AI news readers found stock-market signals—but little left to trade at the open

A model can connect good news with a rising stock and still offer a trader little advantage. The crucial question is whether the price has already moved before the trader enters. A new arXiv preprint, Thinking Fast, Trading Late: System One Decision Models vs the Market, reports that this timing distinction substantially changes the interpretation of AI news signals.

For the models and large US stocks examined, the researchers found no statistically significant ability to predict returns from the opening price to the close. Commercial models did show an association with overnight returns, but those gains were already incorporated into the opening price. The findings concern this particular historical evaluation, not a demonstration that AI can never produce a trading advantage.

Testing more than a promising backtest

The study examines AI decision models: non-generative language models that return probabilities for specified answers rather than writing free-form responses. Its primary model, Laya-Fin, was fine-tuned on news and returns from July 2024 through June 2025. Comparators included three commercial decision models, untuned Laya and conventional baselines.

The researchers used historical S&P 500 membership, two validation quarters in the second half of 2025, and a blinded January–July 2026 test period. The pre-open evaluation contained 13,748 company-news events across 140 sessions.

Their proposed OLIVIA protocol combines timestamp-controlled inputs, identity masking, validation-only selection, a cryptographically frozen analysis plan, logged test-data access and a ledger of attempted variants. The plan was registered in October 2026, after the historical test period had ended, but before the study’s panels, targets or model outputs were created, according to the paper. Test outcomes were subsequently released to the frozen analyses through a logged lock.

At 09:15 Eastern Time, models received news available since the previous close, alongside structured context. Names, tickers, dates and price-move references were masked. Separate readings at 10:00 and after the close assessed the market’s reaction.

The opening price changes the story

The main measure was the daily information coefficient, or IC: a rank correlation between model readings and stock moves across that day’s events, averaged over sessions. It is not a percentage return or a prediction-success rate.

Laya-Fin’s IC with close-to-close returns was 0.055 and 0.065 in the two validation quarters. On the blinded test it fell to 0.020, with a t-statistic of 1.67. None of the five primary hypotheses was supported for Laya-Fin after the specified multiple-testing correction.

The commercial models retained overnight ICs between 0.074 and 0.077, while their open-to-close ICs ranged from −0.001 to 0.002. Untuned Laya showed a similar overnight-versus-intraday split.

There is an important timing qualification: the overnight return runs from the previous close to the opening price, including trading before the 09:15 reading. The analysis does not separate movement before and after that reading. Consequently, the overnight association partly measures alignment with an already unfolding price move, rather than a wholly prospective forecast.

The models also failed to establish that they could distinguish overreaction from underreaction and predict the subsequent correction. Commercial forecasts were overconfident: their stated probabilities exceeded their observed accuracy.

Trading costs and hidden identities

No pre-registered strategy survived both costs and the deflated Sharpe assessment, which adjusts performance for the many trials behind a backtest. At costs of 10 basis points per side, charged on stock positions and the market hedge, the main strategies lost money. A sparse strategy trading model-labelled misreactions remained positive, but failed the selection-adjusted hurdle. Borrow costs and market impact were not modeled.

Masking also provided less protection than removing names might suggest. In a four-candidate identification exercise, commercial models recognized companies from masked news in 62%–74% of cases, versus a matched baseline of 28%. Untuned Laya scored 32%.

Recognition does not prove that a model remembers subsequent returns. Nevertheless, every commercial candidate failed the pre-registered recognition screen. The researchers retained three as explicitly screen-failing comparators, not clean evidence against contamination.

What remains unresolved

All decision models were released after the historical sample ended, so exposure to evaluation-period material cannot be ruled out. The intended anchor model produced nearly constant directional outputs, preventing conclusions about what fine-tuning adds. Training used one seed and one epoch, and some planned robustness analyses remained incomplete.

The sample covers highly followed large companies over only seven test months. The authors also disclose earlier exposure to overlapping aggregate news reactions, and limited assistant analyses that shaped exploratory work. Raw news, prices and membership data are private, although code and reported outputs are available.

A pre-registered forward evaluation using news published after the models’ releases had not been run when the paper was written. Until then, the strongest practical lesson is methodological: an association with a day’s return is not necessarily a forecast a trader could have acted on.