What Predicts U.S. Recessions?

1 This paper presents preliminary findings and is being distributed to economists and other interested readers solely to stimulate discussion and elicit comments. The views expressed in this paper are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank of New York or the Federal Reserve System. Any errors or omissions are the responsibility of the authors. Federal Reserve Bank of New York Staff Reports What Predicts Recessions? Weiling Liu Emanuel Moench Staff Report No. 691 September 2014 What Predicts Recessions? Weiling Liu and Emanuel Moench Federal Reserve Bank of New York Staff Reports, no. 691 September 2014 JEL classification: C52, C53, E32, E37 Abstract We reassess the predictability of recessions at horizons from three months to two years ahead for a large number of previously proposed leading-indicator variables.

2 We employ an efficient probit estimator for partially missing data and assess relative model performance based on the receiver operating characteristic (ROC) curve. While the Treasury term spread has the highest predictive power at horizons four to six quarters ahead, adding lagged observations of the term spread significantly improves the predictability of recessions at shorter horizons. Moreover, balances in broker-dealer margin accounts significantly improve the precision of recession predictions, especially at horizons further out than one year. Key words: recession predictability, ROC, term spread, leading indicators, efficient probit estimator _____ Liu: Harvard Business School (e-mail: Moench: Federal Reserve Bank of New York (e-mail: The views expressed in this paper are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank of New York or the Federal Reserve System.))

3 1 IntroductionAccurately predicting business cycle turning points, and in particular impending economicrecessions, is of great importance to households, businesses, investors and policy makers research has documented that a variety of economic and financial variables contain pre-dictive information about future recessions in the United States. Most prominently, Estrellaand Hardouvelis (1991) and Estrella and Mishkin (1998) have documented that the slope ofthe term structure of Treasury yields has strong predictive power for US output growth andUS recessions at horizons up to eight quarters into the future. Other variables that havebeen considered as leading recession indicators include stock prices (Estrella and Mishkin(1998)), the index of Leading Economic Indicators (Stock and Watson (1989), Berge andJord`a (2011)), credit market activity (Levanon, Manini, Ozyildirim, Schaitkin, and Tanchua(2011)), as well as various employment and interest rate measures (Ng (2014)).

4 In this paper, we reassess the predictability of US recessions since 1959 using a wide varietyof leading indicator variables that have been considered in the academic and practitionerliterature. Consistent with most of the prior literature, we use the business cycle datingchronology provided by the National Bureau of Economic Research (NBER) as the bench-mark series of business cycle turning points. While the NBER recession indicator is a binaryvariable, most leading indicators have continuous distributions. Thus, much of the empiricalliterature has used the nonlinearprobitmodel to map changes in predictor variables intorecession forecasts, and we follow this probability of a recession implied by the probit model is rarely exactly zero or one.

5 Thus,a cutoff is usually adopted such that a predicted probability above the cutoff is classifiedas a recession. In order to objectively evaluate the model s ability to categorize future timeperiods into recessions versus expansions over an entire spectrum of different cutoffs, oneneeds to complement the probit model with a classification scheme. A classification schemethat has long been used in the statistics literature but has only recently found its way intoeconomic research is the receiver operating characteristic (ROC) curve (see, for example,1 Khandani, Kim, and Lo (2010), Jord`a and Taylor (2011), Jord`a and Taylor (2012)). TheROC curve is computed in several steps. First, for a given grid of cutoff values of the im-plied recession probability, one calculates the percentage of true positives and false positivesfor classifying all periods in the sample.

6 One then plots the percentage of true and falsepositives against one another for the entire grid to create the receiver operating curve. Onemethod of comparing the predictive ability of classifiers across a spectrum of cutoff values isto integrate the area under the ROC curve, creating the AUROC. A model which delivers aperfect classification of all time periods into recession and expansion would only have truepositives and no false positives and an AUROC equal to one. In contrast, a model which isthe equivalent of a random guess would have on average an equal number of true and falsepositives, which corresponds to an AUROC equal to Hanley and McNeil (1983) deriveat-test for the hypothesis that the predictive ability of two different classifiers are equalby using their AUROC s.

7 We use their test in order to discriminate between the predictiveability of different recession indicators considered in the main findings can be summarized as follows. The Treasury term spread Predicts bestat horizons of one year and more. That said, some indicators add to the predictive ability ofthe term spread at these horizons. In particular, margin debit at NYSE brokers and deal-ers, a measure of leverage in the financial sector, significantly improves the in-sample andout-of-sample predictive power of the probit model when considered jointly with the termspread at these longer horizons. This highlights the importance of financial intermediarybalance sheet conditions in the transmission of economic shocks (see, for example, Adrianand Shin (2010) and Adrian, Moench, and Shin (2010)).

8 While the importance of financialintermediary leverage for the pricing of risk has been empirically documented by Adrian,Etula, and Muir (2012) and Adrian, Moench, and Shin (2013), to the best of our knowledge,its usefulness for the predictability of recessions has not previously been horizons shorter than one year ahead, we find that adding six-month lagged observationsof the Treasury term spread significantly improves the predictive power of the probit model2to predict recessions. This suggests that at these shorter horizons there is predictive infor-mation not only in the contemporaneous steepness of the Treasury yield curve, but also inthe lagged term structure slope. The negative sign on the coefficient of lagged spread hastwo implications: persistence and change.

9 First, if spreads were negative six-months ago,then there is a higher probability of recession in the future. Second, given the same startingvalue of spread six-months ago, a sharper drop in the spread since then leads to a higherprobability of recession in the future. In addition to the contemporaneous and lagged Trea-sury term spreads, a number of other variables also contain predictive information aboutfuture recessions at horizons less than one year ahead. In particular, the annual return onthe S&P500 stock market index, the Michigan survey of consumer expectations, and againthe margin debit at NYSE brokers and dealers significantly increase the predictive power ofthe probit model when added to the Treasury term paper is related to a large literature on predicting real output growth and recessionsusing financial and macroeconomic leading indicators.

10 Estrella and Hardouvelis (1991) firstpopularized the Treasury term spread as a predictor of future output growth and found that it has greater predictive power than the Leading Indicator Index and outper-forms survey forecasts both in- and out-of-sample. Estrella and Mishkin (1996) and Estrellaand Mishkin (1998) considered the out-of-sample performance of a range of macroeconomicand financial variables both one-at-a-time and in combination. Their findings suggest thatin the short run, stock returns are a valuable leading indicator. However, at horizons of oneyear ahead or more, the Treasury term spread is still the single best performing (1997) revisited the term spread as a leading indicator within the context of theprobit model studied in our paper.