SBCA: cross-modal gated actor-critic with frozen BERT-based sentiment features for multi-asset portfolio optimization

Portfolio optimization is constrained by linear assumptions and limited integration of multi-modal information. Throughout this study, BERT serves only as a text encoder producing a scalar sentiment score from headlines; it is fine-tuned offline once and then held fixed, so its parameters are neither updated nor jointly optimized with the policy, and its scores are pre-computed static inputs. We propose a cross-modal gated Actor-Critic framework (SBCA) for multi-asset portfolio optimization, addressing two gaps in deep reinforcement learning: reliance on a single affine concatenation layer for price-text fusion, and the absence of downside-risk and turnover penalties in reward designs. SBCA learns gated cross-modal interactions and embeds both penalties in its reward. On 11 years of U.S. equity data, SBCA outperforms Equal Weight, Buy & Hold, the Dow Jones index, MVO and PPO on all metrics in the 2- and 4-asset groups, and in the 6-asset group trails only Equal Weight and MVO in absolute return. Block-bootstrap tests give exploratory, uncorrected evidence for the Actor-Critic component over gated fusion in the 2-asset group, though no component-level comparison survives correction; cost sensitivity gives descriptive evidence of robustness. SBCA thus combines offline text preprocessing with a jointly trained policy for dynamic portfolio decision-making.