Return Predictability or Risk Timing? Bootstrap Evidence from a Century of US Equity Data

Abstract

The term spread is one of the most widely cited predictors of US equity returns. Using monthly data from January 1925 to December 2025, we evaluate its out-of-sample forecasting ability via a rolling-window predictive regression against a historical mean benchmark. The term spread produces a negative out-of-sample R 2 of-0.52%, indicating that incorporating the predictor worsens forecast accuracy. The Clark-West test statistic of 2.085 suggests significance under asymptotic inference (p = 0.019), but a moving block bootstrap with 5,000 replications overturns this conclusion (p = 0.557), with the empirical 5% critical value (5.21) far exceeding the asymptotic threshold (1.645). Despite the absence of return predictability, a volatility-targeted portfolio conditioning on the same term spread signal delivers a 29% improvement in Sharpe ratio (0.45 vs 0.35) and a positive certainty equivalent gain of 9.0 basis points per annum. The resolution is that the term spread contains information about risk regimes rather than expected returns: it operates as a risk-timing signal that reduces exposure during adverse macroeconomic states, preserving capital for compounding during recoveries. The break-even transaction cost of 24.2 basis points per unit turnover survives realistic institutional costs but offers limited buffer. These findings have implications for the return predictability literature: many published results that rely on asymptotic inference may not survive finite-sample correction.