Paper: arXiv 2510.03129

Authors: Yoontae Hwang, Stefan Zohren

Abstract

Robust asset allocation is a key challenge in quantitative finance, where deep-learning forecasters often fail due to objective mismatch and error amplification. We introduce the Signature-Informed Transformer (SIT), a novel framework that learns end-to-end allocation policies by directly optimizing a risk-aware financial objective. SIT’s core innovations include path signatures for a rich geometric representation of asset dynamics and a signature-augmented attention mechanism embedding financial inductive biases, like lead-lag effects, into the model. Evaluated on daily S&P 100 equity data, SIT decisively outperforms traditional and deep-learning baselines, especially when compared to predict-then-optimize models. These results indicate that portfolio-aware objectives and geometry-aware inductive biases are essential for risk-aware capital allocation in machine-learning systems. The code is available at: https://github.com/Yoontae6719/Signature-Informed-Transformer-For-Asset-Allocation

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs only partially modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper employs advanced mathematics including Rough Path Theory and path signatures, and integrates a custom attention mechanism with financial inductive biases, giving a high math score. It reports out-of-sample backtests on S&P 100 data, provides code, and optimizes a direct financial objective (CVaR), indicating strong empirical grounding.

Research Flowchart

  flowchart TD
  subgraph A ["Research Goal"]
      goal["Robust Asset Allocation via<br>End-to-End Policy Learning"]
  end

  subgraph B ["Methodology"]
      sig["Path Signatures<br>Geometric Feature Extraction"]
      att["Signature-Augmented Attention<br>Lead-Lag Inductive Bias"]
      sit["Signature-Informed Transformer SIT"]
  end

  subgraph C ["Data & Experiment"]
      data["S&P 100 Daily Equity Data"]
      obj["Risk-Aware Financial Objective"]
  end

  subgraph D ["Outcomes"]
      res["Decisive Outperformance<br>vs. Predict-Then-Optimize & Baselines"]
      conc["Portfolio-Aware Objectives<br>& Geometry Biases are Essential"]
  end

  goal --> sig
  sig --> sit
  att --> sit
  sit --> obj
  data --> obj
  obj --> res
  res --> conc

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes the Signature‑Informed Transformer (SIT), a unified deep‑learning model that extracts path‑signature features and makes portfolio allocation decisions in a single policy.

Source sentence
We propose the Signature Informed Transformer to address this by unifying feature extraction and decision making into a single policy.

Assumes 4

  • The attention gate scalar γ is strictly positive. model
    Source sentence
    c bias matrix is scaled by a learnable, strictly positive scalar gate, γ > 0 (parameterized as γ = softplus(ˆγ))
  • A zero risk‑free rate is used for all risk‑adjusted metrics. market
    Source sentence
    assuming a zero risk‑free rate.
  • Portfolios are long‑only, fully invested, and rebalanced monthly. execution
    Source sentence
    long‑only, fully invested, monthly (k‑step) rebalancing on the 40‑ and 50‑asset universes
  • Transaction costs are one‑way proportional fees applied per dollar traded. execution
    Source sentence
    Transaction costs are one‑way proportional fees of c ∈0, 5, 10 basis points per dollar traded.

Data

Backtest sample: S&P 100 stocks · equities · daily

  • S&P 100 price data · Wharton Research Data Services (WRDS) market-prices commercial
    2000‑01‑01 to 2024‑12‑27 · 30, 40, and 50 S&P 100 companies
    Source sentence
    ... The daily price data was sourced from Wharton Research Data Services (WRDS).
  • Dow30 price data · Wharton Research Data Services (WRDS) market-prices commercial
    2000‑01‑01 to 2024‑12‑27 · 10 and 20 assets from the DOW30
    Source sentence
    We also selected two additional portfolios of 10 and 20 assets from the DOW30 to validate performance against a different index composition
  • CSI 300 price data · Wharton Research Data Services (WRDS) market-prices commercial
    2000‑01‑01 to 2024‑12‑27 · 50 and 100 assets from the CSI 300 index
    Source sentence
    included two portfolios consisting of 50 and 100 assets from the CSI 300 index.

Recognised sources: WRDS

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Partial

Paper text extraction: Modeled - One‑way proportional fees of 0, 5, or 10 basis points per dollar traded are applied in the sensitivity analysis.

Source sentence (the authors' words)
Transaction costs are one‑way proportional fees of c ∈0, 5, 10 basis points per dollar traded.

Finds 3

Reported performance (Signature-Informed Transformer (SIT)): Sharpe 0.77 · max drawdown 0.3%

  • Replacing the CVaR objective with a risk‑neutral mean‑return objective reduces Sharpe from 0.6717 to 0.5691, showing the importance of decision‑focused loss. performance
    Source sentence
    When the Conditional Value‑at‑Risk (CVaR) loss is replaced with a standard risk‑neutral objective (w/o CVaR), the Sharpe Ratio on the 40‑asset universe falls from 0.6717 to 0.5691.
  • Performance peaks at moderate allocation dispersion (τ ≈ 1.3) and degrades linearly with transaction costs, losing about 0.03‑0.04 Sharpe when costs increase from 0 to 10 bps. performance
    Source sentence
    First, performance peaks at moderate dispersion, near τ ≈1.3. Second, frictions compress Sharpe roughly linearly over this range: moving from 0 to 10 bps reduces Sharpe by about 0.03–0.04 at the optimum.
  • SIT decisively outperforms traditional and advanced forecasting baselines across diverse equity universes. performance
    Source sentence
    Experiments across diverse equity universes show our approach significantly outperforms both traditional strategies and advanced forecasting baselines.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 23 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (63)

  1. The signature of a rough path: uniqueness (2016)
  2. A primer on the signature method in machine learning (2016) — link
  3. Signatory: differentiable computations of the signature and logsignature transforms, on both CPU and GPU (2020) — link
  4. Differential equations driven by rough signals (1998)
  5. Differential equations driven by rough paths: Ecole d’Et 'e de Probabilit 'e s de Saint-Flour XXXIV-2004 (2007)
  6. Building diversified portfolios that outperform out-of-sample (2016)
  7. Optimization of conditional value-at-risk (2000)
  8. Deep Learning in Asset Management: Architectures, Applications, and Challenges (2025)
and 55 more
  1. Minimum-variance portfolio composition (2011)
  2. Optimal versus naive diversification: How inefficient is the 1/N portfolio strategy? (2009)
  3. Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
  4. Are transformers effective for time series forecasting? (2023)
  5. Are language models actually useful for time series forecasting? (2024)
  6. Timemixer: Decomposable multiscale mixing for time series forecasting (2024) — link
  7. Timexer: Empowering transformers for time series forecasting with exogenous variables (2024)
  8. itransformer: Inverted transformers are effective for time series forecasting (2023) — link
  9. A globally optimal portfolio for m-sparse sharpe ratio maximization (2024)
  10. Portfolio Selection (1952) — link
  11. Decision-informed Neural Networks with Large Language Model Integration for Portfolio Optimization — → 2502.00828
  12. An Overview of Machine Learning for Portfolio Optimization. (2024)
  13. A universal end-to-end approach to portfolio optimization via deep learning (2021) — link
  14. Deep learning with long short-term memory networks for financial market predictions (2018)
  15. Attention is all you need (2017)
  16. Etsformer: Exponential smoothing transformers for time-series forecasting (2022) — link
  17. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting (2022)
  18. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
  19. Timesnet: Temporal 2d-variation modeling for general time series analysis (2022) — link
  20. A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
  21. Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
  22. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting (2019)
  23. Anatomy of Machines for Markowitz: Decision-Focused Learning for Mean-Variance Portfolio Optimization — → 2409.09684
  24. Accurate multivariate stock movement prediction via data-axis transformer with multi-level contexts (2021)
  25. Returns-Driven Macro Regimes and Characteristic Lead-Lag Behaviour between Asset Classes (2022) — link
  26. Detecting lead-lag relationships in stock returns and portfolio strategies (2023)
  27. Efficient capital markets (1970)
  28. Empirical properties of asset returns: stylized facts and statistical issues (2001)
  29. Robust risk measurement and model risk (2014)
  30. Deep hedging (2019)
  31. Learning to trade via direct reinforcement (2001)
  32. Extreme correlation of international equity markets (2001)
  33. Signature methods in machine learning (2022) — link
  34. Sigformer: Signature transformers for deep hedging (2023)
  35. Machine learning and portfolio optimization (2018)
  36. Pairs trading (2005)
  37. Extracting information from the signature of a financial data stream (2013) — link
  38. Optimal execution with rough path signatures (2020)
  39. Stochastic gradient methods for distributionally robust optimization with f-divergences (2016)
  40. Detecting data-driven robust statistical arbitrage strategies with deep neural networks (2024)
  41. Deep signature transforms (2019) — link
  42. Hist: A graph-based framework for stock trend forecasting via mining concept-oriented shared information (2021) — link
  43. FactorGCL: A Hypergraph-Based Factor Model with Temporal Residual Contrastive Learning for Stock Returns Prediction — → 2502.05218
  44. Temporal fusion transformers for interpretable multi-horizon time series forecasting (2021)
  45. Neural controlled differential equations for irregular time series (2020)
  46. Sig-SDEs model for quantitative finance (2020)
  47. Rough transformers: Lightweight and continuous time series modelling through signature patching (2024)
  48. The signature kernel is the solution of a goursat pde (2021)
  49. Kernels for sequentially ordered data (2019)
  50. Distributionally robust end-to-end portfolio construction (2023)
  51. Deep Learning in Quantitative Trading (2025)
  52. Enhancing time series momentum strategies using deep neural networks (2019) — link
  53. Deep learning for portfolio optimization (2020) — link
  54. A universal end-to-end approach to portfolio optimization via deep learning, 2021
  55. The effects of errors in means, variances, and correlations on the mean-variance framework (2022)

3 of 63 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.