MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning

Researchers have introduced MoCA-Agent, a code agent that uses a market-based system of atomic claims to improve the accuracy and robustness of financial and numerical question answering, achieving strong results across multiple public benchmarks [1]. The system, detailed in a paper by Abdelrahman E.M. Abdallah and colleagues, departs from standard multi-agent debate by decomposing questions into typed, verifiable atomic claims [1]. Specialist trader agents then buy or sell these claims, and their orders are cleared into confidence-weighted accept or reject decisions [2]. From this market-supported evidence, the system synthesizes an executable Python program to derive the final answer [1]. A code-aware verifier subsequently checks the program for execution errors, structural consistency, and common financial reasoning mistakes, with at most one market-aware repair round permitted [3]. This process makes the assumptions behind the program explicit, reducing the risk that a mis-extracted cell, incorrect formula, or scale error silently propagates through the pipeline [4]. MoCA-Agent was evaluated using a fixed Qwen3.6-27B backbone across ten public benchmarks [1]. It achieved 78.3% on FinQA, 76.0% on FinanceMath, 71.2% on MultiHiertt, 86.9% on ESGenius, and an 85.6% average on FinChart-Bench [2]. The benchmarks span financial numerical reasoning, general tabular reasoning, ESG question answering, and multimodal chart reasoning [5]. The paper argues that aggregating evidence at the level of atomic claims, rather than whole answers, improves robustness in high-stakes numerical reasoning [1]. The code and data for the project have been made publicly available [2].