SBCA: Cross-Modal BERT-driven Actor-Critic for Multi-Asset Portfolio Optimization
SBCA: Cross-Modal BERT-driven Actor-Critic for Multi-Asset Portfolio Optimization
The paper introduces SBCA, a cross-modal BERT-driven Actor-Critic framework for multi-asset portfolio optimization. It addresses limitations of traditional models by fusing price time-series data with financial text sentiment using a novel gated fusion mechanism, which adaptively weights price features based on textual context. The reward function incorporates downside risk and turnover penalties to reflect real-world trading constraints, and the framework is validated on an 11-year U.S. stock dataset. SBCA outperforms equal weight, buy-and-hold, and market benchmarks in terms of portfolio value, annual return, Sharpe ratio, and maximum drawdown. Ablation studies confirm the complementary benefits of the cross-modal fusion and Actor-Critic components, while cost sensitivity analysis demonstrates robustness under varying transaction costs. The paper also provides theoretical proofs for the utility consistency of the reward function and the nonlinear expressiveness advantage of the gated fusion over linear concatenation, establishing a rigorous foundation for the approach.
Highlights
- 1Proposes a cross-modal BERT-driven Actor-Critic framework (SBCA) for multi-asset portfolio optimization, integrating price time-series and financial text sentiment via a gated fusion mechanism.
- 2Embeds downside risk and turnover penalty constraints into the reward function to align with real-world trading conditions.
- 3Outperforms equal weight, buy-and-hold, and market benchmarks on an 11-year U.S. stock dataset across portfolio value, annual return, Sharpe ratio, and maximum drawdown.
- 4Provides theoretical proof of utility consistency for the risk-sensitive reward function and nonlinear expressiveness advantage of the gated fusion over linear concatenation.
Methods
- MCross-modal gated fusion mechanism: adaptively integrates price features (via MLP) and BERT-extracted text sentiment features using a gating vector.
- MActor-Critic reinforcement learning: uses policy gradient (Actor) and value function (Critic) for continuous portfolio weight optimization.
- MRisk-sensitive reward function: includes downside risk penalty (squared negative log returns) and turnover penalty to enforce practical constraints.
- MEmpirical validation: 11-year backtest on U.S. multi-asset data with ablation studies and cost sensitivity analysis.
Results
- RSBCA achieves higher portfolio value, annual return, Sharpe ratio, and lower maximum drawdown compared to equal weight, buy-and-hold, and market benchmarks.
- RAblation studies confirm that both the cross-modal fusion module and Actor-Critic mechanism contribute complementary performance gains.
- RCost sensitivity analysis shows the model remains robust under varying transaction costs, maintaining superior risk-adjusted returns.
- RTheoretical analysis proves that the gated fusion can represent any linear concatenation function and can approximate nonlinear functions that linear concatenation cannot.
Analyze Paper
Generate insights from "SBCA: Cross-Modal BERT-driven Actor-Critic for Multi-Asset P...".