Forecast-Agnostic Portfolios

Abstract

We introduce forecast-agnostic (FA) portfolios that exhibit out-of-sample market-timing ability without relying on estimated predictive coefficients. These portfolios go long or short the market based on the level of a predictor variable, thereby avoiding the instability and estimation error that undermine traditional market-timing strategies. Despite using predictor variables that typically deliver negative out-of-sample R2 values (Goyal et al., 2024), FA portfolios deliver significantly positive alphas on average. We explain these seemingly contradictory phenomena by interpreting regression coefficients as portfolio returns: genuine predictability is necessary for high portfolio returns, whereas achieving a positive out-of-sample R2 additionally requires the ability to forecast the returns on the forecast-agnostic portfolios themselves. As these FA portfolio returns could not be too predictable, estimating them substantially penalizes the out-of-sample R2 by the inverse of the estimation sample size. Simulations show that the statistic we propose has power to detect predictability that extends beyond in-sample diagnostics and the out-of-sample R2.

Keywords: Market timing, Efficient Market Hypothesis

JEL Classification: G12, G14, G40