Reinforcement Learning-Driven Optimal Capital Allocation for Intelligent Finance (PLoS ONE)

Summary

• The proposed DAC-CPPO model achieves an overall accuracy of 97.6%, an MAE of 0.074, and an RMSE of 0.081 for portfolio optimization.p.1
• The study evaluates two primary datasets: the Intelligent Finance Assets dataset containing 7,000 records with 32 features and the Massive Yahoo Finance dataset containing 603k records with 9 features.p.1
• Financial dataset partitioning employs an 80% training and 20% testing split structured chronologically to prevent look-ahead bias.p.6
• The DAC-CPPO model yields a 1.91 Sharpe ratio, a 1.12 Cumulative return, and a 0.14 Volatility, outperforming baseline models.p.1
• Using a 5-fold expanding window time-series cross-validation, the DAC-CPPO model achieves a mean accuracy of 97.4% and an F1-score of 96.5%.p.28

Key insights

1. The DAC-CPPO framework integrates Dynamic Actor-Critic parallel learning with a clipped PPO constraint layer to stabilize policy updates during extreme market volatility.p.4
2. Features are preprocessed using Z-score normalization to ensure zero mean and unit variance, preventing training bias across volume and price technical indicators.p.6
3. Linear Discriminant Analysis (LDA) is applied as a dimensionality reduction technique to maximize between-class variance while reducing intra-class financial noise.p.7
4. The reinforcement learning environment utilizes 8 parallel worker agents to maintain computational efficiency and convergence stability on CPU hardware.p.14
5. Sharpe ratios in the study are calculated on an annualized basis across 252 trading days assuming a 2% annual risk-free rate.p.26