Technical Summary: Signature-Informed Transformer for Asset Allocation

1. Problem Statement

Modern quantitative finance faces a central challenge in strategic asset allocation: constructing portfolios robust to the complex, non-linear, and path-dependent nature of financial markets. The predominant deep learning paradigm for this task employs a decoupled, two-stage pipeline: a forecasting model predicts asset returns (typically optimized for statistical metrics like Mean Squared Error), and these predictions are subsequently fed into a downstream portfolio optimizer.

The authors argue this approach suffers from two critical flaws:

1. Lack of Financial Inductive Bias: General-purpose forecasting architectures fail to model idiosyncratic market structures, such as intricate lead-lag relationships between assets, making it difficult to distinguish genuine signals from noise.
2. Objective Mismatch and Error Amplification: Optimizing for prediction accuracy (e.g., minimizing MSE) is agnostic to the downstream goal of portfolio construction. Minuscule prediction errors can be magnified by the optimizer into volatile, impractical, or suboptimal portfolio weights. Furthermore, MSE objectives implicitly incentivize models to favor easily predictable assets, potentially distorting the final allocation by ignoring harder-to-predict assets with larger estimation errors.

The paper posits that a robust solution requires a unified policy that learns an end-to-end mapping from market data directly to portfolio weights, while being architecturally designed to model the known geometric properties of financial time series.

2. Methodology: The Signature-Informed Transformer (SIT)

The authors propose the Signature-Informed Transformer (SIT), a deep learning framework that unifies feature extraction and decision-making. The architecture is built on three synergistic pillars:

A. Path-wise Feature Representation

Instead of raw price inputs, SIT utilizes Rough Path Signatures to generate features. For a price path $X$, the signature $\text{Sig}(X)$ is a sequence of iterated integrals that provides a principled summary of the path's shape, encoding trends and oscillations.

• First-order terms capture net increments.
• Second-order terms encode signed areas, which the authors prove (Proposition 2.1) strictly capture lead-lag relationships between assets. If asset $X_1$ strictly leads $X_2$, the second-level cross-term $A(X) = \int X_1 dX_2 - \int X_2 dX_1$ is strictly positive.
• The model constructs input tokens by fusing these signature embeddings with time embeddings (calendar features) and learnable asset-specific embeddings.

B. Signature-Augmented Attention

To model dependencies between assets, SIT introduces a novel Signature-Informed Self-Attention mechanism.

• Unlike standard attention, this mechanism injects a dynamic bias term into the attention logits based on the cross-signature of asset pairs.
• The bias is computed as the inner product between a dynamic query vector (representing the current informational need of an asset) and a relational embedding derived from the pairwise signature.
• Theoretical Guarantee: Theorem 2.2 proves that the directional derivative of the attention weight with respect to the signature bias is strictly positive. This ensures that when a signature-encoded relationship aligns with an asset's current state, the model rigorously amplifies the attention to that counterpart. A learnable gate $\gamma$ controls the magnitude of this influence.

C. Decision Alignment (Training Objective)

SIT abandons the two-stage "predict-then-optimize" approach. Instead, it is trained end-to-end to directly minimize the Conditional Value-at-Risk (CVaR) of the portfolio's loss distribution.

• The model outputs predicted returns which are converted to portfolio weights via a softmax function.
• The loss function is defined as $\min_\theta E_\omega[\text{CVaR}_\alpha(\{L^{(k)}\}_{k=1}^K)]$, where $L^{(k)}$ is the portfolio loss over $K$ future steps.
• No auxiliary prediction losses (like MSE) are used. This ensures the training objective is perfectly aligned with the financial goal of risk-adjusted return, avoiding the error amplification inherent in decoupled pipelines.

3. Key Contributions

1. Unified Architecture: A single policy that maps market data directly to portfolio weights, eliminating the objective mismatch between forecasting and optimization.
2. Geometric Inductive Bias: The integration of path signatures to explicitly encode lead-lag relationships and higher-order path dependencies, which are mathematically proven to be robust measures of asset interaction.
3. Signature-Informed Attention: A novel attention mechanism that dynamically amplifies attention based on geometric asset relationships, with a theoretical proof of its monotonicity.
4. Risk-Aware Optimization: Direct minimization of portfolio CVaR, ensuring the model learns to hedge against tail events rather than merely minimizing point-wise prediction errors.

4. Experimental Results

The authors evaluated SIT across diverse equity universes (30, 40, 50 assets from S&P 100; 10, 20 from DOW30; 50, 100 from CSI 300) covering training (2000–2016), validation (2017–2019), and testing (2020–2024) periods.

• Performance: SIT significantly outperformed both traditional strategies (Equal Weight, Global Minimum Variance, HRP) and advanced deep learning baselines (Autoformer, DLinear, iTransformer, etc.) across all metrics: Sharpe Ratio, Sortino Ratio, Maximum Drawdown, and Final Wealth Factor.
  • Example: In the 40-asset universe, SIT achieved a Sharpe Ratio of 0.6717, surpassing the next-best deep learning baseline (RFormer at 0.4901) and traditional baselines.
• Ablation Studies:
  • Removing the CVaR objective (replacing it with mean return maximization) caused a significant drop in performance, validating the necessity of decision-focused learning.
  • Removing the signature-based attention or the financial bias term resulted in degraded performance, confirming that geometric inductive biases are superior to learning these relationships from scratch.
  • Removing the learnable gate $\gamma$ also hurt performance, highlighting the need to dynamically modulate financial priors.
• Mechanism Validation: Analysis showed a strong positive correlation (mean ~0.54–0.61) between signature strength and inbound attention weights, confirming that the model actively utilizes signature-derived signals to route information.
• Robustness: SIT maintained superior performance under varying transaction costs and allocation concentration parameters, demonstrating that its gains stem from robust allocation strategies rather than raw prediction accuracy.

5. Significance and Claims

The paper claims that effective quantitative portfolio management requires robust allocation policies rather than just optimized prediction accuracy. The authors argue that the "predict-then-optimize" paradigm is fundamentally flawed due to objective mismatch and error amplification.

By unifying feature extraction (via signatures), interaction modeling (via signature-augmented attention), and decision-making (via CVaR minimization), SIT provides a blueprint for moving machine learning in finance from forecasting towards end-to-end, risk-aware capital allocation. The authors assert that their approach successfully bridges the gap between theoretical financial goals and practical deep learning implementation, offering a model that is architecturally designed to understand the underlying geometry of market dynamics.

The work is presented as a step toward more stable, real-world performance in asset allocation, with the code made available for reproducibility. The authors note that while tested on U.S. and Chinese equity data, the framework is theoretically extensible to higher-frequency and multi-asset markets.