Agent Trading Arena: A Study on Numerical Understanding in LLM-Based Agents

TL;DR

This paper introduces the Agent Trading Arena, a virtual trading environment for LLM-based agents that enables realistic market interactions and improves numerical reasoning and trading performance through visual inputs and reflection modules.

Contribution

We develop a realistic multi-agent trading platform for LLMs that simulates live markets and enhance their numerical reasoning with visual data and reflection modules.

Findings

Visualizations improve numerical reasoning and trading performance.

Reflection modules yield additional performance gains.

Our method outperforms baselines on NASDAQ and CSI datasets.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in natural language tasks, yet their performance in dynamic, real-world financial environments remains underexplored. Existing approaches are limited to historical backtesting, where trading actions cannot influence market prices and agents train only on static data. To address this limitation, we present the Agent Trading Arena, a virtual zero-sum stock market in which LLM-based agents engage in competitive multi-agent trading and directly impact price dynamics. By simulating realistic bid-ask interactions, our platform enables training in scenarios that closely mirror live markets, thereby narrowing the gap between training and evaluation. Experiments reveal that LLMs struggle with numerical reasoning when given plain-text data, often overfitting to local patterns and recent values. In contrast, chart-based visualizations significantly enhance both numerical reasoning and trading performance. Furthermore, incorporating a reflection module yields additional improvements, especially with visual inputs. Evaluations on NASDAQ and CSI datasets demonstrate the superiority of our method, particularly under high volatility. All code and data are available at https://github.com/wekjsdvnm/Agent-Trading-Arena.