Forecasting Recessions: Can We Do Better on MARS(TM)?

Macroeconomists spend much of their time developing theories and building models to demonstrate how shocks propagate and affect the overall level of economic activity. Both policymakers and the private sector maintain a keen interest in understanding the state of business affairs and the most likely path the economy will take over a planning horizon. Although there are a number of economic events that concern the authorities—including excessive inflation and unemployment—considerable attention is paid to the forecasting of recession. If policymakers can anticipate a recession, they take preemptive corrective action. The private sector uses this information to shelter itself from the vagaries of the business cycle and the most likely reaction of policymakers. Recently a number of studies have examined the ability of financial variables to forecast recessions. Many analysts find that financial indicators contain information that can be used to increase forecast accuracy. Estrella and Mishkin (1998) found that the slope of the yield curve helped predict recessions beyond one quarter. Haubrich and Dombrosky (1996), Bernard and Gerlach (1996), Dueker (1997), and Atta-Mensah and Tkacz (1998) reported similar results.1 Many of these studies employed probit models to estimate the probability of recession. Probit models are sometimes used when economists model the behavior of a dependent variable which takes on two values, e.g., recession = 1, no recession = 0. The traditional approach to probit modeling requires the researcher to choose the variables that will be included in the equation, determine their level of interaction, and assume each variable plays the same role across all recessions in the sample period. These assumptions imply that the causal nature of recessions remains fixed over time, which we know to be at odds with the stylized facts of American business cycles in the twentieth century.2 Consequently, probit models may not adequately capture the underlying processes related to recession. The purpose of this paper is to revisit the information contained in financial variables using nonlinear, nonparametric methods, in particular, multivariate adaptive regression splines (MARS).3 As with the probit specification, MARS models provide probability forecasts that lie between zero and one, yet they admit a much wider range of possible relationships in the data. The MARS approach allows the series to enter both individually and in combination. Given the idiosyncrasies of the American business cycle, this nonlinear, nonparametric approach may provide greater insight into the factors contributing to recession while avoiding some of the pitfalls associated with the probit specification.