Do Macroeconomic Predictors Matter More in Recessions? Evidence from Penalized Regression and Out-of-Sample Tests

Sammanfattning

This thesis examines whether macroeconomic predictors of the equity premium exhibit stronger forecasting power during economic recessions. Building on the Elastic Net framework of Zou and Hastie (2005), we introduce state-dependent interaction terms between macroeconomic variables and a recession indicator, capturing regime-specific predictive relationships across the business cycle. Using 126 monthly and 245 quarterly predictors from FRED-MD and FRED-QD, we forecast CRSP equal- and value-weighted excess returns over an expanding window from January 1980 to November 2024, under two recession definitions: the ex-post NBER indicator and the real-time Sahm rule. Under the NBER indicator, the interaction model reduces forecast error overall and during recession months for equal-weighted returns, supported by the Clark-West test, the Diebold-Mariano statistic confirms the result is not mainly driven by the nested-model bias correction. Value-weighted returns show no improvement under any specification. Sahm-rule gains are significant but concentrated in expansion rather than recession months, providing weaker support for the state-specific predictability hypothesis. A mean-variance portfolio exercise shows the interaction model improves Sharpe ratios relative to the baseline across all reported specifications. At high risk aversion, certainty-equivalent returns exceed the buy-and-hold benchmark for equal-weighted strategies. The results suggest that recession-dependent macroeconomic predictability is real but narrow, concentrated in small-cap equities and sensitive to the choice of recession indicator.