Authors - Meshak Ratshikombo, Mehrdad Ghaziasgar Abstract - Deep Reinforcement Learning has emerged as a promising approach for algorithmic trading, but trading performance remains highly sensitive to reward design. This study investigates how reward engineering shapes learned trading behavior by training agents exclusively on FTSE market data under identical conditions while varying only the reward formulation. Evaluation across multiple unseen equity indices demonstrates that reward functions induce distinct behavioral biases governing exposure timing, volatility sensitivity, and downside risk. Profit-oriented rewards encourage aggressive trading and higher variability, whereas risk-aware and sparse rewards produce more stable policies with improved drawdown control. The findings show that reward engineering acts as a strong inductive bias influencing both trading behavior and cross-market generalization in reinforcement-learning-based trading systems.
Open Zoom