Beyond Black-Box AI:  A Theory Of Interpretable Transformers For Asset Pricing

Abstract

Artificial intelligence (AI) is transforming asset pricing by uncovering complex return patterns that traditional models struggle to capture. Transformer-based Artificial Intelligence Pricing Models (AIPMs) have shown empirical promise, yet their complexity raises concerns about economic interpretability. This paper develops a general theoretical framework that quantifies the trade-off between predictive power and transparency in AIPMs. I introduce an attention-based interpretability metric and derive a lower bound on its trade-off with Rademacher complexity, demonstrating that transformer-based models can be both powerful and economically meaningful. Furthermore, I establish that interpretable transformers universally approximate pricing kernels, offering a robust foundation for AI-driven asset pricing. By linking attention mechanisms to classic portfolio theory, I show how linear attention extends Markowitz optimization, while nonlinear attention enhances cross-asset dependencies. Monte Carlo simulations confirm the framework’s practical value, linking AI innovation to financial theory for scalable, interpretable pricing.

Biography

Dr Fallahgoul is a Senior Lecturer at the School of Mathematics at Monash University. He has a strong background in finance and economics, having completed postdoctoral positions at both the Swiss Finance Institute at EPFL and the European Center for Advanced Research in Economics and Statistics (ECARES) at the Free University of Brussels.

He holds a PhD in Applied Probability and Statistics and has over two years of experience in risk management in financial indutries. Dr Fallahgoul's research interests are focused on financial econometrics, asset allocation, and the application of AI/ML in economic and finance. He also serves on the editorial board of the Journal of Financial Data Science.