Can AI predict the stock market?

Only weakly, and only on average. Machine-learning models can rank stocks a little better than chance across many names, but in a widely cited study of US stocks the best models explained well under 1% of the monthly variation in individual stock returns (Gu, Kelly and Xiu, 2020). A small edge spread across thousands of stocks can still matter. No model knows whether a particular stock will rise, and published patterns tend to weaken: McLean and Pontiff (2016) found anomaly returns about 58% lower after publication.

What is the difference between an AI stock picker and quantitative research?

An AI stock picker delivers a list of names, often with little on how the method was tested. Rigorous quantitative research states the question, the data, the method and the result on data the model did not see, and reports the ideas that failed as well as those that worked. Useful checks: how many variations were tried, whether the data were point-in-time, whether the forecasts carry a stated uncertainty, and how they fared once they matured.

Can ChatGPT or other generative AI pick stocks?

It can help read and summarise text such as filings and earnings calls, and one study found that ChatGPT's readings of news headlines were related to next-day stock returns (Lopez-Lira and Tang, 2023). But a language model can state wrong figures with confidence, and testing it on the past is hard: it may have read about the period being tested during training, which leaks the future into the test. Its outputs need the same out-of-sample testing as any other signal.

Does Opulence Alpha use generative AI?

No. The platform calls no ChatGPT, Gemini or Claude service, and no language model writes, summarises or decides anything in its research. The only language model in the system, FinBERT, is a classifier: it labels the tone of SEC filing text as positive, negative or neutral, and those scores are one input among many to the forecasting models.

FAQ · AI investment research

What is AI investment research, and what are its limits?

Reviewed against the platform's code on Sep 26, 2026

AI investment research uses machine learning to analyse markets. Predictive models learn from prices, filings and economic data to estimate quantities such as a stock's future return or risk; generative models such as large language models read and summarise text. Neither knows the future: done properly, every estimate is tested on data the model never saw.

Why it matters

Stock returns are mostly noise, and the patterns in them change. A flexible model will find structure in almost any dataset, including structure that is pure chance, so an impressive backtest from an AI model is weak evidence on its own. Markets are also non-stationary: relations shift as rates, rules and participants change, and a pattern that becomes widely known tends to be traded away (McLean and Pontiff, 2016). Generative models add their own risks: they can state wrong figures fluently, and a backtest of one can leak the future, because the model may have read about the test period during training. What turns a model's output into evidence is a test on data it never saw.

How it works

Predictive research turns a question into a quantity to estimate, such as each stock's return over the next month relative to other stocks. Models such as gradient-boosted trees and neural networks learn from inputs built only from data public at the time, and are scored on later periods they did not train on, by the information coefficient or interval coverage. In a study of US stocks over six decades, trees and neural networks forecast returns better than linear models, largely by capturing interactions between predictors, yet the best explained well under 1% of the monthly variation in individual returns (Gu, Kelly and Xiu, 2020). Generative AI is used mainly to read text, and its outputs need the same testing.

How Opulence Alpha applies it

Opulence Alpha uses predictive machine learning, not generative AI. Its return forecasts for 1, 5, 21 and 63 trading days draw on gradient-boosted decision trees (CatBoost), LSTM neural networks (PyTorch) and XGBoost quantile models, whose ranges are calibrated by conformalized quantile regression. Bernstein Online Aggregation combines their forecasts, shifting weight each trading day toward the models whose matured forecasts ranked stocks better. A model that fails a statistical gate on held-out data is still scored every day but does not shape the published forecasts. FinBERT scores the tone of SEC filings as one input and writes nothing. The forecasts are model estimates with a stated range, not recommendations to buy or sell.

Which models the platform uses, and which it does not →

Related concepts

Backtest overfittingBacktest overfitting is when a strategy is tuned to the noise of past data, so its backtest looks strong but its live performance decays. How to detect and avoid it.Out-of-sample testingOut-of-sample testing judges a strategy on data it never saw while being built: in-sample vs out-of-sample, holdouts, walk-forward and the reuse trap.Ensemble forecastingEnsemble forecasting combines several models' forecasts into one. Why a combination beats a typical single model, and how online aggregation sets the weights.Strategy decayStrategy decay (alpha decay) is the loss of a trading edge over time through overfitting, crowding and regime change. How to detect it, and how we watch for it.

All concepts →Sector behaviour in every market regime →

Questions

Can AI predict the stock market?: Only weakly, and only on average. Machine-learning models can rank stocks a little better than chance across many names, but in a widely cited study of US stocks the best models explained well under 1% of the monthly variation in individual stock returns (Gu, Kelly and Xiu, 2020). A small edge spread across thousands of stocks can still matter. No model knows whether a particular stock will rise, and published patterns tend to weaken: McLean and Pontiff (2016) found anomaly returns about 58% lower after publication.
What is the difference between an AI stock picker and quantitative research?: An AI stock picker delivers a list of names, often with little on how the method was tested. Rigorous quantitative research states the question, the data, the method and the result on data the model did not see, and reports the ideas that failed as well as those that worked. Useful checks: how many variations were tried, whether the data were point-in-time, whether the forecasts carry a stated uncertainty, and how they fared once they matured.
Can ChatGPT or other generative AI pick stocks?: It can help read and summarise text such as filings and earnings calls, and one study found that ChatGPT's readings of news headlines were related to next-day stock returns (Lopez-Lira and Tang, 2023). But a language model can state wrong figures with confidence, and testing it on the past is hard: it may have read about the period being tested during training, which leaks the future into the test. Its outputs need the same out-of-sample testing as any other signal.
Does Opulence Alpha use generative AI?: No. The platform calls no ChatGPT, Gemini or Claude service, and no language model writes, summarises or decides anything in its research. The only language model in the system, FinBERT, is a classifier: it labels the tone of SEC filing text as positive, negative or neutral, and those scores are one input among many to the forecasting models.

References

Gu, S., Kelly, B. & Xiu, D. (2020). Empirical Asset Pricing via Machine Learning. Review of Financial Studies, 33(5), 2223–2273.
McLean, R. D. & Pontiff, J. (2016). Does Academic Research Destroy Stock Return Predictability? Journal of Finance, 71(1), 5–32.
Lopez-Lira, A. & Tang, Y. (2023). Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models. Working paper, arXiv:2304.07619.
Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv:1908.10063.

Educational content about research methods. Not investment advice.