Forecasting U.S. Recessions with Probit Stepwise Regression Models
The Result Is Better Forecasting Performance
By John Silvia, Sam Bullard, and Huiwen Lai
John Silvia has been Chief Economist of the Wachovia Corporation since 2002. Previously, he was with the U.S. Senate Joint Economic Committee, the U.S. Senate Banking, Housing and Urban Affairs Committee, Kemper Funds, and Scudder Kemper Investments, Inc. Silvia has been a director of NABE and serves as a member of the Blue Chip Panel of Economics Forecasters and on an informal advisory group of the Federal Reserve Bank of Philadelphia. He serves on the Economic Development Board for the State of North Carolina and has served on a number of other advisory committees. He holds a B.A. and Ph.D. in economics from Northwestern University and a Master’s degree from Brown University.
Sam Bullard is an economist at the Wachovia Corporation. His coverage responsibilities include U.S. macroeconomic and financial market developments as well as the financial services industry. He is also responsible for the Economic Group’s strategic development and daily operations. He received his B.B.A in finance from the University of Georgia and his M.B.A. from Wake Forest University.
Huiwen Lai is an econometrician for the Wachovia Corporation and a visiting associate professor of economics at the University of International Business and Economics in Beijing. Before joining Wachovia, He was with Bank of America, SAS Institute Inc., and Guangxi University. He has published in several economics journals. He received his Ph.D. in economics from the University of Toronto, a Master’s degree in quantitative economics from Northeastern University of Finance and Economics in China, and a Bachelor’s degree in mathematics from Sichuan University.
Yield spreads have been repeatedly used in the literature as the top candidates in predicting future recessions. In this paper, we show that existing model specifications are good but fall short of the performance of more complete models. Applying a probit stepwise regression procedure to a large number of economic indicators, we find models that dramatically outperform those used in the literature. Due to a time series that only began in 1964Q1 and very few historical recessions, any model specification may capture only a few of the economy's many aspects and thus can potentially be biased. Nevertheless, models with better statistical properties should have a better chance to capture the occurrence of recession. Our chosen models are not immune to statistical limitations but should forecast better than the existing models in the literature.
