Financial AI needs more than an accurate answer.
Reproducing a result is a different question from getting a good score.
The research question
Why can a financial-AI workflow give results that are difficult to reproduce? The authors examine tabular models, graph networks and language-model workflows.
Method and reported findings
The survey combines literature analysis with experiments on public financial datasets. It describes instability in explanations, changing graph-model predictions and differing language-model outputs under different execution settings.
The markets connection
Our reading: evaluate whether a result can be reproduced alongside its predictive quality. Record the data, software and execution conditions. A high benchmark score alone does not answer an audit question.
Keep it in context
Preprint, version 4. These are the authors’ reported results; Market Mechanics has not independently reproduced the experiments. First submission and revision dates are shown separately.
From Accuracy to Auditability: A Survey of Determinism in Financial AI Systems Zhou et al. · arXiv · 11 May 2026 · revised 26 AugustArticle history
- 18 September 2026 — First web edition. Pinned to version 4; added the revision date.