Paper: arXiv 2404.05101
Authors: Dat Mai
Abstract
This paper introduces StockGPT, an autoregressive ``number’’ model trained and tested on 70 million daily U.S.\ stock returns over nearly 100 years. Treating each return series as a sequence of tokens, StockGPT automatically learns the hidden patterns predictive of future returns via its attention mechanism. On a held-out test sample from 2001 to 2023, daily and monthly rebalanced long-short portfolios formed from StockGPT predictions yield strong performance. The StockGPT-based portfolios span momentum and long-/short-term reversals, eliminating the need for manually crafted price-based strategies, and yield highly significant alphas against leading stock market factors, suggesting a novel AI pricing effect. This highlights the immense promise of generative AI in surpassing human in making complex financial investment decisions.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 6.0/10 (capped from 9.0: transaction costs only partially modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 9.0)
Why this score: The paper employs advanced generative AI/transformer architectures with attention mechanisms (high math complexity) and demonstrates rigorous out-of-sample backtesting across 23 years with 70M data points, detailed portfolio construction, and statistical significance testing (high empirical rigor).
Research Flowchart
flowchart TD A["Research Goal: Can Generative AI<br>surpass human stock prediction?"] --> B["Methodology: Autoregressive<br>Attention-Based Model"] B --> C["Input: 70M Daily US Stock<br>Returns (1926-2023)"] C --> D["Process: Train StockGPT to<br>predict next return tokens"] D --> E["Output: Daily/Monthly<br>Long-Short Portfolios"] E --> F["Key Findings:<br>Strong Out-of-Sample Returns<br>Significant Alphas vs. Factors<br>Novel AI Pricing Effect"]
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 StockGPT, an autoregressive “number” model trained on 70 million daily U.S. stock returns to predict future returns.
Source sentence
This paper introduces StockGPT, an autoregressive “number” model trained and tested on 70 million daily U.S. stock returns over nearly 100 years.
Assumes 5
- The model is trained only once using data up to 2000 and then applied off‑the‑shelf to the out‑of‑sample period. model
Source sentence
Notably, the model is trained only once using the training sample and applied off-the-shelf to the out-of-sample period.
- Trading is done at the market close with no explicit slippage or market impact beyond removing small stocks. execution
Source sentence
To avoid trading only micro stocks since these stocks are illiquid and incur high transaction and market impact costs, before forming the portfolio, I remove stocks below the 10th percentile market value at the market close.
- Only common stocks with share codes 10 or 11 listed on NYSE, AMEX, and NASDAQ are included. data
Source sentence
I include only common stocks with a share code of 10 or 11 traded on three main exchanges NYSE, AMEX and NASDAQ.
- Returns are discretized into integer basis‑point intervals for the transformer input. model
Source sentence
Accordingly, I first convert returns into integer basis points by multiplying them by 10,000 and keeping the integer portion. Next, I cut the basis points into intervals of 50, closed on right.
Show 1 more
- The training sample (1926‑2000) is split into a parameter‑optimization period (1926‑1990) and a hyper‑parameter tuning period (1991‑2000). data
Source sentence
Within the training sample, data from 1926 to 1990 is used for parameter optimization and data from 1991 to 2000 for hyperparameter tuning and evaluation.
Data
- CRSP daily stock returns market-prices public
Source sentence
Stock return data comes from Center for Research in Security Prices (CRSP) that collects all historical U.S. stock returns from 1926 to 2023.
Costs
Partial
Stocks below the 10th percentile market value are removed to avoid high transaction and market impact costs; no explicit cost model is applied to headline returns.
Source sentence
To avoid trading only micro stocks since these stocks are illiquid and incur high transaction and market impact costs, before forming the portfolio, I remove stocks below the 10th percentile market value at the market close.
Finds 5
- The equal‑weighted daily rebalanced long‑short portfolio earns an average annualized return of 119% with a Sharpe ratio of 6.5 over 2001‑2023. performance
Source sentence
Under the equal-weighting scheme where each stock receives an equal weight in the portfolio, this daily rebalanced long-short portfolio earns an average annualized returns of 119% from 2001 to 2023, achieving a Sharpe ratio of 6.5.
- The value‑weighted version of the daily portfolio achieves an average annualized return of 27% and a Sharpe ratio of 1. performance
Source sentence
Under the value weighting scheme where stock weights in the portfolio are proportional to their market values, the StockGPT-based portfolio achieves an average annualized return of 27% and a Sharpe ratio of 1.
- Cross‑sectional Fama‑MacBeth regression shows an average slope of 0.5, implying a 100‑bp prediction difference signals a 50‑bp realized return difference; R² is 1.2%. predictive-accuracy
Source sentence
Accordingly, the average slope coefficient is 0.5, indicating that a cross‑sectional difference of 100 basis points (i.e, 1%) in StockGPT return predictions signals a difference of 50 basis points in realized returns. Moreover, the average cross‑sectional R2 is 1.2%.
- Monthly long‑term forecasts (20‑day horizon) yield an average slope of 3 (300‑bp realized return per 100‑bp prediction) and cross‑sectional R² of 0.55%. predictive-accuracy
Source sentence
the average slope coefficient is 3 (significant at 5%), indicating that a cross‑sectional difference of 100 basis points (i.e, 1%) in StockGPT return forecasts signals a difference of 300 basis points in realized returns. Moreover, the average cross‑sectional R2 is 0.55%.
- Monthly long‑short decile portfolios based on 20‑day forecasts outperform standard stock factors and generate highly significant alphas, though performance is less impressive than the daily model. performance
Source sentence
Overall, while the investment performance of the monthly StockGPT model is far less impressive than that of the daily model, it still outperforms all of the standard stock factors and yields highly significant alphas.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.
6 papers in the archive
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 (31)
- The social impact of generative ai: An analysis on ChatGPT (2023)
- Generative AI: Here to stay, but for good? (2023)
- Artificial intelligence in developing countries: The impact of generative artificial intelligence (AI) technologies for development (2023)
- Sentiment Trading with Large Language Models — → 2412.19245
- Expected returns and large language models (2022)
- Can ChatGPT forecast stock price movements? Return predictability and large language models — → 2304.07619
- Attention is all you need (2017)
- Language models are few-shot learners (2020)
and 23 more
- BERT: Pre-training of deep bidirectional transformers for language understanding (2018) — link
- OPT: Open pre-trained transformer language models (2022) — link
- BloombergGPT: A large language model for finance (2023) — link
- FinBERT: A Pretrained Language Model for Financial Communications (2020) — link
- Financial machine learning (2023)
- ClinicalGPT: large language models finetuned with diverse medical data and comprehensive evaluation (2023) — link
- Chronos: Learning the Language of Time Series (2024) — link
- Empirical asset pricing via machine learning (2020)
- (Re-) Imag (in) ing Price Trends (2023)
- Risk, return, and equilibrium: Empirical tests (1973)
- Hypothesis testing with efficient method of moments estimation (1987)
- A five-factor asset pricing model (2015)
- An augmented q-factor model with expected growth (2021)
- Does the stock market overreact? (1985)
- Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency (1993)
- Evidence of predictable behavior of security returns (1990)
- A new anomaly: The cross-sectional profitability of technical analysis (2013)
- Efficient capital markets (1970)
- Experimental evidence on the productivity effects of generative artificial intelligence (2023)
- Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (2023)
- Deep order flow imbalance: Extracting alpha at multiple horizons from the limit order book (2023)
- Investor sentiment and the cross-section of stock returns (2006)
- The uneven impact of generative AI on entrepreneurial performance (2023)
2 of 31 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.