Reinforcement learning meets technical analysis: combining moving average rules for optimal alpha

A common issue faced by investors who use technical analysis is the reconciliation of conflicting trading signals, especially when these signals are highly correlated, such as those generated by multiple moving averages. This study expands on a model-free algorithm inspired by reinforcement learning to address the challenge of reconciling trading signals while taking transaction costs into account. The algorithm is trained to optimize alpha, a widely used measure of risk-adjusted return. Principal component analysis is utilized to reduce the dimensionality of a modified version of moving average signals, which are then used to define the input state. A policy network, represented by a feedforward neural network, is trained using historical data to convert states into trading actions. An evaluation network calculates and optimizes alpha by adjusting the policy network’s parameters. The algorithm utilizes a zero-arbitrage portfolio to accurately isolate alpha from the underlying asset’s return. By combining 199 simple moving average signals in a systematic manner, the algorithm was able to maximize the q5 asset pricing model alpha using a high-volatility United States stock portfolio as a risky asset. The algorithm demonstrates superior performance compared to both individual moving average signals and existing combination algorithms.This study significantly advances market‑timing strategies by introducing a reinforcement‑learning‑inspired algorithm that reconciles highly correlated yet conflicting trading signals—such as simple moving averages—to maximize alpha while explicitly accounting for transaction costs. At its core is a policy network, implemented as a feedforward neural network, trained simultaneously across all in‑sample time slices via a complementary evaluation network. The original trading signals are first transformed to achieve mean‑stationarity, then standardized and compressed with principal component analysis to reduce dimensionality and information redundancy. Algorithmic returns are rigorously isolated from underlying asset returns through a zero‑arbitrage portfolio framework. Because the method is model‑free, it avoids reliance on predefined return models or specific functional forms for the trading signals. The algorithm can also be readily extended to other asset‑pricing models and to similar technical—or non‑technical—signals. This comprehensive, adaptive, and principled approach equips investors with a robust tool for enhanced decision‑making in dynamic financial environments.