Source: Levy (2024) — “Caution Ahead: Numerical Reasoning and Look-ahead Bias in AI Models”
The part that bothers me is how calm the language sounds. It makes the risk feel distant, when the real problem is already sitting inside today’s systems and decisions.
There’s a recurring pattern in financial AI research: a new model demonstrates extraordinary performance on a prediction task, the paper circulates widely, and practitioners get excited. Then more careful examination reveals that the performance partially reflects a flaw in the experimental design rather than genuine predictive ability. This paper is an important reality check on AI outperformance in accounting and finance. The author identifies two specific problems that affect a significant portion of published work. The first is poor numerical reasoning. LLMs, it turns out, are surprisingly bad at arithmetic and numerical reasoning in general. When financial prediction tasks require quantitative computation — not just pattern matching on qualitative text — LLMs make systematic errors. The implication is that applications which require actual numerical reasoning not just retrieving or characterising patterns should be treated with additional scepticism. The second, and more consequential, problem is look-ahead bias. Large language models are trained on data with a specific cutoff date, but that cutoff includes data about events that hadn’t yet been publicly known at earlier points in the training window. A model trained on data through December 2023 will have seen news, analysis, and commentary about events from 2020, 2021, and 2022 that was written after the fact — incorporating information that wouldn’t have been available to an investor in real time.
For leaders, the lesson is simple: if the risk timeline changes, the attention timeline has to change too. Waiting until everyone agrees it is urgent is usually how organisations arrive late.