Paper: arXiv 2412.09394

Authors: Sebastien Valeyre, Sofiane Aboura

Abstract

Recently, LLMs (Large Language Models) have been adapted for time series prediction with significant success in pattern recognition. However, the common belief is that these models are not suitable for predicting financial market returns, which are known to be almost random. We aim to challenge this misconception through a counterexample. Specifically, we utilized the Chronos model from Ansari et al.(2024) and tested both pretrained configurations and fine-tuned supervised forecasts on the largest American single stocks using data from Guijarro-Ordonnez et al.(2022). We constructed a long/short portfolio, and the performance simulation indicates that LLMs can in reality handle time series that are nearly indistinguishable from noise, demonstrating an ability to identify inefficiencies amidst randomness and generate alpha. Finally, we compared these results with those of specialized models and smaller deep learning models, highlighting significant room for improvement in LLM performance to further enhance their predictive capabilities.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: reported Sharpe ratio of 4.21 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.5)

Why this score: The paper employs advanced deep learning architectures (LLMs/Transformers with >11M parameters) and sophisticated financial modeling (residual returns, long/short portfolios), reflecting high mathematical complexity. It also features substantial empirical work using large-scale US stock data, pre-trained/fine-tuned models, and performance simulations, indicating strong data and implementation rigor.

Research Flowchart

  flowchart TD
  Goal["Research Goal:<br>Challenge belief that LLMs<br>cannot predict financial returns<br>(quasi-random data)"] --> Method["Methodology:<br>Utilize Chronos Model<br>(Pretrained & Fine-tuned)"]
  Method --> Data["Data Input:<br>Largest American Single Stocks<br>(Guijarro-Ordonnez et al., 2022)"]
  Data --> Process["Computational Process:<br>1. Time Series Prediction<br>2. Generate Trading Signals<br>3. Construct Long/Short Portfolio"]
  Process --> Findings["Key Findings & Outcomes:<br>1. LLMs identify inefficiencies in noise<br>2. Generate Alpha (Excess Returns)<br>3. Compare with specialized/smaller DL models<br>4. Room for performance improvement"]

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 using the Chronos large language model, both pretrained and fine‑tuned, to forecast residual returns of large US single stocks and build a long/short portfolio.

Source sentence
Specifically, we utilized the Chronos model from Ansari et al.(2024) and tested both pretrained configurations and fine‑tuned supervised forecasts on the largest American single stocks using data from Guijarro‑Ordonnez et al.(2022).

Assumes 5

  • The performance metric ignores slippage and trading costs. execution
    Source sentence
    We based our results on the gross Sharpe ratio metric of the portfolio which corresponds to the outsample annualized average of daily returns of the portfolio divided by the annualized standard deviation of daily returns, meaning that we do not consider slippage and trading costs.
  • Only large, liquid stocks (e.g., S&P 500 constituents) are used to ensure robustness. market
    Source sentence
    Their methodology underscores the importance of focusing on larger, more liquid stocks, such as those in the S&P 500, to ensure robustness in empirical findings.
  • Residual returns are less correlated and thus suitable for deep learning. data
    Source sentence
    Using residual returns allows for a less correlated dataset, which is crucial for deep learning.
  • The zero‑shot Chronos model can be applied without a larger dataset because it is pretrained on non‑financial data. model
    Source sentence
    Also as we mainly use the pretrained case of Chronos, we did not need a larger dataset for the zeros shot version or the fine tuned one.
Show 1 more
  • A 3 basis‑point per‑trade cost approximates typical market impact and brokerage fees. execution
    Source sentence
    A transaction cost of 3 basis points per trade is consistent with the average market impact—typically between 2 and 3 basis points for very small orders—and brokerage fees of approximately 1 basis point.

Data

Backtest sample: 2002 – 2016 · US single stocks · equities · daily

  • IPCA residual returns · Guijarro‑Ordonnez et al. (2021) – GitHub market-prices public
    January 1978 to end of 2016 · CRSP securities universe
    Source sentence
    The three different datasets of daily residual returns (IPCA, PCA and FF) provided by Guijarro‑Ordonnez et al. (2021) are all based on the same raw universe of securities from the CRSP dataset spanning January 1978 to the end of 2016.
  • PCA residual returns · Guijarro‑Ordonnez et al. (2021) – GitHub market-prices public
    January 1978 to end of 2016 · CRSP securities universe
    Source sentence
    The three different datasets of daily residual returns (IPCA, PCA and FF) provided by Guijarro‑Ordonnez et al. (2021) are all based on the same raw universe of securities from the CRSP dataset spanning January 1978 to the end of 2016.
  • FF residual returns · Guijarro‑Ordonnez et al. (2021) – GitHub market-prices public
    January 1978 to end of 2016 · CRSP securities universe
    Source sentence
    The three different datasets of daily residual returns (IPCA, PCA and FF) provided by Guijarro‑Ordonnez et al. (2021) are all based on the same raw universe of securities from the CRSP dataset spanning January 1978 to the end of 2016.

Recognised sources: CRSP

Costs

Partial Costs appear only as a robustness check, a flat fee without spread or impact, or outside the headline numbers.

A flat 3 basis‑point slippage cost per trade is applied in a robustness check, turning positive gross Sharpe ratios into negative net Sharpe ratios.

Source sentence
trading costs are prohibitive, as including a 3 basis point slippage

Finds 4

Reported performance (Chronos (LLM) with resized fine‑tuned weights): Sharpe 4.21

  • The pretrained Chronos model with α = 0.3 achieves a gross Sharpe ratio above 3.17 for the PCA dataset over a 15‑year period. performance
    Source sentence
    We observe that the pre‑trained Chronos model with the autoregressive coefficient α = 0.3 effectively identifies opportunities in the financial market (see Table 2), achieving a Sharpe ratio above 3.17 for PCA over a 15‑year period, which corresponds to a t‑statistic of 3.17 √15 = 12.27.
  • When a 3‑basis‑point transaction cost is applied, the net Sharpe ratio for PCA drops from 3.17 to –1.49. negative
    Source sentence
    With α = 0.3, the gross Sharpe ratio for PCA is 3.17, but it drops to -1.49 after accounting for a 3 basis point trading cost.
  • Chronos becomes highly correlated with the Short‑Term Reversal (STR) strategy (β = 0.3) but does not outperform it. performance
    Source sentence
    However, it ultimately gets very correlated to the Short‑Term Reverseal (STR) strategy with β = 0.3 but fails to outperform it.
  • The STR strategy outperforms AutoARIMA, showing higher Sharpe ratios across factor datasets. performance
    Source sentence
    That assertion is supported by the case of the strong outperformance of the STR compared to AutoARIMA, which can capture more complexity but whose noisy fit makes it suboptimal and overly erratic.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

https://github.com/dkyol/Asset-Pricing-Model — the authors' own repository

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