A Neural Network Architecture for Maximizing Alpha in a Market Timing Investment Strategy

In finance, assuming more risk often corresponds to the expectation of higher, compensating returns. In this setting, alpha stands out as one of the most prevalent and refined measures of risk-adjusted return ever postulated, allowing for the estimation of the excess return that cannot be explained by the risk factors impacting an asset. This article introduces a neural network architecture designed to formulate an investment strategy with the explicit goal of maximizing alpha. The strategy, centered around market timing, determines on a daily basis, based on past returns of the risky asset, whether to fully invest in the risky asset or opt for the risk-free alternative. The neural network architecture comprises two components: a policy network for strategy implementation and an evaluation network for long-term alpha computation during parameter optimization. Employing value-weighted US size decile portfolios as risky assets, the study achieves significant out-of-sample alphas ranging from 3.6% to 8.2% per year under the $q^{5}$ asset pricing model (with a transaction cost assumption of one basis point). By construction, these alphas are not generated by risky asset growth. Robustness tests yield similar results with equal-weighted decile portfolios or under the Fama and French six-factor asset pricing model. Variations in transaction cost, number of past returns used as inputs, policy network design, or training sample size produce similar outcomes. This study underscores the effectiveness of reinforcement learning-inspired techniques in uncovering alpha in financial markets.