Construction of Quantitative Trading Framework in Taiwan Stock Market Integrating Multi Factor Model with Deep Reinforcement Learning

Abstract

Heightened global financial uncertainty has exposed two persistent weaknesses in traditional factor investing: temporal decay of factor returns and the limited adaptability of fixed-threshold selection rules to shifting market regimes. To address these limitations, this study develops a quantitative trading framework that couples a multi-factor model with reinforcement learning, applied to 863 companies listed on the Taiwan Stock Exchange and spanning seven primary categories—market, valuation, quality, growth, macro, momentum, and sentiment—comprising twenty-four secondary factors. Information-coefficient analysis identifies sales growth as the most stable predictor, with a positive predictive ratio of 70.34%. Among five reinforcement-learning algorithms benchmarked under identical trading constraints, the Twin Delayed Deep Deterministic Policy Gradient (TD3) attains the strongest risk-adjusted performance, with an annualised return of 67.51% and a Sharpe ratio of 1.92. Over a six-month out-of-sample window, the strategy generates a 28.05% cumulative return and exceeds the Taiwan Capitalisation-Weighted Stock Index (TAIEX) by 23.15 percentage points while remaining nearly market-neutral. The empirical results indicate that the framework not only captures the dynamic structure of the Taiwan equity market and supports adaptive stock selection and risk management, but also delivers a return profile that is competitive with established benchmarks—offering a forward-looking methodology for quantitative investing.