Paper: arXiv 2510.26228

Authors: Nikolas Anic, Andrea Barbon, Ralf Seiz, Carlo Zarattini

Abstract

This paper investigates whether large language models (LLMs) can improve cross-sectional momentum strategies by extracting predictive signals from firm-specific news. We combine daily U.S. equity returns for S&P 500 constituents with high-frequency news data and use prompt-engineered queries to ChatGPT that inform the model when a stock is about to enter a momentum portfolio. The LLM evaluates whether recent news supports a continuation of past returns, producing scores that condition both stock selection and portfolio weights. An LLM-enhanced momentum strategy outperforms a standard long-only momentum benchmark, delivering higher Sharpe and Sortino ratios both in-sample and in a truly out-of-sample period after the model’s pre-training cut-off. These gains are robust to transaction costs, prompt design, and portfolio constraints, and are strongest for concentrated, high-conviction portfolios. The results suggest that LLMs can serve as effective real-time interpreters of financial news, adding incremental value to established factor-based investment strategies.

Complexity vs Empirical Score

  • Math Complexity: 2.5/10
  • Empirical Rigor: 9.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s core mathematical complexity is moderate, focusing on standard momentum signal construction and portfolio weighting, with no advanced statistical theory or heavy derivations. However, it demonstrates high empirical rigor by utilizing high-frequency news and equity datasets, conducting in-sample and out-of-sample backtests with robustness checks against transaction costs and prompt variations, and reporting performance metrics like Sharpe and Sortino ratios.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Can LLMs enhance<br>cross-sectional momentum<br>strategies using news?"]

  B["Data & Inputs<br>Daily S&P 500 Returns +<br>High-Frequency News Data"]

  C["Core Methodology<br>ChatGPT Prompt Engineering<br>Stock enters momentum portfolio?"]

  D["Computational Process<br>LLM Evaluation Score<br>Conditions Selection & Weights"]

  E["Key Outcome 1<br>LLM-Enhanced Strategy<br>Outperforms Benchmark"]

  F["Key Outcome 2<br>Higher Sharpe & Sortino Ratios<br>Robust Out-of-Sample Results"]

  A --> B
  B --> C
  C --> D
  D --> E
  D --> F

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 LLM‑enhanced cross‑sectional momentum strategy that uses ChatGPT scores from firm‑specific news to condition stock selection and portfolio weights.

Source sentence
The LLM then evaluates whether recent news supports a continuation of past returns, producing a score that conditions both stock selection and portfolio weights.

Assumes 5

  • Transaction costs are assumed to be 2 basis points per trade. execution
    Source sentence
    These performance gains remain robust even after accounting for conservative estimates of transaction costs of 2 bps.
  • Monthly rebalancing is used because it yields lower turnover and costs than weekly rebalancing. execution
    Source sentence
    We find that a monthly rebalancing frequency yields significantly better results than a weekly one.
  • The baseline and LLM‑enhanced portfolios share the same universe and rebalancing frequency, isolating the effect of the LLM. market
    Source sentence
    Both strategies are described in Section 2. To ensure comparability, the baseline universe and the rebalancing frequency are the same for both portfolios.
  • The LLM model and hyperparameters are selected solely on validation‑set performance to avoid overfitting. model
    Source sentence
    The LLM model, together with its hyperparameters, is selected based solely on Validation Set performance.
Show 1 more
  • Out‑of‑sample period starts after the model’s pre‑training cutoff, so results are not due to information leakage. data
    Source sentence
    Importantly, the out‑of‑sample period begins after the model’s pre‑training cut‑off, ensuring that the observed gains cannot be attributed to information leakage.

Data

Backtest sample: S&P 500 stocks · equities

  • Finreon internal dataset · Finreon (internal) market-prices proprietary
    U.S. equities (S&P 500 constituents)
    Source sentence
    First, we use an internal Finreon dataset, similar in structure and coverage to the CRSP database, which provides daily total returns and market capitalization data for U.S. equities.
  • Stock News API · Stock News API text-news commercial
    Firm‑level news for S&P 500 constituents
    Source sentence
    Second, we obtain high-frequency stock‑specific news from the Stock News API. This dataset includes publication timestamps with second‑level precision...
  • FRED risk‑free rate series · Federal Reserve Economic Data (FRED) macro public
    Source sentence
    Third, we retrieve the U.S. risk‑free rate from the Federal Reserve Economic Data (FRED) database.

Recognised sources: FRED CRSP

Costs

Modeled Costs are part of the reported headline results.

2 basis points per trade (conservative estimate for highly liquid S&P 500 stocks)

Source sentence
All results are computed under the assumption of 2 basis points of transaction costs per trade, which is a conservative estimate given the high liquidity of S&P 500 stocks in our sample.

Finds 5

Reported performance (LLM-enhanced momentum strategy): Sharpe 0.69 (benchmark 0.57) · annual return 18.0% · max drawdown 31.0%

  • In the full sample the LLM‑enhanced portfolio achieves a Sharpe ratio of 0.69 versus 0.57 for the baseline. performance
    Source sentence
    In the full‑sample evaluation, the LLM‑enhanced portfolio exhibits a Sharpe ratio of 0.69, compared to 0.57 for the baseline.
  • Out‑of‑sample Sharpe improves to 1.06 from 0.79, and Sortino to 1.28 from 0.93, with returns rising from ~24% to 30%. performance
    Source sentence
    ... the Sharpe ratio improves from 0.79 for the baseline to 1.06 for the LLM‑enhanced strategy, while the Sortino ratio increases from 0.93 to 1.28. The total return increases from approximately 24% to 30%.
  • Volatility declines slightly (full sample 24% vs 26%; out‑of‑sample 22% vs 24%) and maximum drawdown improves (‑31% vs ‑33%). performance
    Source sentence
    Volatility decreases slightly from 26% to 24%, while the maximum drawdown improves from −33% to −31%.
  • Turnover is higher for the LLM‑enhanced strategy but net returns remain superior after accounting for the 2 bps costs. performance
    Source sentence
    all reported results are net of transaction costs, the increased turnover is offset by the stronger predictive signals.
  • A time‑series regression yields an annualized alpha of 3.26% for the LLM‑enhanced returns, significant only at the 10% level. statistical
    Source sentence
    a time‑series regression of LLM‑enhanced returns on baseline returns yields a positive annualized alpha of 3.26%, but the estimate is statistically significant only at the 10% level.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 56 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 (15)

  1. Chronologically Consistent Large Language Models — → 2502.21206
  2. Returns to buying winners and selling losers: Implications for stock market efficiency (1993)
  3. Momentum strategies (1996)
  4. On persistence in mutual fund performance (1997)
  5. Post-earnings-announcement drift: delayed price response or risk premium? (1989)
  6. A unified theory of underreaction, momentum trading, and overreaction in asset markets (1999)
  7. Giving content to investor sentiment: The role of media in the stock market (2007)
  8. When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks (2011)
and 7 more
  1. Text mining for market prediction: A systematic review (2014)
  2. Predicting returns with text data (2019)
  3. Empirical asset pricing via machine learning (2020)
  4. Can chatgpt forecast stock price movements? return predictability and large language models — → 2304.07619
  5. Financial statement analysis with large language models — → 2407.17866
  6. Automated machine learning (2019)
  7. Explanatory model analysis: explore, explain, and examine predictive models (2021)

3 of 15 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.