The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence, pp. 39-46, Mar. 21, 2026
The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence (SIG-FIN)
We propose a construction method for an artificial market model using LLM-based agents for evaluating the impact of trading systems in financial markets. In conventional rule-based agent artificial markets, investors' decision rules and their distributions must be manually designed by researchers, which entails high design costs and arbitrariness. In this work, we propose pseudocode-driven agents (LBA-P) that pre-generate and store LLM-generated investment strategies as pseudocode, and determine orders by executing the pseudocode. This enables semi-automatic introduction of diverse strategy distributions mimicking real investor distributions while retaining strategies as text, and allows post-hoc strategy analysis. Experimental results show that the pseudocode condition satisfies the stylized facts of the market, while using Python code for strategy representation failed to satisfy some indicators. Post-hoc analysis of strategy logic reveals that differences in output format produce systematic biases in strategy distributions, clarifying that the choice of strategy representation is critical in LLM-based semi-automatic market modeling.
Artificial Market; LLM; Pseudocode; Agent;
10.11517/jsaisigtwo.2026.FIN-036_39
@inproceedings{Itakura2026-sigfin36,
title={{Artificial Market Model using LLM-based Agent Driven by Pseudocode [in Japanese]}},
author={Ryoma Itakura and Masanori Hirano and Kentaro Imajo and Hiroki Sakaji and Itsuki Noda},
booktitle={The 36th meeting of Special Interest Group on Financial Informatics of Japanese Society for Artificial Intelligence},
pages={39-46},
doi={10.11517/jsaisigtwo.2026.FIN-036_39},
year={2026}
}