Paper: arXiv 2510.17165

Authors: Jinrui Zhang

Abstract

Foundation models - already transformative in domains such as natural language processing - are now starting to emerge for time-series tasks in finance. While these pretrained architectures promise versatile predictive signals, little is known about how they shape the risk profiles of the trading strategies built atop them, leaving practitioners reluctant to commit serious capital. In this paper, we propose an extension to the Capital Asset Pricing Model (CAPM) that disentangles the systematic risk introduced by a shared foundation model - potentially capable of generating alpha if the underlying model is genuinely predictive - from the idiosyncratic risk attributable to custom fine-tuning, which typically accrues no systematic premium. To enable a practical estimation of these separate risks, we align this decomposition with the concepts of uncertainty disentanglement, casting systematic risk as epistemic uncertainty (rooted in the pretrained model) and idiosyncratic risk as aleatory uncertainty (introduced during custom adaptations). Under the Aleatory Collapse Assumption, we illustrate how Monte Carlo dropout - among other methods in the uncertainty-quantization toolkit - can directly measure the epistemic risk, thereby mapping trading strategies to a more transparent risk-return plane. Our experiments show that isolating these distinct risk factors yields deeper insights into the performance limits of foundation-model-based strategies, their model degradation over time, and potential avenues for targeted refinements. Taken together, our results highlight both the promise and the pitfalls of deploying large pretrained models in competitive financial markets.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 6.5: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 6.5)

Why this score: The paper introduces a novel CAPM extension with heavy formalization (mean-variance optimization, epistemic/aleatory uncertainty disentanglement, Monte Carlo dropout) and sophisticated theoretical derivations, warranting a high math score. While it includes empirical experiments on real assets and backtesting frameworks for risk decomposition, the described implementation is more conceptual and tool-agnostic than fully code- or data-heavy, justifying a moderate-high rigor score.

Research Flowchart

  flowchart TD
  RQ["Research Goal: Disentangle Systematic vs. Idiosyncratic Risk<br>in Foundation Model Trading Strategies"]
  
  Inputs["Data/Inputs: Equity Market Data<br>Foundation Model Embeddings"]
  
  Method["Methodology: Extension of CAPM<br>+ Uncertainty Disentanglement"]
  
  Comp["Computational Process:<br>Monte Carlo Dropout for<br>Epistemic vs. Aleatory Risk Estimation"]
  
  Out["Key Findings:<br>Transparent Risk-Return Profiles<br>Identified Model Degradation<br>Refinement Avenues"]

  RQ --> Inputs
  Inputs --> Method
  Method --> Comp
  Comp --> Out

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 an extension to CAPM that separates systematic risk from a shared foundation model and uses uncertainty disentanglement (e.g., MC dropout) to measure epistemic risk.

Source sentence
In this paper, we propose an extension to the Capital Asset Pricing Model (CAPM) that disentangles the systematic risk introduced by a shared foundation model—potentially capable of generating alpha if the underlying model is genuinely predictive—from the idiosyncratic risk attributable to custom fine-tuning

Assumes 2

  • Aleatory Collapse: observable aleatory uncertainty is arbitraged away, so uncertainty estimator converges to epistemic uncertainty. market
    Source sentence
    Assumption 3.1. Aleatory Collapse. In a competitive market, any readily observable aleatory uncertainty is swiftly arbitraged away, due to its’ correlation with return.
  • The empirical study restricts to single‑asset trading to avoid portfolio aggregation effects. execution
    Source sentence
    we restrict our scope to single-asset trading. This choice avoids the added complexity of portfolio aggregation effects

Data

Backtest sample: U.S. equities · equities · intraday

  • U.S. equity NBBO · Interactive Brokers API order-book commercial
    2023-01-03 to 2025-01-31 · S&P 500 and Russell 2000 constituents
    Source sentence
    US Equity: One-second aggregates of the National Best Bid and Offer (NBBO) for a selection of S&P 500 and Russell 2000 constituents, collected through the ibkr API
  • Binance crypto BBO · Binance exchange order-book public
    2023-01-01 to 2025-01-31 · FDUSD‑quoted pairs
    Source sentence
    Cryptos: Best Bid and Offer quotes for FDUSD‑quoted pairs on Binance

Recognised sources: Binance

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Our backtest assumes orders are filled at the next tick’s best ask—a convention frequently described as a tick‑by‑tick backtest or immediate exec

Finds 3

Reported performance (TimesFM-v2): Sharpe 3.87

  • Isolating epistemic and aleatory risk yields deeper insights into performance limits, model degradation, and refinement avenues for foundation‑model strategies. theoretical
    Source sentence
    Our experiments show that isolating these distinct risk factors yields deeper insights into the performance limits of foundation‑model‑based strategies, their model degradation over time, and potential avenues for targeted refinements.
  • Foundation Sharpe ratios (PML slopes) are unstable for conventional baselines but more evident for pretrained architectures where shared signals dominate risk‑return. performance
    Source sentence
    of “shared risk” is not universally applicable—rather, it becomes most evident in pretrained architectures where common signals genuinely dominate the risk–return profile.
  • A rolling‑window analysis shows a gradual decay of alpha as foundation‑model edges become widely recognized, with observed Sharpe ratios falling below theoretical upper bounds. performance
    Source sentence
    ... the overall downward trend remains consistent, underscoring a gradual but persistent decay in alpha once a foundation model’s edge becomes widely recognized.

Checked by others

Supported by

  • Look-Ahead-Bench: a Standardized Benchmark of Look-ahead Bias in Point-in-Time LLMs for Finance (2026)
    On the question: Do foundation‑model strategies exhibit alpha decay over time?
    Compare the findings

    This paper: A rolling‑window analysis shows a gradual decay of alpha as foundation‑model edges become widely recognized, with observed Sharpe ratios falling below theoretical upper bounds.

    That paper: Standard foundation models show severe alpha decay, exceeding –15 percentage points between the in‑sample and out‑of‑sample periods, indicating strong look‑ahead bias.

    Both report decay: A shows gradual alpha loss with Sharpe below theoretical bounds; B shows a –15 pp drop from in‑sample to out‑of‑sample, supporting A.

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 (67)

  1. CAPITAL ASSET PRICES : A THEORY OF MARKET EQUILIBRIUM UNDER CONDITIONS OF RISK (1964)
  2. The Capital Asset Pricing Model : Theory and Evidence (2004)
  3. StockMixer: A Simple Yet Strong MLP - Based Architecture for Stock Price Forecasting . (2024)
  4. Risk models-at-risk (2014)
  5. Model risk of risk models (2016)
  6. Problems of Monetary Management : The UK Experience (1984)
  7. Unveiling the Influence of Artificial Intelligence and Machine Learning on Financial Markets : A Comprehensive Analysis of AI Applications in Trading , Risk Management , and Financial Operations (2023)
  8. AI in Finance : Challenges, Techniques , and Opportunities . (2023)
and 59 more
  1. Stock market’s price movement prediction with LSTM neural networks. (2017)
  2. Deep Learning for Portfolio Optimization (2020)
  3. Data- Driven Market - Making via Model - Free Learning . (2020)
  4. Stock price prediction using LSTM , RNN and CNN -sliding window model (2017)
  5. Predicting Stock Closing Prices in Emerging Markets with Transformer Neural Networks : The Saudi Stock Exchange Case (2021)
  6. Hierarchical Multi - Scale Gaussian Transformer for Stock Movement Prediction . (2020)
  7. Deep Direct Reinforcement Learning for Financial Signal Representation and Trading . (2017)
  8. Mastering Stock Markets with Efficient Mixture of Diversified Trading Experts (2023)
  9. Uncertainty- Aware Lookahead Factor Models for Quantitative Investing . (2020)
  10. Quantitative stock investment by routing uncertainty-aware trading experts: A multi-task learning approach (2022) — link
  11. BloombergGPT: A Large Language Model for Finance (2023)
  12. FinGPT: Open-Source Financial Large Language Models — → 2306.06031
  13. Financial Fine -tuning a Large Time Series Model — → 2412.09880
  14. Aleatory or epistemic? Does it matter? (2009)
  15. A decoder-only foundation model for time-series forecasting (2024)
  16. Chronos: Learning the Language of Time Series (2024)
  17. Unified Training of Universal Time Series Forecasting Transformers (2024)
  18. Timer: Generative Pre -trained Transformers Are Large Time Series Models (2024)
  19. Tiny Time Mixers ( TTMs ): Fast Pre -trained Models for Enhanced Zero / Few - Shot Forecasting of Multivariate Time Series (2024)
  20. Stock Price Prediction via Discovering Multi - Frequency Trading Patterns . (2017)
  21. Deep Learning with Gated Recurrent Unit Networks for Financial Sequence Predictions (2018)
  22. A Dual - Stage Attention - Based Recurrent Neural Network for Time Series Prediction . (2017)
  23. Stock market value prediction using neural networks. (2010)
  24. Understanding alpha decay (2022)
  25. Active Portfolio Management: A Quantitative Approach for Providing Superior Returns and Controlling Risk (2000)
  26. Deep reinforcement learning for automated stock trading: an ensemble strategy. (2020)
  27. A comprehensive evaluation of ensemble learning for stock-market prediction. (2020)
  28. Risk assessment and risk management: Review of recent advances on their foundation (2016)
  29. Value at risk-new approaches to risk management (1996)
  30. Value at risk when daily changes in market variables are not normally distributed (1998)
  31. Evaluating value-at-risk models with desk-level data (2011)
  32. Model Risk (1996)
  33. Mean-modified value-at-risk optimization with hedge funds (2002)
  34. Expected shortfall: a natural coherent alternative to value at risk (2002)
  35. How to manage risk (after risk management has failed) (2010)
  36. Introduction to applied Bayesian statistics and estimation for social scientists (2007)
  37. Bayesian forecasting of value at risk and expected shortfall using adaptive importance sampling (2010)
  38. Uncertainty in value-at-risk estimates under parametric and non-parametric modeling (2006)
  39. Understanding and Managing Model Risk: A practical guide for quants, traders and validators (2011)
  40. Hedge fund modelling and analysis using Excel and VBA (2012)
  41. Supervisory Guidance On Model Risk Management (2011)
  42. Dropout as a bayesian approximation: Representing model uncertainty in deep learning (2016)
  43. What uncertainties do we need in bayesian deep learning for computer vision? (2017)
  44. Aleatory and epistemic uncertainty in probability elicitation with an example from hazardous waste management (1996)
  45. Portfolio Selection (1952) — link
  46. A Practical Bayesian Framework for Backpropagation Networks . (1992)
  47. Bayesian learning for neural networks (2012)
  48. Simple and scalable predictive uncertainty estimation using deep ensembles (2017)
  49. Simple and principled uncertainty estimation with deterministic deep learning via distance awareness (2020)
  50. A simple baseline for bayesian uncertainty in deep learning (2019)
  51. Laplace redux-effortless bayesian deep learning (2021)
  52. Uncertainty estimation using a single deep deterministic neural network (2020)
  53. Deep deterministic uncertainty: A new simple baseline (2023)
  54. On calibration of modern neural networks (2017)
  55. Decomposition of uncertainty in Bayesian deep learning for efficient and risk-sensitive learning (2018)
  56. Quantifying aleatoric and epistemic uncertainty in machine learning: Are conditional entropy and mutual information appropriate measures? (2023)
  57. Benchmarking uncertainty disentanglement: Specialized uncertainties for specialized tasks (2024) — link
  58. Evidential deep learning to quantify classification uncertainty (2018)
  59. FinVision: A Multi - Agent Framework for Stock Market Prediction — → 2411.08899

3 of 67 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.