Micro Alphas

Abstract

We introduce the concept of micro alphas, weak signals that may not reach statistical significance individually, but when combined, generate predictability for the equity risk premium. Unlike previous predictability studies born out of the Goyal and Welch (2008) critique, predictors do not have to exhibit statistical significance consistently over the sample for them to be considered. Though the recent update from Goyal, Welch, and Zafirov (2024) is more nuanced than the original, it still evaluates candidate variables against an only partially relevant criteria. Borrowing from the machine learning literature, we further apply transformations and feature selection in a cross-validated walk-forward way, in order to capture potentially non-linear and timedependent effects. We present a strategy based on an elastic net model. This strategy has been implemented in a public fund and has given consistent excess returns over the S&P500. We also suggest possible improvements to the model.

Keywords: Equity risk premium, market timing, tactical asset allocation, predictability, weak signals, machine learning