Reproduce “ARCADIA: Scalable Causal Discovery for Corporate Bankruptcy Analysis Using Agentic AI” on OpenResearch

About this paper

Fabrizio Maturo, Donato Riccio, Andrea Mazzitelli, Giuseppe Bifulco, Francesco Paolone, Iulia Brezeanu

ARCADIA introduces a framework leveraging agentic large language models and statistical validation to discover causal directed acyclic graphs for corporate bankruptcy analysis. This approach consistently generates structurally valid and interpretable causal models, addressing the limitations of purely correlational predictive systems in finance.

Key insights

• The iterative, agentic approach, combining LLM reasoning with rigorous statistical and structural validation, enables the discovery of causal graphs that are 'born valid' and require minimal post-hoc pruning.
• Leveraging the LLM's domain knowledge allows the framework to propose theory-grounded proxy confounders and address unobserved confounding, a common challenge in complex financial datasets.
• ARCADIA's design prioritizes theory-driven causal discovery, ensuring that the discovered structures are not merely statistically optimal but also theoretically coherent and justifiable within financial economics.

Summary by alphaXiv.