference between the 12 month Treasury bill rate and the overnight federal funds rate (term spreads are yield differences between two pub lic debt instruments with different maturities). This indicates that different types of informa tion are important for forecasting growth at different forecast horizons. Methodology As noted above, the primary focus of this article is the examination of various data series as indicators of changes in real economic activ ity, which we measure as annualized quarterly log changes in real GDP, except in the sections of the article which focus on issues of timing, in which case the annualized monthly log changes in employment are used. Since the employment data series is available at the monthly frequency, it allows for more precise estimation of the pattern of impact over time. Throughout the article the indicators are used to produce forecasts of economic activity. The specific functional form of the forecasting equation is always the same. One year of data for the indicator and one year of lagged eco nomic activity are included in the regression. Thus, the exercise is strictly equivalent to a bivariate vector autoregression (VAR) with one year of lags: four quarters of lags for the real GDP models and twelve months of lags for the employment models. The models are estimated in log differences and rates of change are annu alized. Interest rates, interest rate spreads, and some of the composite indicators are used in their level form. In many of the tables an addi tional forecast is provided with the label 4 “NONE.” In this case, the forecast is based solely on the past history of economic activity, that is, a pure autoregressive model with one year of lagged data. This pure autoregressive forecast is referred to as the no-indicator fore cast. When the horizon of the forecast is var ied, we simply change the dependent variable in the regression rather than dynamically iterate the one period ahead forecast. This optimizes the parameterization for the forecast horizon in question, rather than multiplicatively combin ing estimation errors forward. Symbolically the forecasting equation can be written: 0 ) Y,+k - Y, = A (W Y ' , + B (L )It_] + cd ; where T is the log of economic activity at time t, / is the indicator at time t, k is the number of periods in the forecast horizon, and A(L) and B(L) are polynomials in the lag operator L of order one year. The indicators are split into four groups, which we call families. Each family is meant to represent a natural division of indicators into groups which are likely to share similar charac teristics. The first family we examine is inter est rate levels, the second is money based mea sures, the third is interest rate spreads, and the fourth is composite indicators, such as the De partment of Commerce Composite Index of Leading Indicators and the Standard and Poor’s 500 Stock Index. The fourth group also con tains those series which do not fit neatly into the overall classification scheme. The idea is to first examine the indicators within a family, characterize the information, and find out which indicators within each fami ly produce the best forecasts and contain the most independent information. Then we take these “best” indicators from all four families and examine what is to be gained by mixing the information from different families. This serves a number of purposes. First, breaking the large list of potential indicators into smaller groups makes each examination more manage able. Second, using natural groupings allows us to look at questions such as what is the best interest rate or the best money measure in a natural way. Third, one key issue for indicators is the degree to which they actually contain independent information. Focusing on groups which are already thought to have similar infor mation provides a natural way to learn if these ECONOMIC PERSPECTIVES preconceptions are accurate or if some of these groups contain more than one type of informa tion. Lastly, by first selecting the best indica tors at the family level and then mixing be tween families, we can produce a mixed fore cast which, as noted above, closely approxi mates the way indicators are used in practice. Each family of indicators is subjected to the same analysis. First, each family of indica tors is described. Then each of the indicators is subjected to four evaluations: classical goodness-of-fit rankings; indicators’ performance in practice; characterization of fit; and encom passing tests. The results of our evaluations are summarized in tables numbered as follows: the first digit in the table’s number refers to the family of indicators (for example, interest rate levels constitute our first family), while the second digit refers to the type of statistics dis cussed (for example, multiperiod forecast re sults are summarized in the second table of each family). For example, Table 1.2 is the second table in our first family of indicators. The first part of our analysis focuses on classical goodness-of-fit statistics, which are based on simple full sample regressions esti mated on data from January 1962 to December 1991. The results are presented in Table _ .l2 of each family analysis section. In this table we report the correlation coefficients produced by the regression, and we rank the indicators in each family according to their R2s. The idea is that the best indicators are the ones that pro duce the best fit as measured by the R2of the regression. This closely approximates the stan dard notions of evaluating indicators of eco nomic activity. It is also closely linked to the notion of Granger causality, which statistically measures whether or not the indicator actually helps forecast economic activity. The probabil ity value for this test is also included in the table. Low probability values, especially below .05, are normally thought to indicate that a variable is valuable in generating forecasts. The second evaluation switches the focus to how well the indicators are likely to work in practice. To this end, goodness-of-fit is reinter preted in a way closer to the way forecasts are actually used. First, Table _.2 shows goodnessof-fit rankings recalculated for a series of fore cast horizons using standard regression analysis to provide a bench mark for evaluating out-ofsample forecasts. The one quarter horizon used in Table _.l is first presented and then a two FEDKRAI. RESERVE RANK OF CHICAGO quarter forecast horizon evaluation and a four quarter forecast horizon evaluation.3 Table _.3 in each section then repeats this analysis using forecasting equations which do not contain any prior information. Specifically, the forecasting equations are estimated using Kalman filtering techniques which recursively compute minimum mean squared errors using only data available prior to the forecasting period. This analysis provides a more accurate assessment of how an indicator is likely to perform in practice, since this is the regression an analyst would have actu ally estimated just prior to making the forecast, rather than the regression the analyst would gen erate today using all of the data since the forecast period. These forecasts are then ranked by the root mean squared error (RMSE) (the average size of the error) of the forecasts from July 1973 onward. To see how the indicators perform under different circumstances, we look at Kal man forecasts in recessions and expansions, and re-rank the indicators according to their RMSEs, as shown in Table _.4. Next, Figure _. 1 in each section graphs the cumulative residuals for the Kalman forecasts. These charts allow us to determine if these fore casts tend to perform badly during recessions or if there was some particular point in the past where they did especially well or poorly. It also tells us if the forecasts have tended to miss in some systematic fashion over time. The residu als are measured as the actual growth in econom ic activity minus the forecasted growth. There fore, although a flat cumulated residuals’ slope indicates good overall performance, a path con sistently close to the zero horizontal line would be ideal. On the other hand, a downward trend in the cumulative residuals would indicate a period of overpredicting growth in activity, while an upward trend would indicate a period of under forecasting. The third evaluation seeks to characterize the type of information in the indicator. Typical ly the question can be thought of as follows: if the indicator goes up today how does that change my expectations about economic activity in the future? This is analyzed by calculating the dy namic response path of employment for each of the indicator forecasting equations, which shows how a one standard deviation4 increase in the indicator changes expectations about the future growth rate of employment for each month for the next 36 months.5 This allows us to character ize the information in the indicator based on how 5 fast economic activity responds, how much it responds and how long the change in activity lasts. Figure _.2 in each family section graphs the dynamic response path for selected indica tors in the family, as well as the two standard deviation bands on the estimates of the dynam ic response paths to show the amount of uncer tainty about the response path. The fourth evaluation switches the focus to independence of information. As noted earlier, one of the most important factors to understand about indicators is whether or not they contain independent information relative to some other indicator. This allows the analyst to assess whether a new piece of information actually contains any additional information or whether it is simply the same information with a differ ent label. This is evaluated through a set of techniques called encompassing tests. In the context of this paper, indicator A is said to encompass indicator B if, given the forecast implicitly based on A, there is no additional information in indicator B. Indicator A is said to dominate indicator B if A encompasses B and B does not encompass A. The simplest way to test this is to run a regression with eco nomic activity as the dependent variable and the forecast of activity based on indicator A and the forecast of activity based on indicator B as the independent variables. Symbolically this can be written: (2) AGDP= <J)for(A)' + ( l - Wor(B)t + e; where for(A)i and for(B)i are the forecasts of GDP based on indicators A and B respectively and ()) is the relative weight an ordinary least squares (OLS) regression assigns to for(A)t and for(B)r If (]) is significantly different from 0 then we can reject that for(A) is encompassed by for(B). Likewise if 1 —<)>is significantly different from 0 then we can reject that for(B) is encompassed by for(A). If neither is encom passed then both indicators contain independent information and a better forecast can be ob tained by mixing both sets of information with the relative weights given by (J). If only one is encompassed, then it is said to be dominated and only the other is necessary to produce an efficient forecast. If both are encompassed then either indicator alone can produce an efficient forecast. This occurs when there is a very high degree of collinearity and the standard error of the parameter estimate is large. In this case the 6 indicator which has the best historical track record would likely be the superior choice. The generalization to longer horizons is straightfor ward, though the calculations of the standard errors are more complicated since the errors are no longer independent. Table _.5 in each family section contains the encompassing tests. The table is read as follows. The indicators are listed both along the top and along the side of the matrix. The numbers in the table refer to the test that the indicator listed along the side is encompassed by the indicator along the top. The statistics reported are the significance levels for the test that the indicator along the top does in fact contain all the infor mation in the indicator along the side. Values below .05 indicate substantial independent infor mation possessed by the indicator listed along the side. For the sake of readability, such values are replaced by a dash in the table. In general, the lower the number, the more likely it is that the indicator listed along the side possesses inde pendent information and the higher the number, the more likely it is that the indicator listed along the top encompasses the indicator along the side. The way to interpret Table _.5 is that a side indicator whose row is blank contains informa tion that is independent of every other indicator in the family. A top indicator whose column is full of high numbers is said to encompass the indicators on the side. An indicator that did both would be said to dominate the family. In gener al, we search for the set of indicators in each family which contains all the information in the family using as few indicators as possible. This will mean that the best variable from the previ ous tests will be included together with addition al indicators which contain independent informa tion, that is, the indicators that add the most. Formally, this means that we include all indica tors that are not encompassed by any other indi cators in the family plus whatever additional indicators are necessary to fully encompass or cover all of the other indicators in the family. This is analogous to finding a set of minimally sufficient statistics. The indicators that make it through this process will then be tested in the mixing model section of the article in between-family encom passing tests, which examine whether or not there is independent information between fami lies. Then a set of “best” indicators will be se lected in order to develop mixing models of indicators which contain independent informa- ECONOMIC PERSPECTIVES tion for each of the forecasting horizons. These models will contain estimates of the ap propriate relative weights that should be applied to the individual indicator-based forecasts. Completing the circle of policy forecasts, the mixing models will be time varying to see if there is any gain from adjusting the weight ap plied to these individual forecasts based on re cent performance. 1. In te re s t ra te levels TABLE 1.1 Classical goodness-of-fit statistics R2 Correlation with real GDP P-value Rank FF 0.338 -0.353 0.0000 3 TB03 0.293 -0.299 0.0001 6 Indicator TB06 0.304 -0.295 0.0000 5 CM01 0.309 -0.282 0.0000 4 CM03 0.279 -0.257 0.0002 7 As shown in Table 1.1, we selected the 0.268 -0.251 0.0003 8 CM05 following levels of interest rates for investiga -0.237 0.0009 10 0.253 CM10 tion: the federal funds rate (FF); the 3 and 6 0.354 -0.352 0.0000 1 EUR03 month Treasury bill rates (TB03 and TB06); the 2 0.348 -0.342 0.0000 CP6 1, 3, 5, and 10 year Treasury constant maturity 0.0007 BAA -0.269 9 0.258 bond rates (CM01, CM03, CM05, and CM 10); the 3 month eurodollar rate (EUR03); the 6 NOTE: Sample period is January 1962 - December 1991, quarterly data. month commercial paper rate (CP6); and the BAA corporate bond rate (BAA). Goodness-offit tests show that all of these interest rates are To determine how interest rates would actu negatively correlated with real GDP, which ally perform as indicators of economic activity, indicates that an increase in interest rates this we use Kalman filtering techniques to produce period is associated with a decline in real output. out-of-sample forecasts using only data available The eurodollar rate, the commercial paper prior to the forecasting period. When we rank the rate, and the federal funds rate have the three resulting RMSEs in Table 1.3 it becomes clear, largest absolute correlation coefficients with once again, that the overall performance of short real GDP and produce the best fit to the model term interest rates improves when we expand the as measured by their individual R2s, ranking forecast horizon. FF continues to perform best at first, second, and third, respectively. The the one year forecast horizon, while maintaining strength of such relationships is not surprising a standing similar to the in-sample results at given the role that these instruments play in money markets. For exam TABLE 1.2 ple, because the federal funds rate Multiperiod forecasts, in-sample is a key instrument of monetary policy and a bench mark for other Real GDP money market interest rates, fluctu 1 quarter 2 quarters 4 quarters ations in the rate are strongly asso Indicator R2 Rank R2 Rank R2 Rank ciated with future movements in FF 0.338 3 0.463 3 0.530 1 real economic activity. TB03 0.293 6 0.402 5 0.496 3 The predictive power of our 0.304 4 0.487 TB06 0.406 5 5 interest rate family is then tested at different forecast horizons using 4 CM01 0.397 6 0.443 6 0.309 standard regression analysis over 7 0.377 7 0.279 7 0.350 CM03 the full sample period. The in0.332 8 0.346 8 CM05 0.268 8 sample results of Table 1.2 show 0.296 0.307 CM10 0.253 10 10 10 that while EUR03 loses some of its 0.354 1 0.471 2 0.490 4 EUR03 strength as the forecast horizon CP6 0.348 2 1 0.516 2 0.475 increases, as shown by the recalcu lated rankings, the fit of both CP6 BAA 0.258 9 0.329 9 0.315 9 and FF improves at longer forecast NONE 11 11 11 0.118 0.123 0.076 horizons, with FF having the stron gest predictive power at the four NOTE: Sample period is January 1962 - December 1991, quarterly data. quarter forecast horizon. FEDERAL RESERVE BANK OF CHICAGO 7 er maturity bonds, such as the 3, 5, and 10 year Treasury bonds. Kalman multiperiod forecasts, out-of-sample Once the general strength of Real GDP an indicator is established, it be 1 quarter 2 quarters 4 quarters comes important to determine how Rank RMSE Rank Indicator RMSE Rank RMSE the indicator would perform under different economic circumstances, 1 FF 2 3 2.160 3.793 2.859 and Table 1.4 tells us how well or TB03 3.969 9 3.075 6 2.260 5 how poorly our interest rate family 4 3.862 4 2.251 TB06 3.000 5 performs during recessions and 4 2.356 6 CM01 3.826 3 2.996 expansions. The strength of FF 7 3.094 7 2.483 CM03 3.876 5 deteriorates somewhat during both 3.144 2.552 recessions and expansions, when 3.936 7 8 8 CM05 compared to other interest rates. 3.949 3.249 10 2.683 9 CM10 8 On the other hand, EUR03 contin 1 2.754 1 2.222 3 EUR03 3.622 ues to perform strongly especially 2.827 2 2 2.216 CP6 3.880 6 during recessions, and CP6’s rank 3.197 9 2.725 10 BAA 4.006 10 ing improves during expansionary periods. It is also interesting to 11 11 2.819 11 NONE 4.015 3.358 note that our autoregressive indica NOTE: Sample period is July 1973 - December 1991, quarterly data. tor “NONE” ranks first in the Kalman forecasts during expan sions. This result demonstrates that sometimes indicators can be misleading shorter horizons. CP6, on the other hand, experi during expansionary periods. ences an out-of-sample deterioration at the one The cumulated residuals from the Kalman quarter horizon, but ranks second at both the two forecasts in Figure 1.1 show that, overall, the quarter and one year forecast horizons. In gener indicators in our interest rate family consistently al, our results indicate that shorter maturity instru underforecasted real GDP between 1974 and ments, namely FF and EUR03, outperform long 1982. The upward trend in the cumulated residuals during this TABLE 1.4 period can be explained in part by an unprecedented increase in infla Kalman 1 quarter ahead forecasts in tion, which caused interest rates to recessions and expansions rise without the normally anticipat Real GDP ed decline in output. On the other Actual Recession Expansion hand, between 1983 and 1989, FF, Rank RMSE Rank RMSE Rank RMSE Indicator CP6, EUR03, and all of the Trea 4 4 3.801 FF 3.793 2 3.753 sury bill rates performed well, as 3.941 3.969 9 4.108 8 9 shown by the flattening of their TB03 cumulated residuals’ slopes during 3.862 4 6 TB06 3.780 6 3.878 this period. Between 1990 and 2 3.857 CM01 3.826 3 3.663 5 1991, however, the indicators’ 7 CM03 3.876 5 3.722 3 3.905 performance deteriorated again, as 7 3.814 7 3.959 10 CM05 3.936 all of the interest rates missed the 11 3.949 8 3.766 5 3.983 CM10 1990-91 recession and consistently 3.622 1 1 3.625 2 EUR03 3.605 overforecasted real GDP. Figure 1.2 shows the dynamic 3.642 3.880 6 4.928 10 3 CP6 response of the forecasted growth BAA 10 4.377 9 3.930 8 4.006 rate of employment when FF in creases. Because the response 11 5.817 11 1 NONE 4.015 3.563 paths of our interest rate family are NOTE: Sample period is July 1973 - December 1991, quarterly data. virtually identical across all indica- 8 TABLE 1.3 ECONOMIC PERSPECTIVES TABLE 1.5 Multiperiod encompassing tests (Probability value for null hypothesis: X is encompassed by Y) Real GDP (1 quarter) Y FF TB03 TB06 CM01 CM03 CM05 CM10 0.796 0.830 0.061 — — — 0.403 0.723 0.803 0.601 EUR03 CP6 0.932 0.856 0.391 0.262 BAA Maximum P-value X FF n.a. TB03 0.482 n.a. TB06 0.945 0.204 n.a. 0.947 - — CM01 0.677 0.103 0.342 n.a. — — CM03 0.682 0.343 0.949 0.264 n.a. 0.065 CM05 0.637 0.375 0.910 0.412 0.251 n.a. 0.066 0.906 CM10 0.464 0.371 0.798 0.684 0.508 0.563 n.a. 0.818 EUR03 0.119 — — — — n.a. 0.272 — — — CP6 — — BAA 0.326 0.221 0.638 0.485 0.407 0.253 — — Ci.431 — 0.932 — 0.830 0.241 — 0.947 0.524 — 0.723 0.053 0.949 0.702 0.154 0.910 0.976 0.380 0.976 0.240 — 0.240 0.659 n.a. — 0.659 0.666 0.783 n.a. 0.783 Real GDP (2 quarters) FF — n.a. — — — 0.605 0.867 — 0.867 TB03 0.090 n.a. 0.925 0.310 — — — 0.340 — — 0.925 TB06 0.337 0.448 n.a. 0.220 — — 0.250 — — 0.448 CM01 0.582 0.515 0.864 n.a. — — — — 0.293 — — 0.864 CM03 0.617 0.959 0.443 0.109 n.a. — — 0.360 0.107 — 0.959 CM05 0.694 0.975 0.520 0.191 0.132 n.a. — 0.450 0.197 0.137 0.975 CM10 0.665 0.763 0.418 0.210 0.096 — n.a. 0.491 0.263 0.794 0.794 EUR03 0.231 — — — — — — — — — n.a. 0.598 — 0.598 0.837 0.429 0.228 CP6 0.214 — BAA 0.574 0.302 — — 0.635 0.340 n.a. — 0.340 0.989 0.742 n.a. 0.989 0.044 Real GDP (4 quarters) FF n.a. — — — — — — TB03 0.963 n.a. 0.15'2 — — — — 0.139 0.661 — 0.963 TB06 0.920 0.910 n.a. — n.a. — — — 0.255 0.662 — 0.920 — — 0.980 0.157 — 0.980 — 0.623 0.166 — 0.623 0.506 0.140 — 0.541 0.555 0.211 0.419 0.588 n.a. 0.785 — 0.785 0.052 0.767 n.a. 0.456 — 0.746 0.895 — CM01 0.596 0.373 — CM03 0.593 0.363 — — CM05 0.541 0.302 — — CM10 0.588 0.362 0.130 — 0.074 — n.a. — 0.072 — n.a. EUR03 0.539 0.263 '0.173 CP6 BAA 0.746 0.845 — — 0.534 — — 0.692 — — 0.895 — — 0.101 — 0.776 — 0.507 — n.a. NOTES: Values less than or equal to 0.05 are marked with a dash. Sample period is tors, we chose the federal, funds rate as an ex ample of how a one standard deviation increase in the interest rate today changes the growth rate of employment during the next 36 months. The forecasted growth rate of employment increases for approximtately two months and then falls, plunging to very deep negative val ues especially during the first year. Eventually, the growth rate moves very close to zero as The horizon expands, indicating that the change i n FEDERAL RESERVE BANK OF CHICAGO n.a. 1962 - December 1991, qu arterly data. FF does not impact employment forecasting after approximately two years. Finally, as shown in Table 1.5, our encom passing tests indicate that both FF and EUR03 contain significant information but neither of them dominates. This indicates that both inter est rates are close substitutes, and using both would not improve the forecasting results since either interest rate contains all of the necessary information. For example, at the one quarter 9 FIGURE 1.1 Interest rate levels Fed funds(FF) 5 year Tre asury bond (CM05) cumulated Kalman residuals 100 cumulated I Caiman residuals 75 r 3 month Treasury bill (TB03) 10 year "I Treasury bond (CM 10) 100 r- 75 r r — L_i I I ,1 1 1 .1 i i i 6 month Treasury bill (TB06) 3 month eurodollar (EUR03) 100 r 1 year Treasury bond (CM01) 6 month cornmerciaT paper (CP6) 100 r 100 r 75 h 75 - 50 50 - 25 25 |- 0 -25 3 year Treasury bond (CM03) 100 10 BAA corporate bond (BAA,1 100 75 - 50 - 25 - jl973 76 79 i i i i i i » i FIGURE 1.2 TABLE 2.1 Dynamic response of employment to FF Classical goodness-of-fit statistics annualized percent growth rates R2 Correlation with real GDP P-value Rank MBSTL 0.166 0.034 0.1744 7 MB 0.145 0.013 0.4734 9 M1 0.172 0.157 0.1284 5 M2 0.219 0.236 0.0084 4 M3 0.169 0.246 0.1483 6 Indicator L 0.164 0.239 0.1993 8 DBTNF 0.124 0.180 0.9352 10 M1R 0.250 0.297 0.0012 2 M2R 0.346 0.353 0.0000 1 NBRX 0.249 0.154 0.0012 3 NOTE: Sample period is January 1962 - December 1991, quarterly data. forecast horizon EUR03 encompasses all of the other indicators, but at the same time, EUR03 is encompassed by FF and CP6. However, because EUR03 ranked first in the in-sample forecasts at the one quarter horizon, and in the out-of-sample forecasts at the one and two quarter horizons, it is selected as our best indi cator at both the one and two quarter forecast horizons. Similarly, FF is chosen as the best indicator at the one year forecast horizon for its strong performance in-sample and out-of-sam ple when the forecast horizon increases. the indicators in our family of money based measures are annualized log differences. Goodness-of-fit statistics in Table 2.1 show that all of the money based indicators are posi tively correlated with real GDP. Not surpris ingly, as the endogenous components of the monetary aggregate increase, the contempora neous correlation with economic activity rises. Moreover, the broader monetary aggregates seem to impact real GDP more than the narrow- 2 . M o n e y based m easures Table 2.1 lists the monetary indicators we selected for investiga tion: a measure of the monetary base developed by the Federal Reserve Bank of St. Louis (MBSTL); the Board of Governors’ monetary base6 (MB); Ml; M2; M3; L;7 long term debt of domestic nonfinancial institutions (DBTNF); real Ml (MIR) and real M2 (M2R) both deflated by the consumer price index; and NBRX, which is the ratio of nonborrowed reserves at time t to total reserves at time t - 1. Strongin (1991) found that this normalized reserve aggregate (NBRX) contains much of the in formation about monetary policy actions which Sims (1991) at tributes to innovations in the federal funds rate. Except for NBRX, all of FEDERAL RESERVE BANK OF CHICAGO TABLE 2.2 Multiperiod forecasts, in-sample Indicator ___________ 1 quarter R2 Rank Real GDP 2 quarters R2 Rank 4 quarters R2 Rank MBSTL 0.166 7 0.154 8 0.102 8 MB 0.145 9 0.144 9 0.121 5 Ml 0.172 5 0.183 7 0.096 10 M2 0.219 4 0.249 4 0.186 4 M3 0.169 6 0.189 5 0.107 7 L 0.164 8 0.184 6 0.097 9 DBTNF 0.124 10 0.133 10 0.121 6 M1R 0.250 2 0.288 3 0.244 3 M2R 0.346 1 0.447 1 0.514 1 NBRX 0.249 3 0.327 2 0.292 2 NONE 0.118 11 0.123 11 0.076 11 NOTE: Sample period is January 1962 - December 1991, quarterly data. 11 er measures of money. This is TABLE 2.3 probably due to the fact that broad Kalman multiperiod forecasts, out-of-sample er money measures consist of a Real GDP larger number of components, each 1 quarter 2 quarters 4 quarters associated with movements in eco RMSE Rank RMSE Rank Indicator RMSE Rank nomic activity. M2R, M 1R, M3, and L have the largest correlation MBSTL 3.474 2.904 4.108 7 10 8 coefficients with GDP, and M2R MB 4.114 7 7 8 3.426 2.840 and M 1R also show the strongest fit 4.149 M1 10 3.455 9 2.992 11 to the model, as their R2s rank first 3.944 4 M2 3 3.252 3 2.809 and second, respectively. NBRX 3.394 M3 4.073 5 6 2.948 10 and M2 are also statistically signifi L 3.432 4.136 9 8 2.926 9 cant, ranking third and fourth, re spectively. DBTNF 4.242 11 11 3.495 2.820 6 The predictive power of our M1R 4 4.097 6 3.285 2.775 3 money based indicators is then 3.674 M2R 1 2.844 1 2.219 1 tested at different forecast horizons, NBRX 2 2 2 3.799 3.003 2.550 and in-sample results shown in Table 2.2 indicate that M2R, MIR, NONE 4 2.819 4.015 3.358 5 5 NBRX, and nominal M2 all contin NOTE: Sample period is July 1973 - December 1991, quarterly data. ue to perform well, providing addi tional information to the forecasts as the horizon increases. M2R, however, clearly has the strongest predictive Figure 2.1 provide another perspective of the out-of-sample performance of our family of power at all forecast horizons (ranking always first), while M IR’s ranking slightly deteriorates money based measures. In our case, the best as the forecast horizon increases. On the other indicator is again M2R as its cumulated residu als’ path clearly stays near zero values, except hand, NBRX’s performance improves at the two quarter and four quarter horizons, ranking for isolated periods of large forecast errors in second in both. 1978 and 1981, when M2R underforecasted Once again, to see how the indicators would actually perform TABLE 2 .4 using only data prior to the fore Kalman 1 quarter ahead forecasts in casting period, we use Kalman recessions and expansions filtering techniques. Out-of-sample Kalman forecast results in Table Real GDP 2.3 show M2R and NBRX to have Actual Recession Expansion Indicator RMSE Rank RMSE Rank RMSE Rank the strongest fit at all horizons, as shown by their individual RMSEs, MBSTL 4.108 7 5.774 8 3.700 6 while MIR’s performance greatly MB 4.114 8 5.534 7 3.777 7 improves in the long run. As M1 4.149 10 4 5.245 3.901 9 shown in Table 2.4, M2R also con M2 3.944 3 11 3.402 6.011 2 sistently performs well under differ ent circumstances, and especially M3 4.073 5 5.848 10 3.631 5 during expansionary periods. On L 4.136 9 5.256 5 3.883 8 the other hand, while MIR is a DBTNF 4.242 11 5.400 6 11 3.980 good predictor during recessions, its M1R 4.097 1 3.949 6 4.793 10 performance considerably worsens M2R 3.674 1 5.109 2 1 3.326 during expansions. NBRX’s per NBRX 2 3.799 5.228 3 3.454 formance is noticeably consistent 3 during recessions and expansions, NONE 4 4.015 5.817 9 3.563 4 as it ranks third during both. The cumulated residuals from NOTE: Sample period is July 1973 - December 1991, quarterly data. the Kalman forecasts shown in 12 ECONOMIC PERSPECTIVES FIGURE 2.1 Money based measures St. Louis monetary base (MBSTL) Nominal L (L) cumulated Kalman residuals cumulated Kalman residuals FRB monetary base (MB) Nominal nonfinancial debt (DBTNF) Nominal M2 (M2) Real M2 (M2R) Nominal M3 (M3) FEDERAL RESERVE BANK OF CHICAGO NBR/TR ratio (NBRX) 13 TABLE 2.5 Multiperiod encompassing tests (Probability value for null hypothesis: X is encompassed by Y) Real GDP (1 quarter) Y MBSTL MB M1 M2 M3 L DBTNF M1R M2R NBRX Maximum P-value 0.763 X n.a. 0.064 0.150 0.462 0.178 0.094 0.763 0.759 0.411 MB 0.726 n.a. 0.307 0.569 0.296 0.224 0.075 0.682 0.936 0.500 0.936 Ml 0.055 — n.a. 0.506 0.105 0.054 — 0.658 0.855 0.671 0.855 MBSTL M2 — — — n.a. — — — 0.098 0.954 — 0.954 M3 0.138 — 0.135 0.733 n.a. 0.174 — 0.327 0.653 0.149 0.733 0.449 L 0.119 — 0.136 0.407 0.324 n.a. — 0.322 0.449 0.286 DBTNF 0.694 0.755 0.669 0.771 0.716 0.825 n.a. 0.829 0.970 0.755 0.970 MIR — — — — — — — n.a. 0.924 — 0.924 M2R — — — — — — — — NBRX n.a. — 0.000 0.286 n.a. 0.286 Real GDP (2 quarters) n.a. 0.266 0.760 0.817 0.484 0.359 — 0.959 1.000 0.595 n.a. 0.516 0.654 0.477 0.445 0.167 1.000 0.686 0.954 MB 0.994 0.722 0.994 M1 — — n.a. 0.667 0.112 — — 0.803 0.970 0.845 0.970 M2 — — — n.a. — — — 0.173 0.833 0.119 0.833 M3 — — 0.064 0.603 n.a. 0.197 — 0.323 0.560 0.193 0.603 0.333 MBSTL — — — 0.258 0.294 n.a. — 0.274 0.284 0.333 0.490 0.715 0.604 0.691 0.533 0.697 n.a. 0.774 0.973 0.745 0.973 M IR — — — — — — — n.a. 0.752 0.101 0.752 M2R — — — — — — — — L DBTNF NBRX n.a. — 0.000 0.133 n.a. 0.133 Real GDP (4 quarters) n.a. 0.930 0.341 0.604 0.525 0.344 0.659 0.840 0.782 0.896 0.930 MB 0.336 n.a. 0.126 0.248 0.228 — 0.330 0.362 0.817 0.464 0.817 0.987 MBSTL M1 0.658 0.693 n.a. 0.914 0.669 0.439 0.517 0.987 0.841 0.958 M2 — — — n.a. — — — 0.263 0.430 0.400 0.430 M3 0.452 0.392 0.424 0.612 n.a. 0.442 0.375 0.776 0.918 0.746 0.918 L 0.521 0.523 0.396 0.523 0.626 n.a. 0.652 0.802 0.824 0.975 0.975 DBTNF 0.196 0.331 0.072 0.230 0.209 0.089 n.a. 0.300 0.836 0.334 0.836 — — — — — — — n.a. 0.257 0.305 0.305 M1R M2R — — — — — — — — n.a. — 0.000 NBRX — — — — — — — — 0.473 n.a. 0.473 NOTES: Values less than or equal to 0.05 are marked with a dash. Sample period is January 1962 - December 1991, quarterly data. economic activity. M2R’s performance was again noticeably good between 1990 and 1991, when most of the other money based indicators clearly failed to predict the recession. NBRX was relatively stable from 1973 to 1981, but has shown a consistent pattern of overforecasting output growth since 1982. This deterioration may be due to increasing reluctance on the part of banks to borrow from the discount window. The performance of other monetary aggregates is less reliable and clearly more volatile than 14 the behavior of M2R and NBRX. For example, the two measures of the monetary base and M 1 consistently underforecasted real GDP between 1974 and 1977, as shown by their upward slop ing paths. Overall, the path of nominal aggre gates plunged during the credit control program of 1980, overpredicting output growth during the mild recession. From 1983 to 1988, these nominal aggregates performed fairly well, exhibiting uncharacteristic stability, except for M 1 which did substantially worse between ECONOMIC PERSPECTIVES contains unique information and that adding another money based indicator to the model would add no additional information. 3. Interest rate spreads 1983 and 1984. Finally, between 1990 and 1991, there was a considerable deterioration in the performance of M l, L, and the two measures of the monetary base, as they consistently overpre dicted economic growth. To see how changes in money based mea sures affect the forecaster’s expectations over time, we look at the dynamic response of em ployment to our strongest indicator, M2R. Fig ure 2.2 shows the response of the forecasted growth rate of employment when M2R increases by one standard deviation. In general, a positive impulse in a money based indicator leads to an increase in employment growth rates. In our case, the response to a one standard deviation increase in M2R is quick and persistent over a period of appro ximately 15 months, with the maximum impact occurring within the first year. These results indicate that the impact of changes ii M2R on real economic activity is very strong, afhough somewhat short lived. Finally, we test our family of money based measures to determine the degree of independent information they contribute to the model individ ually. As shown in Table 2.5, our encompassing tests slow that M2R is clearly the dominant indicator within our family of monetary aggre gates. Ift fact, M2R is not encompassed by any of the other indicators at all forecast horizons. The row labeled M2R in the table has dashes, indicating that the hypothesis that M2R is en compassed by any of the other indicators is con sistently rejected. Similarly, the high signifi cance levels in the column labeled M2R indicate that M2R encompasses all of the other indicators at all forecast horitons. This suggests that M2R FEDERAL RESERVE HANK OF CHICAGO Recent research on financial market indica tors of economic activity has brought renewed attention to interest rate spreads. Laurent (1988), Bemanke (1990), Estrella and Hardouvelis (1991) , Friedman and Kuttner (1992), Kashyap, Stein, and Wilcox (1991), and Stock and Watson (1989b) all have suggested and tested various interest rate spreads as predictors of economic activity with significant success. The idea behind most of these spreads is that the difference in yields between two different debt instruments has a greater informational content than interest rate levels. The two primary types of interest rate spreads that have been used are risk spreads which measure the difference in yield between a private debt instrument and a government bond of equivalent maturity, and term spreads which measure the difference in yield between two government debt instruments of different maturities. Typically, risk spreads contain information useful to the forecaster because the return on the private debt instrument is a measure of the mar ket’s assessment of the near term risk in the relevant business environment, and higher re turns are usually associated with higher per ceived business risk. Friedman and Kuttner (1992) have argued that this interpretation is probably flawed since the spreads are typically too large to be explained by any reasonable estimate of the risk inherent in the private debt instruments. Therefore, they suggest that liquidi ty considerations play a significant role in the pricing of private/public spreads. Following their lead, we will also refer to these spreads as private/public spreads. Term spreads seek to measure the market’s perception of the relative availability of credit through time. The convention is that the yield on the debt instrument with the shorter maturity is subtracted from the yield on the instrument with the longer maturity. Thus, a positive spread would indicate that short term funding is cheaper than long term funding, therefore boosting cur rent economic activity. An alternative explana tion is that the higher long term yields may sig nal expectations of higher future credit demand resulting from increased economic activity. An additional interpretation is that by taking the difference between long and short term interest 15 rates, the short term rate is corrected for changes TABLE 3.1 in inflationary expectations and taxes, leaving a Classical goodness-of-fit statistics better measure of short run credit conditions. In Correlation any case, all of these term spread regressions Indicator R2 with real GDP P-value Rank have the counterintuitive implication that a rise in long term interest rates is good for the near TB3FF 0.327 0.449 0.0000 3 term outlook of the economy. Estrella et al. TB6FF 0.321 0.442 0.0000 4 (1991) and Strongin (1990) attempt to reconcile TB12FF 0.330 0.425 0.0000 2 the term spread results with current theory, how CM05FF 0.302 0.321 0.0000 6 ever with limited success. CM10FF 0.309 0.309 0.0000 5 As shown in Table 3.1, we tested seven term TB12TB3 0.238 0.225 0.0026 9 spreads and three private/public spreads.8 Five of the seven term spreads are based on the feder CM10CM1 0.284 0.170 0.0001 8 al funds rate (FF), and they are: the 3 month EUROTB3 0.294 -0.378 0.0001 7 Treasury bill rate less FF (TB3FF); the 6 month CP6TB6 0.339 -0.431 0.0000 1 Treasury bill rate less FF (TB6FF); the 12 month BAACM10 0.234 -0.297 0.0033 10 Treasury bill rate less FF (TB12FF); the 5 year Treasury constant maturity bond rate less FF NOTE: Sample period is January 1962 - December 1991, quarterly data. (CM05FF); and the 10 year Treasury constant maturity bond rate less FF (CM 1OFF). TB3FF is a short term spread; TB6FF is a medium term spread; and TB12FF, CM05FF, and CM 1OFF are coefficient in absolute terms. An increase in all long term spreads. Our term spreads also the yield on private debt instruments may signal a riskier economic environment, which is then include two intermediate spreads: the difference associated with a decline in investment and a between the 12 month and the 3 month Treasury drop in output. In this case, if the return on bill rates (TB12TB3), and the difference between the 10 year and the 1 year Treasury constant public instruments is unchanged, the private/ public spread increases while economic activity maturity bond rates (CM10CM1). The three private/public spreads we investi declines. CP6TB6 has also the strongest fit to the model, as shown by its R2, followed by gated are: the 3 month eurodollar rate less the 3 TB12FF and TB3FF. month Treasury bill rate (EUROTB3); the 6 month commercial paper rate less the 6 month Treasury bill rate TABLE 3.2 (CP6TB6); and the BAA corporate Multiperiod forecasts, in-sample bond rate less the 10 year Treasury constant maturity bond rate Real GDP 1 quarter 2 quarters 4 quarters (BAACM10).9 Indicator R2 Rank R2 Rank R2 Rank Goodness-of-fit statistics in Table 3.1 indicate that all of our TB3FF 0.327 3 0.446 3 0.437 5 term spreads are positively associat TB6FF 4 0.321 2 0.459 0.490 4 ed with real GDP, with the short TB12FF 2 1 0.330 0.470 0.518 1 and medium spreads showing the CM05FF 0.302 6 0.428 6 0.498 2 strongest correlation coefficients. The positive association is not CM10FF 4 0.309 5 0.435 0.491 3 surprising given that short term TB12TB3 0.238 9 0.333 9 0.383 7 interest rates tend to be more vola CM10CM1 0.284 8 0.374 7 0.396 6 tile than long term interest rates, 0.294 7 0.364 EUROTB3 8 0.230 9 and that a decline in short term 0.2FJ9 CP6TB6 0.339 1 0.429 5 8 interest rates is typically associated 0.234 BAACM10 10 0.175 10 0.138 10 with a steepening of the yield curve. On the other hand, private/ NONE 11 11 0.118 0.123 0.076 11 public spreads are negatively corre lated with GDP, with CP6TB6 NOTE: Sample period is January 1962 - December 1991, quarterly data. having the strongest correlation 16 ECONOMIC PERSPECTIVES The pred ictive power of our famil y of int erest rate spreads is Kalman multiperiod ffoneoastey out-of^sampib next tested at different forecast Real GDP horizons, and in-sample results in 1 quarter 2 quarters 4 quarters Table 3 Si show a strong deteriora RMSE Rank Rank RMSE Rank RMSE Indicator tion in \he performance of CP6TB6 at the two and four quarter forecast 5 2.253 2.674 1 1 TB3FF 3.609 horizons, while the strength of 2 2.081 2.691 2 TB6FF 3.691 3 TBUiFF improves considerably in 1 2.015 2.754 6 3 TB12FF 3.753 the kong run. In general, the predic 3 2.111 6 5 2.811 CM05FF 3.745 tive power of medium and long 4 2.161 7 2.785 5 3.763 CM10FF team spreads seems to improve as 6 2.370 'the forecast horizon increases. 3.187 4.197 11 9 TB12T B3 Also, term spreads perform better 7 2.389 8 CM10'CM1 3.857 8 2.970 than private/public spreads across 2.721 8 7 4 2.886 3.698 EUROTB3 horizons, except for CP6TB6, ;2.744 9 4 2.760 3.656 2 CP6T B6 which is the strongest indicator at 11 2’.846 11 9 3.485 BAACM10 3.983 the one quarter forecast horizon. This scenario is virtually unchanged 10 2.819 10 3.358 10 NONE 4.015 in the out-of-sample Kalman fore NOTE: Sample period is July 1973 - December 1991, quarterly data. casts shown in Table 3.3. As we test the actual performance of our indicators using only data available formed fairly well from 1973 to 1980, tlv;y prior to the forecasting period, we see tha t clearly failed during the last three recessio ns. In CP6TB6 remains very strong in the short run, fact, they all underforecasted economic activity although its ranking somewhat deteriorates between 1980 and 1982, and then overpredicted when compared to in-sample results. Although real GDP between 1990 and 1991. Between the out-of-sample performance of TB12FF at 1982 and 1989, their path was conspicuously short term horizons considerably worsens, its strength increases at the four quarter forecast flat. This suggests that these spreads do well in horizon, as its RMSE ranks first. Under different circumstances, we TABLE 3 .4 see that overall, private/public spreads, such as CP6TB6 and H&lftianiUquarter ahead 1fftnceastft in EUROTB3, perform better during reressions ami cocpamions expansionary periods than our term Real GDP spreads, as shown in Table 3.4. Actual Recession Expansion On the other hand, term spreads Rank Indicator RMSE Rank RMSE Rank RMSE outperform private/public spreads TB3FF 1 1 3.447 3 3.609 4.353 during recessions, as TB3FF and TB6FF 4 3.691 3 4.634 3 3.479 TB12FF rank first and second, respectively, according to their TB12FF 2 3.753 6 4.599 3.566 9 individual RMSEs. CM05FF 3.745 5 4.714 5 3.527 6 The cumulated residuals from CM10FF 7 3.521 5 3.763 4.823 6 the Kalman forecasts in Figure 3.1 4.197 11 TB12TB3 11 5.172 8 3.980 show some striking similarities in CM10CM1 3.857 4.707 4 3.670 10 8 the overall forecasting performance 7 2 EUROTB3 3.698 4 4.987 3.393 of our family of interest rate spreads. Except for TB3FF, CP6TB6 2 5.727 10 3.099 1 3.656 TB6FF, and TB12FF, all of our 3.557 7 BAACM10 3.983 5.698 9 9 spreads tend to overforecast real GDP, as shown by their consistent NONE 5.817 11 3.563 8 4.015 10 ly negative residuals. While NOTE: Sample period is July 1973 - December 1991, quarterly data. TB3FF, TB6FF, and TB12FF per FEDERAL RESERVE BANK OF CHICAGO 17 FIGURE 3.1 12 month T bill less 3 month T bill (TB12T&) cumulated Kalman residuals 75 r 3 month eurodollar less 3 month T bill (EUROTB3) f i i » i i i i i i i » i 5 y e a rT bond less fed funds (CM05FF) i ■ i i i 6 mo. commercial paper less 6 mo. T bill (CP6TB6) 25 r fcMatttiffc p § iftp § e ff*f§ FIGURE 3.2 Dynamic response of employment to interest rate spreads 3 month T bill less fed funds (TB3FF) 12 month T bill less 3 month T bill (TB12TB3) annualized percent growth rates annualized percent growth rates 0.80 r 0.75 r- 0.50 - 0.25 - 0.40 0.20 0.00 0.00 0.25 6 month commercial paper less 6 month T bill (CP6TB6) 5 year T bond less fed funds (CM05FF) 0.25 0.00 0.25 0.50 0.75 1 11■ ■ ■ ■ 1■ * ‘ ‘ ■■' ■■■ ■■‘ * ■ ■ ■ ■ -1.00 ■*‘ BAA corporate bond less 10 ye a rT bond (BAACM10) 0.22 0.45 0.00 0.15 - 0.22 0.00 -0.44 - 0.66 -0.15 ■ i . i i i i i i i i i i i i i i i i i i i i i i i i i i i i i i i i ■ .......................... ■ I I 1 I 1 I I 1 I I I I I 1 1 1 t 1 1.1 L J 0 5 FEDERAL RESERVE 10 15 20 months BANK OF CHICAGO 25 30 35 0 5 10 15 20 25 30 35 months 19 clearly overpredicted real GDP. All of the private/public spreads followed the same gener al pattern of mediocre performance from 1973 to 1981, and persistent overprediction of economic activity thereafter. In general, we conclude that, although a persistent bias in forecasting exists in all of the interest rate spreads we investigated, some of them did fairly well during most of our sample period, but failed during periods of large scale financial restructuring. forecasting “normal” periods of economic ac tivity, but periodically fail in predicting reces sions. Although CM05FF and CM 1OFF follow a similar pattern between 1973 and 1981, after 1982 their cumulated residuals’ path never stabilized but plunged to persistently negative values. Our intermediate term spreads (TB12TB3 and CM10CM1) failed during all of the recessions in our sample period (including the 1973-1975 recession), and developed a consistently negative bias after 1982, as they T A B L E 3 .5 M u l t i p e r i o d e n c o m p a s s in g te s ts ( P r o b a b ility v a lu e f o r n u ll h y p o th e s is : X is e n c o m p a s s e d b y Y ) Real GDP (1 quarter) Y TB3FF TB6FF TB12FF CM05FF CM10FF TB12TB3 CM10CM1 EUROTB3 CP6TB6 BAACM10 Maximum P-value X TB3FF TB6FF TB12FF CM05FF n.a. 0.462 0.167 0.105 0.185 n.a. 0.227 — — 0.093 0.109 0.389 0.062 0.333 0.231 0.797 — 0.106 EUR0TB3 — 0.125 0.215 0.430 0.454 — — — — CP6TB6 0.066 — — — — — — — — — — — — — 0.072 0.053 CM10FF TB12TB3 CM10CM1 BAACM10 0.999 n.a. — — n.a. — — 0.540 n.a. 0.413 0.699 0.156 0.109 — — — — — — — — — 0.098 0.077 n.a. 0.115 — — — — — n.a. — — 0.055 — n.a. 0.186 0.053 — — — — — — 0.185 0.999 0.227 0.540 0.231 0.797 — 0.699 — 0.186 n.a. — 0.104 n.a. 0.066 0.104 Real GDP (2 quarters) TB3FF TB6FF n.a. 0.070 0.569 0.467 — — — — — — — 0.569 n.a. — — — — — — — — — 0.155 0.092 0.798 n.a. 0.337 — — — — — — — — — 0.798 0.155 0.665 0.055 0.206 0.665 n.a. — — — — n.a. 0.271 — — — — — 0.271 — — 0.214 0.256 0.755 0.370 0.353 n.a. 0.071 — 0.126 0.876 — — — — n.a. — — — — — — 0.580 — — — — — — — 0.459 0.710 — — n.a. — 0.222 — — 0.908 0.991 0.436 0.935 0.807 n.a. 0.936 0.322 n.a. 0.548 0.197 — — — — n.a. 0.176 0.144 0.092 — — — — 0.131 0.056 — — — — — — — — — — — n.a. 0.720 0.170 n.a. — — — — — 0.576 0.062 0.333 0.593 0.261 0.230 — — — — — — 0.979 0.965 0.783 0.930 TB12FF CM05FF CM10FF TB12TB3 CM10CM1 EUR0TB3 CP6TB6 BAACM10 0.093 0.545 n.a. 0.755 0.876 0.222 0.093 0.991 Real GDP (4 quarters) TB3FF TB6FF n.a. — — TB12FF CM05FF CM10FF — — — — — 0.142 EUR0TB3 — 0.752 — 0.989 CP6TB6 0.883 0.870 0.840 0.899 0.774 BAACM10 0.500 0.392 0.442 0.863 TB12TB3 CM10CM1 — — n.a. — — 0.094 n.a. 0.428 — 0.111 0.548 0.197 0.027 0.176 0.720 0.576 0.593 — — — 0.560 — 0.115 n.a. — n.a. — 0.989 0.883 0.973 0.569 0.575 n.a. 0.973 NOTES: Values less than or equal to 0.05 are marked with a dash. Sample period is January 1962 - December 1991, quarterly data. 20 ECONOMIC PERSPECTIVES To see how changes in interest rate spreads may affect a forecaster’s analysis of real eco nomic activity over time, we look at the dynamic response of forecasted employment growth rates to a one standard deviation increase in our family of spreads. The paths depicted in Figure 3.2 show substantial differences between the re sponse to changes in the term spreads and chang es in the private/public spreads. The response of forecasted employment growth rates when BAACM 10 increases is a quick dip in the first two months followed by a fast jump which peaks after eight months and then dies quickly. The response to the two shorter term private/public spreads (EUROTB3 and CP6TB6) follows an exact opposite path, first declining rapidly for approximately ten months, and then rapidly flat tening. With the exception of TB12TB3, the response paths of the term spreads are all very similar, with employment growth rates increasing slowly, peaking at approximately ten months, and then flattening thereafter. This means that either a decline in short term interest rates or a rise in long term interest rates would cause fore casters to increase their predictions of future economic activity. The scenario depicted thus far indicates that the strength of the BAACM10 spread is in very short forecast horizons, as its impact on real economic activity dies fairly quickly compared to other spreads. On the other hand, our analysis shows that the strength of CP6TB6 is in the short and medium forecast horizons, while term spreads’ overall impact on real economic activity is extremely persistent. The results of our encompassing tests shown in Table 3.5 are exactly what we would have expected, given our analysis thus far. That is, we need to look at both a private/public spread and a term spread to obtain all of the information nec essary for forecasting economic activity using interest rate spreads. This is due to the fact that term spreads usually perform better at longer horizons, while private/public spreads have a stronger predictive power at shorter horizons. CP6TB6 and TB12FF dominate their respective groupings. At the four quarter horizon, CP6TB6 no longer contains additional information beyond that contained in TB12FF. Now, however, a longer horizon term spread such as CM 1OFF is also necessary to fully cover the information set. It is interesting to note that the analysis of all of the encompassing results indicates that the sepa ration between the private/public spreads and the term spreads is not very clear. In fact, at some FEDERAL RESERVE BANK OF CHICAGO forecast horizons the results reverse. This indi cates that there are common multiple driving forces in the determination of these spreads, and that the driving factors associated with longer horizons of economic activity predominate in the term spreads, while the common factors that drive short run performance dominate the pri vate/public spreads. 4. Composite indicators Table 4.1 lists the composite indicators we investigated: the National Bureau of Economic Research (NBER) Experimental Leading Index (XLI); the NBER Nonfinancial Experimental Recession Index10 (XRI2); the Department of Commerce (DOC) Composite Index of Leading Indicators (LEAD); the Purchasing Managers’ Index (PMI) of the National Association of Pur chasing Management (NAPM); the Standard and Poor’s 500 Stock Index (S&P); the percent change in sensitive materials prices (SMPS);" and the Kashyap-Stein-Wilcox “mix” (KSWMIX), which is the ratio of bank lending to the sum of bank lending and commercial paper lending [see Kashyap et al. (1991)]. Note that the NBER Experimental Leading Index includes the 10 year Treasury bond/1 year Treasury bond spread and the 6 month commercial paper/6 month Treasury bill spread, while the Depart ment of Commerce Composite Index of Leading Indicators includes real M2, all of which have been discussed in previous sections. The two composite leading indicators and the NBER Nonfinancial Experimental Recession Index are designed to predict economic activity at a six month horizon, although the optimization for the Department of Commerce Index is not as specifTABLE 4.1 Classical goodness-of-fit statistics R2 Correlation with real GDP P-value XLI 0.455 0.547 0.0000 1 XRI2 0.385 -0.649 0.0000 3 LEAD 0.405 0.600 0.0000 2 PMI 0.265 0.632 0.0005 4 S&P 0.205 0.185 0.0222 7 SMPS 0.232 0.278 0.0045 6 KSWMIX 0.243 0.316 0.0023 5 Indicator Rank NOTE: Sample period is January 1963 - December 1991, quarterly data. 21 ic as either of the NBER indices. TABLE 4.2 Except for S&P and LEAD, which Multiperiod forecasts, in-sample are annualized log differences, all Real GDP of the indicators in our family of 'I quarter 2 quarters 4 quarters composite indicators are used in Indicator Rank R2 Rank R2 R2 Rank levels. Also, because data on the XRI2 start in January 1962, our 1 1 1 XLI 0.455 0.568 0.401 sample period for this family of XRI2 0.382 3 0.316 3 0.168 6 indicators starts in January 1963. 0.341 LEAD 0.405 2 2 0.247 2 Goodness-of-fit tests in Table PMI 4 7 0.265 0.203 5 0.173 4.1 show that, except for XRI2, our 7 7 S&P 0.205 0.216 5 0.152 composite indicators have a posi SMPS 0.232 6 0.206 6 0.229 3 tive correlation with contemporane ous economic activity. XRI2 has 0.249 4 KSWMIX 0.243 4 5 0.193 the strongest correlation with real NONE 0.117 0.117 0.072 8 8 8 GDP in absolute terms, while XLI has the strongest fit to the model as NOTE: Sample period is January 1963 - December 1991, quarterly data. it ranks first according to its R2. LEAD and XRI2 also show consid erable strength as their R2s rank second and third, sionary periods. As expected, XRI2 is our best performer during recessions. respectively. The predictive power of our family The cumulated Kalman residuals in Figure of composite indicators is then tested in-sample 4.1 show some striking similarities and some at different forecast horizons. The results report ed in Table 4.2 show that XLI and LEAD contin differences in actual performance across these indicators. Except for KSWMIX, all of our ue to perform very well at all forecast horizons, while XRI2 loses strength at the four quarter composite indicators have overforecasted real horizon. PMI and S&P continue to show GDP over time, as their cumulated residuals are consistently negative. This bias is clearly weakness, especially in the long run, while evident during recessions and becomes more SMPS’ performance slightly improves at the four quarter horizon. dramatic after 1980. After 1982, while the negative bias is exacerbated in XLI and S&P, The results of out-of-sample Kalman tests the path becomes somewhat more stable for in Table 4.3 show a picture very similar to the in-sample results, as XLI continues to rank first most of our indicators. XRI2 is our best per former during this period, which is not surprisacross horizons. LEAD continues to rank sec ond, except for a slight deteriora tion in the four quarter forecast TABLE 4 .3 horizon where it ranks third. XRI2 has again a strong predictive power Kalman multiperiod forecasts, out-of-sample in the short run, while its perfor ______ Real GDP mance worsens at the four quarter 1 quarter 2 quarters 4 quarters horizon. XRI2’s behavior is ex Indicator RMSE Rank RMSE Rank RMSE Rank pected, however, as the indicator XLI 1 1 1 3.246 2.376 2.392 was created to forecast recessions XRI2 3.427 3 3.026 with a six month horizon. 3 2.758 5 Under different circumstances LEAD 3.307 3.024 2 2 2.669 3 we notice that XLI loses some of its PMI 4 4 3.838 3.319 6 2.736 strength outside of “normal” eco 3.964 4 S&P 6 3.253 2.758 6 nomic activity, as shown in Table SMPS 3.914 5 3.306 5 2.612 2 4.4. That is, XLI’s predictive pow KSWMIX 3.377 4.078 8 8 2.846 8 er is slightly weaker during both recessions and expansions. On the NONE 4.052 7 7 7 3.369 2.799 other hand, LEAD performs well NOTE: Sample period is July 1973 - December 1991, quarterly data. during expansions, although its performance worsens during reces 22 ECONOMIC PERSPECTIVES The dynamic responses of fore casted employment growth rates to Kalman 1 qu^termb^d'fp^eie^s in changes in our composite indicators in Figure 4.212 show somewhat simi _________________ Raal GDP ._ lar patterns for XLI and LEAD, Actual Recession Expansion where the response peaks quickly RMSE Rank Indicator RUrf RankRMSE_ —Bank. within approximately five months. 1 4.65‘n XLI 3.246 From the peak, both graphs exhibit 2 2 2.895 significantly different behaviors. 3.427 4.321 3 XRI2 3 1 3.226 The path in the XLI graph stabilizes 3.307 2 5.148 LEAD 1 4 2.814 for four to five months and then 4 5.879 PMI 3.838 4 3 3.300 drops off before the end of the year, 8 3.964 5.919 6 S&P 6 3.460 while the path in the LEAD graph 3.914 5.711 5 5 SMPS 5 3.460 falls more quickly and more dramat 4.724 3 KSWMIX 4.078 8 ically, until the impact of the indica 3.941 8 tor on real economic activity disap 5.894 7 4.052 7 NONE 7 3.588 pears. The path of XRI2 is inverted instead when compared to the path daU NOTE: Sample period is July 1973 December 1991, quarterly of the two leading indicators. In fact, as the graph shows, the reing since the index was originally developed in ponse path plunges very rapidly during the first response to the failure of XLI to forecast the tve months, then increases for another six 1990-1991 recession. months, and finally stabilizes thereafter. The FEDERAL RESERVE TAtJLE 4.4 BANK OF CHICAGO 23 FIGURE 4.1 NBER Experimental Leading Index (XLI) S&P 500 Stock Index (fi&P) cumulated Kalman residuals cumulated Kalman 25 r 25 -25 -75 NBER Nonfinancial Recession Index (XRI2) Change in sensitive materials prices (SMPS) 50 r 50 r - ■ 25 - -50 - -50 ,75 * I ■ I ..............I l I I I I I I I 1_ i DOC Composite Index of Leading Indicators (LEAD) Bank lending/(bank lending + CP) ratio (KSWMIX) 50 r -25 - Purchasing Managers’ Index (PMI'; 50 r response path of employment to changes in PMI and SMPS shows dramatic jumps in fore casted growth rates within the first two months. Employment growth then steadily falls in PMI while it flattens in SMPS. The S&P graph shows a path similar to that depicted in the PMI graph, except for a rapid drop in the first month. It is interesting to note that all of these dynamic response paths are virtually insignifi 24 cant at the one year mark, although the initial impact on real economic activity is fairly strong and well defined. Finally, as a group, these indicators seem to hold a lot of information about short run changes in economic activity, with most of that information centered at the three to nine month horizon. The encompassing results in Table 4.5 show that XLI strongly dominates this entire ECONOMIC PERSPECTIVES FIGURE 4.2 Dynamic response of employment to composite indicators NBER Experimental Leading Index (XLI) Purchasing Managers’ Index (PMI) annualized percent growth rates annualized percent growth rates NBER Nonfinancial Recession index (XRI2) DOC Composite Index of Leading Indicators (LEAD) family of indicators, especially at the two and four quarter forecast horizons. At the one quar ter horizon, both LEAD and XRI2 are not en compassed by any of the other indicators. These results are not surprising in light of the statistical results discussed earlier and the fact that XLI was designed to provide the “best” forecast of economic activity at a six month horizon, using virtually all of the macroeco nomic data available. 5. M ix in g m o d els fo r real G D P This section analyzes those indicators drawn from the previous sections that contain independent information and did well in the out-of-sample Kalman rankings. The indicators are subjected to another round of encompassing tests and rankings. Finally, the usefulness of FEDERAL RESERVE BANK OF CHICAGO Change in sensitive materials prices (SMPS) these final indicators is assessed in the context of a time varying forecast mixing model. Table 5.1 presents the Kalman forecast RMSEs for the one, two, and four quarter horizon forecasts of real GDP. For the one quarter hori zon the best indicators are the NBER composite indicators (XLI and XRI2), and the Department of Commerce Composite Index of Leading Indi cators (LEAD). The spreads and real M2 (M2R) do the worst at this short horizon, but all of the remaining indicators do contribute information beyond the own past history of GDP (NONE). At the two quarter horizon, the best indicator is the NBER Experimental Leading Index (XLI) with the 12 month Treasury bill/federal funds spread (TB12FF) coming in a distant second: XLI is 14 percent more accurate than TB12FF. This is not surprising since XLI was constructed by Stock and Watson to produce the “best” fore- 25 as the best monetary aggregate considered here. Finally, notice Kalman residuals for surviving indicators that the 6 month commercial paper/6 month Treasury bill Real GDP 1 quarter 2 quarters 4 quarters spread (CP6TB6) did not make Indicator RMSE Rank RMSE Rank RMSE Rank the final list at the two quarter forecast horizon, but it is a com 4 2.754 EUR03 3.622 3 n.a. n.a. ponent of XLI. FF n.a. n.a. n.a. 2.160 2 n.a. At the four quarter horizon, 3.674 4 M2R 6 2.844 5 2.219 three indicators are undominated: 4 CP6TB6 3.656 5 2.760 n.a. n.a. FF, M2R, and TB12FF. The NBER Experimental Leading TB12FF 7 2.002 1 3.753 2.751 2 Index (XLI) does not contain CM10FF n.a. n.a. 2.161 3 n.a. n.a. independent in formation beyond XLI 1 1 2.392 3.246 2.376 5 these indicators. CM 1OFF is XRI2 3.427 3 n.a. n.a. n.a. n.a. included in the final list for three LEAD 3.307 2 n.a. n.a. n.a. n.a. reasons: it is undominated at the 15 percent significance level, it 4.052 2.799 NONE 8 3.369 6 6 covers the NBER Experimental NOTES: n.a.: The indicator is not an initial survivor at this forecast horizon. Leading Index better than the Sample period is July 1973 - December 1991, quarterly data. shorter end of the term structure (TB12FF), and it is interesting to include a long term spread at this horizon since cast of the growth in economic activity over the Stock and Watson found a long term spread six month horizon considered here. Turning to useful at the two quarter horizon. the four quarter horizon, it seems surprising that The next step is to combine these forecasts XLI comes in last after TB12FF, the federal into a forecasting model (for each horizon) funds rate (FF), the 10 year Treasury bond/ which allows the weights on the indicators to federal funds spread (CM10FF), and M2R. vary over time depending upon their recent This demonstrates again that the choice of performance. Essentially we would like the economic indicators depends critically upon the model to take the following form: horizon being forecasted: at the four quarter growth horizon, a different collection of interest (3) F, = K f or(Ah + <t\ J°r(B )x + §Mfor (C)t ; rate spreads than the ones selected by Stock and Watson is useful. where for(A) represents a forecast based upon New encompassing results are displayed in indicator A and Ft is the combined forecast. Table 5.2. At this point, the purpose of these The weights <t>.f should be nonnegative and sum tests is to narrow the list of indicators in a struc to one: in this case, the indicator’s weight is a tured manner. However, a rigid adherence to a direct measure of its importance for the fore statistical significance level is not maintained if cast. When the weights vary over time accord an indicator is relatively useful and of indepen ing to their forecast accuracy, the time path of dent interest. At the one quarter horizon, XLI, the weights provide a direct measure of the XRI2, and LEAD are each undominated and indicators’ reliability over time. We implement together sufficient. The two quarter horizon is this model in the following way. Let eit2be the more interesting. Three indicators are clearly sum of (recent) squared forecast errors based necessary. XLI is undominated, and TB12FF is upon indicator /’s model. In this paper, we take undominated at the 10 percent level. The 3 “recent” to be one year of known forecast errors month eurodollar rate (EUR03) is not covered (4 quarters). Let avgr(e.-2) be the average of the by these two indicators, and it is not dominated e.(2s at time t and p. be the average of e.(2at the 11 percent significance level. M2R is avgt( e 2) over time. Then (ft is defined to be: also included in this final cut for two reasons: it is only covered by XLI at the 14 percent (4) <J>,= a. - (3, (e,2- avg/e.'2) - \i.), a , P,> 0 ; significance level and it is of inherent interest 26 TABLE 5.1 ECONOMIC PERSPECTIVES TABLE 5.2 M i x e d m u l t i p e r i o d e n c o m p a s s in g te s ts (P r o b a b ilit y v a lu e f o r n u ll h y p o th e s is : X is e n c o m p a s s e d b y Y ) Real GDP (1 quarter) Y EUR03 FF M2R CP6TB6 TB12FF CM10FF XLI XRI2 LEAD Maximum P-Value X n.a. 0.100 — — — — 0.107 — — 0.107 0.958 n.a. — 0.067 — 0.144 — — 0.958 — — n.a. 0.055 0.168 n.a. — — — — — — 0.168 CP6TB6 — — — — 0.288 — — 0.288 TB12FF 0.186 0.193 — — n.a. — 0.453 — — 0.453 CM10FF 0.168 0.098 — — 0.260 n.a. 0.809 — — 0.809 — — — — — — n.a. — — 0.001 EUR03 FF M2R XLI XRI2 — — — — — — — n.a. — 0.012 LEAD — - — — — — — — n.a. 0.030 — 0.161 — — n.a. 0.228 — — — — — — — n.a. — — — — — — — — — n.a. — — — — — — — — — — n.a. 0.110 — — — 0.139 0.304 0.062 0.514 n.a. 0.370 0.761 Real GDP (2 quarters) n.a. 0.110 — — FF 0.868 n.a. — — M2R — — 0.064 0.076 — — — — 0.082 — — — — — n.a. — — — — 0.066 0.230 — n.a. — — — — 0.088 XRI2 n.a. — — 0.270 — — — 0.791 0.609 n.a. — 0.327 — — — 0.817 LEAD 0.102 0.122 EUR03 CP6TB6 TB12FF CM10FF XLI XRI2 LEAD 0.868 0.139 0.304 0.082 0.514 0.000 0.370 0.761 Real GDP (4 quarters) EUR03 FF M2R CP6TB6 TB12FF CM 10FF XLI j — — n.a. 0.420 — — 0.105 0.959 0.960 — — — n.a. — — — 0.364 — — — 0.850 n.a. 0.147 0.157 0.839 — — — 0.779 — n.a. 0.298 0.711 — — — 0.401 — — n.a. 0.939 — — — — — — — n.a. — — — — — — — 0.690 0.609 0.023 0.007 0.850 0.011 0.147 0.298 0.959 — 0.240 0.300 0.420 — n.a. 0.960 NOTES: Values less than or equal to 0.05 are marked with a dash. Sample period is January 1963 - December 1991, quarterly data. where the parameters a and (3 can be estimated by a linear regression model if the nonnegativi ty constraints are ignored, or nonlinear meth ods if the constraints are imposed. Since e.f2avg((e.(2) - (j.. is mean zero by construction, the time variation due to the (3s nets out to zero over time. Consequently, the a estimates repre sent the average weight associated with each indicator forecast. However, over short periods of time when an indicator’s forecast misbe haves, its errors e.2 will be larger than the aver age errors; this will lead to the indicator’s forecast receiving a temporarily smaller weight. Table 5.3 displays the estimated a weights for these models. The one quarter results indi FEDERAL RESERVE BANK OF CHICAGO cate that XLI is the most reliable, having an average weight of .533 in the combined fore cast. The other indices (XRI2 and LEAD) received about equal shares of the remaining weight. The (3s in this case are estimated to be zero; that is, there is no significant contribution to the forecast accuracy by allowing the weights to vary over time. The two quarter results are more interest ing. As was expected from the encompassing results, XLI receives the bulk of the weight in the final forecast (62 percent). This agrees with the analysis of Stock and Watson who con structed the NBER Experimental Leading In- 27 TABLE 5.3 Relative weights in mixing regressions Indicator _____________ Real GDP_____________ 1 quarter 2 quarters 4 quarters * 0.093 (0.260) n.a. n.a. n.a. 0.105 (0.209) M2R * 0.414 (0.178) CP6TB6 * 0.187 (0.227) * TB12FF * 0.103 (0.238) 0.368 (0.259) CM10FF n.a. n.a. XLI 0.533 (0.174) 0.617 (0.197) 0.114 (0.212) * XRI2 0.214 (0.155) n.a. n.a. LEAD 0.253 (0.206) n.a. n.a. EUR03 FF n.a. NOTES: Numbers in parenthesis are standard errors, n.a.: The indicator is not an initial survivor at this forecast horizon. (*): The indicator is encompassed by other indicators at this horizon. dex explicitly for its ability to forecast at this two quarter horizon. We do find that M2R receives a substantial weight (19 percent), while the TB12FF spread is at 10 percent and EUR03 is at 9 percent. Figure 5.1 graphs the time path of the (|) weights for these four indicators, as well as the two quarter GDP forecast and actual. Notice first that the NBER Experimental Lead ing Index forecasts have been quite reliable, only once dropping below a 50 percent weight in the combined forecast. M2R, however, has varied dramatically in its usefulness, going negative on two occasions: in 1976 and imme diately following the 1981-82 recession. During that recession, M2R did not forecast negative growth at any time (although it did in the 1980 recession), whereas EUR03, TB12FF, and XLI did forecast negative growth during some por tion of this recession.13 This poor performance is captured in the time varying model by de creasing the weight on the M2R forecast tempo rarily until it begins to improve. On the other hand, during the most recent recession M2R has gone above a 50 percent weight (keep in mind that the average weight for M2R is .19). During this time, M2R has grown only slowly and this 28 led to a forecast of slow growth during 1991 (see Figure 5.1). At this same time, EUR03, TB12FF, and XLI signalled substantially higher growth than was realized. Each of these indica tors is currently receiving less than its average weight. Consequently, the time varying mixing model finds that M2R has been an unusually useful indicator during the recent recession, despite its generally erratic performance at this horizon versus its relative failure at the twelve month horizon. By contrast the four quarter horizon results in Figure 5.2 appear to be a picture of stability. M2R and TB12FF receive the largest uncondi tional weights, 41 percent and 37 percent re spectively. FF and CM 1OFF receive consider ably less (around 10 percent each). The graphs of the time varying weights indicate that, at this horizon, M2R and TB 12FF have been reason ably reliable indicators, always staying near their unconditional weight. On the other hand, CM 1OFF has been extremely unreliable, going to zero or negative in 1987-88 and during the recent recession. The contrast between the dominance of XLI at the two quarter forecast horizon and its submissiveness at the four quarter horizon dem onstrates strongly the need for a different set of indicators for each forecast horizon. The useful ness of TB 12FF and M2R for forecasting real GDP at the one year horizon indicates that a different index would be constructed if this forecast horizon was the relevant objective. A note on standard errors is in order. Examination of Table 5.3 indicates that the standard errors associated with the parameters of these mixing models are fairly large. This is not surprising in light of the high degree of collinearity that would be expected of a set of reasonably suc cessful forecasts. In fact, it is typically the case that only the strongest indicator at a given hori zon is statistically significant. All this is saying is that the relative weights among successful indicators are subject to substantial uncertainty and that the marginal information after the first one or two indicators quickly drops toward 0. Nevertheless, the point estimates and time paths of these relative weights provide a useful bench mark, even though the precision with which they are estimated would not change strongly held prior beliefs. ECONOMIC PERSPECTIVES FIGURE 5.1 Mixing results 2 quarter ahead forecast vs. actual 3 month eurodollar (EUR03) Real M2 (M2R) annualized growth rates annualized growth rates 9.0 r 6.0 3.0 0.0 -3.0 II - 6.0 I I I I l I ........... NBER Experimental Leading Index (XLI) 9.0 r 3.0 0.0 -3.0 - 6.0 1973 76 Forecast reliability weight 3 month eurodollar (EUR03) Real M2 (M2R) weight weight NBER Experimental Leading Index (XLI) C onclusion Four things became clear as the preceding analysis developed. First, the forecast horizon is an essential aspect of choosing and evaluat ing indicators. Second, substantial information FEDERAL RESERVE BANK OF CHICAGO resides in the term and private/public spreads and both of these seemingly very different types of spreads seem to include common as well as independent information. Third, while composite indicators may be extremely useful, 29 they are only as good as their design allows. The NBER Experimental Leading Index does very well at precisely what it was designed for, that is, forecasting economic activity at a six month horizon. Its usefulness beyond this hori 30 zon is far more limited than prior analysis would have suggested. Fourth, the analysis also suggests that the type of general purpose target variable that the old monetary targeting litera ture sought probably does not exist, at least in ECONOMIC PERSPECTIVES terms of real economic activity. Policymakers will continue to need to mix information ac cording to their current focus. Mixing models of the sort used in this article are meant to be preliminary work in this regard. FOOTNOTES 'The NBER Experimental Leading Index (XLI) developed by James Stock and Mark Watson is a clear exception, since it was created as a single “best” indicator of economic activity [see Stock and Watson, (1989b)]. 2The following examples illustrate the notation we will use in the Methodology section to indicate different classes of tables: Table 1 refers to the first table in each family of indicators, Table _.2 refers to the second table in each family, and so forth. 3It should be noted that these are not iterated VAR fore casts, rather, the forecast parameters are chosen to maxi mize performance at the forecast horizon specified. This can be thought of either as a state space estimation mini mizing the t+ k forecast variance or as a simple OLS regres sion with the t+ k growth rate as the dependent variable. This avoids any problem that might result from an indicator that performs poorly at high frequencies interfering with longer frequency forecasting. 4The standard deviation measure used is the one from a bivariate VAR for the indicator and the measure of eco nomic activity. This is used to approximate the average size of the movement in the indicator series. 5This is basically the same as an impulse response function except that the identifying assumption is not derived from a specific decomposition of the error matrix, but from the assumed path of the actual series, that is, the indicator changes given the level of current activity. This is arithmet ically equivalent to an impulse response function using a Choleski decomposition with the indicator ordered last. 6The monetary base is the sum of reserve balances at the Federal Reserve Banks and currency in circulation. 7L is the broadest monetary aggregate, consisting of M3 plus the nonbank public holdings of U.S. savings bonds, short term Treasury securities, commercial paper, and bankers’ acceptances, net of money market mutual fund holdings of these assets. "These are the only commonly used spreads available for the entire sample period. 9We used the 10 year Treasury constant maturity bond rate because the 7 year bond rate, which might be preferred, is not available for the entire sample period. l0The NBER Nonfinancial Experimental Recession Index, which estimates the probability that the economy will be in a recession six months later, is based on a set of nonfinan cial leading indicators. (See NBER Press Release, January 30, 1991.) "SMPS is calculated as the quarterly average of the month ly changes in sensitive materials prices, smoothed. The sources for the monthly data are: U.S. Department of Commerce, U.S. Department of Labor, and the Commodity Research Bureau, Inc. 12The dynamic response graph for KSWMIX is not shown because data on the mix are available only on a quarterly basis, while employment data are monthly. l3It is useful to remember that the primary components of the NBER Experimental Leading Index are the 6 month commercial paper/6 month Treasury bill spread and the 10 year Treasury bond/1 year Treasury bond spread. So it should not be surprising that the NBER Experimental Leading Index misbehaved during this period when the 3 month eurodollar rate and the 12 month Treasury bill/ federal funds spread also misbehaved. REFERENCES Bernanke, Ben S., “On the predictive power of interest rates and interest rate spreads,” New England Economic Review, November-December, 1990, pp. 51-68. Chong, Y. and D. Hendry, “Econometric evaluation of linear macro-economic models,” Review of Economic Studies, 53, 1986, pp. 671-690. Estrella, A. and G. Hardouvelis, “The term structure as a predictor of real economic activi ty,” Journal of Finance, 46, 1991, pp. 555-576. FEDERAL RESERVE BANK OF CHICAGO Friedman, B. and K. Kuttner, “Why does the paper-bill spread predict real economic activi ty?” forthcoming in James H. Stock and Mark W. Watson eds., New Research in Business Cycles, Indicators and Forecasting, University of Chicago Press and the NBER, 1992. Kashyap, A,, J. Stein and D. Wilcox, “Mone tary policy and credit conditions: evidence from the composition of external finance,” Federal Reserve Board, Working Paper No. 154, 1991. 31 Laurent, Robert D., “An interest rate-based indicator of monetary policy,” Economic Per spectives, Federal Reserve Bank of Chicago, January/February, 1988, pp. 3-14. National Bureau of Economic Research, Press release, January 30, 1991. Sims, Christopher A., “Interpreting the macroeconomic time series facts: the effects of mon etary policy,” manuscript, 1991. Stock, J. and M. Watson, “Interpreting the evidence on money-income causality,” Journal of Econometrics, Vol. 40, 1989a, pp. 161-182. ___________ , “New indexes of coincident and leading economic indicators,” in NBER Macroeconomics Annual, edited by O. Blan chard and S. Fischer, The MIT Press, 1989b, pp. 351-409 Strongin, Steven, “Macroeconomic models and the term structure of interest rates,” Federal Reserve Bank of Chicago, Working Paper No. 90-14, 1990. ____________, “The identification of monetary policy disturbances: explaining the liquidity puzzle,” Federal Reserve Bank of Chicago, Working Paper No. 91-24, 1991. Shaping the Great Lakes Economy Conference on the Region’s Economy and Development Strategies Indianapolis, Indiana October 15, 1992 In conjunction with Indiana University’s Institute for Development Strategies and the Great Lakes Commission, the Federal Reserve Bank of Chicago will hold a conference at the University Place Conference Center and Hotel in Indianapolis. The 1992 conference will focus on the state of the region’s economy and on its strategies to promote economic growth and development. Topics featured will include: ■ the state of the region’s economy and its directions in the 1990s ■ state and regional development policies and the Federal policy environment ■ the profound changes now under way in the manufacturing sector’s organization, technology, and labor force 32 If you are interested in receiving further information and registration materials, please contact: Great Lakes Commission The Argus II Building 400 Fourth St. Ann Arbor, Michigan 48103-4816 Phone: (313) 665 9135 FAX: (313)665 4370 ECONOMIC PERSPECTIVES ECONOMIC PERSPECTIVES BULK RATE Public Information Center Federal Reserve Bank o f Chicago P.O. Box 834 Chicago, Illinois 60690-0834 U.S. POSTAGE PAID CHICAGO, ILLINOIS PERMIT NO. 1942 Do N o t F orw ard A d d ress C o rrec tio n R equested R eturn P o stag e G uaranteed FEDERAL RESERVE BANK OF CHICAGO