Machine confirming: Validating financial theories with transfer learning

Economists develop structural financial models to explain economic phenomena and provide interpretable causal inferences, and related theories are typically validated by formulating hypotheses and translating them into econometric models, often at the expense of predictive accuracy. In this paper, we introduce a transfer learning-based retroduction framework, called machine confirming, designed to validate theories using empirical observations in financial contexts. This standardized framework enables direct comparison between structural model outputs and real-world data. We apply it to a case study in derivative pricing to illustrate its effectiveness, demonstrating that incorporating theoretical structure improves predictive performance and that the theory is validated through the framework. Moreover, features that are statistically correlated but theoretically irrelevant are automatically excluded during the learning process.