Daily AI Trading Paper

Imagine you want to know whether an AI can read company news before the market does, and not just in a backtest that quietly cheats. That is the core problem in this paper. It matters because a huge amount of real trading research now leans on language models reading headlines, press releases, and other news faster than humans can. If that signal is real, it could be useful for short-horizon trading, for event-driven strategies, and for understanding how quickly prices incorporate information. But if the signal only works because the model has seen the future, memorized the company, or because the researcher tried enough variants until something looked good, then the apparent edge is fragile and not tradable. This paper sits right in the middle of a long finance literature on news and prices. We already know that markets react quickly to firm-specific news, often within minutes. We also know that news tone can predict short-horizon returns, and that some news-driven moves continue while others reverse. More recently, large language models have been used to read headlines and articles, and several studies have found predictive power. But finance has a recurring problem: the more flexible the model and the more ways you can slice the data, the easier it is to fool yourself. A strong backtest can reflect look-ahead bias, memorization of the sample, or selection among many trials rather than a genuine signal. The authors respond to that with a protocol they call OLIVIA. The basic idea is to make the test look much more like a real trading decision. First, the inputs are point-in-time, so the model only sees the news that was available at the decision moment. Second, the company identity is masked. They remove names, tickers, people, dates, and any mention of the price move, so the model cannot simply recognize the stock from the text. Third, the analysis plan is fixed in advance, and later changes are logged. Fourth, model selection is done only on validation data, not on the final test. Fifth, the test period is blinded until the plan is frozen, and then unblinded only once. In other words, the paper is trying to remove the usual ways a backtest can overstate performance. The model class they study is also important. They use what they call AI decision models, which are not generative chatbots in the usual sense. Instead of producing free-form text, these models take structured inputs and typed questions, then return probabilities over answers. That makes them attractive for finance because they can be used like probabilistic forecasters: you can ask whether a stock will rise or fall, or whether a reaction is too strong or too weak, and then score the probabilities directly. The main open-weights model in the study is fine-tuned on 2024 to 2025 data. They compare it with three commercial models, the untuned base model, and standard baselines. The data are the point-in-time members of the S&P 500 from July 2024 through July 2026. The sample is split into training, two validation quarters, and a blinded test period in 2026. The authors are careful about timing. For each company and trading session, they construct the news state as of 9:15 a.m. Eastern, before the open. Then they ask the model for a pre-open reading: how will the stock perform once the news is absorbed? They also ask a post-reaction question at 10:00 and after the close: was the move so far too small or too large for the news? That lets them separate the part of the return that happens before the open from the part that happens after the open. That distinction is crucial, because if news is already priced by the open, then a model that only predicts the overnight move may not be useful to a trader who can act at the open. The main result is that the market seems to get there first. The commercial models’ readings line up with the overnight move from the previous close to the open, with information coefficients around 0.074 to 0.077, but they do not line up with the move after the open, which is about zero. The untuned model shows the same split. So the models can read the news in a way that matches the price adjustment that happens before trading starts, but they do not predict the part of the move that remains available after the open. That is a very important trading interpretation: if the predictable component is already embedded in the opening price, then the edge is not accessible to someone who reads the news at the open. The fine-tuned open-weights model, Laya-Fin, looks promising in validation but weakens in the blinded test. In the final validation quarter, its daily information coefficient with the day’s move is 0.065, with a t-statistic of 3.6. But in the blinded January through July 2026 test, that falls to 0.020, with a t-statistic of 1.7. The paper emphasizes that every model with a clearly positive validation signal drops in the test. So the pre-registered signal does not survive honest out-of-sample evaluation. And when the authors translate that into trading terms, no pre-registered strategy survives costs or a deflated Sharpe ratio adjustment. In other words, once you account for transaction costs and for the number of trials behind the backtest, the apparent edge disappears. They also test whether the models can tell overreaction from underreaction. This is a subtle but important question. It is not enough to predict that a stock will move; a useful trading model would ideally say whether the market has overreacted and is likely to mean-revert, or underreacted and is likely to continue. The pre-registered regression does not support that. The news coefficient is not positive, and the moves the models label as underreactions do not later catch up. So the models are not successfully diagnosing the quality of the market’s first response. Another useful measurement result is about how much of a stock’s own daily move the news explains. Out of sample, same-day news explains only a very small fraction of variance, from 0.25% for Laya-Fin to 0.82% for one commercial model. That is not nothing, but it is modest. It also means that a lot of the daily move is still coming from other factors, broader market effects, or information not captured in the news state. The authors also note that the commercial models’ stated probabilities are too optimistic relative to actual accuracy, which is a calibration issue. In finance, that matters because a forecast is only useful if its confidence is meaningful. One of the most striking negative results is about masking. The company names are removed, but the models still identify the company surprisingly well. Given masked news, the commercial models pick the correct company out of four same-sector candidates in 62% to 74% of cases, compared with a 28% matched baseline. The open-weights model does worse, at 32%, but still above chance. So masking is not anonymity. The content, style, product details, management references, and other clues are often enough for the model to infer who the story is about. That is a key warning for anyone trying to use anonymized text as a clean test of generalization. What should we take away from all this? First, there is evidence that these models can read news in a way that is aligned with the immediate market reaction. But the reaction they are aligned with is mostly the one that happens before the open, meaning the market has already incorporated the news by the time a trader can respond in regular hours. Second, a signal that looks real in validation may not survive a truly blinded test once look-ahead, memorization, and selection are controlled. Third, the ability to identify the company from masked text is itself a major confound in language-model finance research. The limitations are also clear. The study is focused on S&P 500 names and on a specific 2024 to 2026 window. The models tested are decision models, not every possible language-model architecture. The results are strongest for the exact setup they define, and some exploratory findings do not hold up in the pre-registered test. The authors also mention a pre-registered forward test on news published after the models’ release, which is the right direction for future work because it further reduces the risk of contamination. So the bottom line is not that AI cannot read markets. It can. The stronger conclusion is that the market may already be reading the news first, and that once you force the evaluation to be genuinely out of sample, the tradable edge becomes much harder to find. This paper is valuable less because it finds a new profitable strategy, and more because it shows how to test these models in a way that is much closer to the real trading problem.