Coverage-valid uncertainty for numeric answers from financial language models.
When a language model produces a number, the number alone is not enough: you also need to know how confident the model is, in a form that can be measured and audited. Conformal-Finance is a research framework that wraps a model's numeric output with a coverage-valid prediction interval whose guarantee holds without distributional assumptions on the data or the model.
The EU AI Act (Article 15) requires high-risk AI systems to demonstrate "appropriate levels of accuracy, robustness and cybersecurity" with measurable uncertainty controls, with obligations phasing in from 2026. Conformal prediction is a natural primitive for that uncertainty layer. This project explores it for numeric financial question answering.
Research preview. A technical paper preprint is in preparation; a link will be added here on release.
To be notified when the preprint and code are released, watch this repository (the Watch button above) or open an issue.
Built by an independent researcher.