DAC-CPPO model reports stronger risk-adjusted portfolio allocation performance

A reinforcement learning model developed for portfolio allocation reported stronger risk-adjusted performance than the models compared in the study. The proposed Dynamic Actor-Critic-enriched Clipped Proximal Policy Optimization model, abbreviated DAC-CPPO, achieved reported accuracy of 97.6%, a mean absolute error of 0.074, a root mean square error of 0.081, a Sharpe ratio of 1.91, cumulative return of 1.12 and volatility of 0.14.

The authors framed capital allocation as a problem requiring an adaptive approach to dynamic market conditions, risk factors and complex financial data. At the center of DAC-CPPO is a Dynamic Actor-Critic module that evaluates actions and learns portfolio policies intended to maximize long-term returns. The Clipped Proximal Policy Optimization mechanism is used to control policy updates and support stable, risk-conscious portfolio optimization.

Evaluation used two financial datasets collected from publicly available Kaggle sources. The Intelligent Finance Assets dataset contained 7,000 records with 32 features, while the Massive Yahoo Finance dataset contained 603,000 records with nine features. The researchers applied data cleaning and normalization to integrate the datasets for model readiness. They also used linear discriminant analysis to extract discriminative financial features intended to improve portfolio allocation prediction performance, with the experimental implementation conducted in Python.

The reported Sharpe ratio, cumulative return and volatility values indicate that the evaluation considered both return and risk rather than allocation accuracy alone. Within the study’s comparison, DAC-CPPO produced the strongest reported risk-adjusted performance, with higher returns while limiting volatility. However, the supplied research description does not identify the financial instruments, evaluation period, transaction-cost assumptions, validation design or the precise meaning of the accuracy measure. It also does not report results from live trading, so the findings should be interpreted as the study’s reported evaluation rather than evidence of future investment performance.

Subject of Research: Evaluation of a reinforcement learning model for portfolio allocation and risk-adjusted performance

Article Title: Reinforcement learning-driven optimal capital allocation for intelligent finance

Article References: Luo, Y., Qi, M., & Cheng, M. (2026). Reinforcement learning-driven optimal capital allocation for intelligent finance. PLOS One, 21(10), e0348010. https://doi.org/10.1371/journal.pone.0348010

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0348010

Keywords: reinforcement learning, portfolio optimization, capital allocation, dynamic actor-critic, clipped policy optimization, financial datasets, risk-adjusted returns, Sharpe ratio, volatility, machine learning, Reinforcement, learning-driven