The paper introduces the Signature-Informed Transformer (SIT), a deep learning framework for dynamic asset allocation that addresses the fundamental mismatch between forecasting and portfolio optimization in traditional two-stage pipelines. SIT integrates three key innovations: (1) path signature features that encode complex temporal dynamics and cross-asset interactions from price histories, (2) a novel signature-augmented attention mechanism that injects geometric inductive biases about lead-lag relationships into the attention scores, and (3) direct optimization of portfolio weights by minimizing the Conditional Value-at-Risk (CVaR) of the portfolio loss distribution. The authors prove that the signature-based attention bias positively amplifies attention weights for aligned asset pairs, providing a theoretical foundation for the mechanism.

Empirically, SIT is evaluated on multiple equity universes (30, 40, 50 assets from S&P100; 10, 20 from DOW30; 50, 100 from CSI300) over a test period from 2020 to 2024. It consistently outperforms traditional strategies (equally weighted, GMV, HRP, CVaR optimization) and state-of-the-art deep learning forecasters (Autoformer, DLinear, FEDformer, iTransformer, NSformer, PatchTST, TimesNet, RFormer) combined with a downstream optimizer. For example, in the 50-asset S&P100 universe, SIT achieves a Sharpe ratio of 0.7715 and Sortino ratio of 0.9743, compared to the best baseline Sharpe of 0.5315. Ablation studies confirm that each component—CVaR objective, asset attention, financial bias, and the learnable gate—contributes significantly to performance. The work demonstrates that aligning model architecture and training objective with financial goals yields robust and superior portfolio allocations.