ForexAgent: Identifying Trading Strategies in Forex Markets with Large Language Models

Abstract

Identifying traders’ underlying strategies from account level logs is vital for market surveillance and risk control, yet threshold rules and conventional learners often fail when behaviors deviate from textbook patterns or combine multiple styles. We introduce ForexAgent, a large language model (LLM) framework that classifies strategies such as Martingale and Kelly directly from trade sequences. ForexAgent couples a realistic data generation pipeline, which replays canonical strategies on historical foreign exchange prices to produce labeled, account level records, with an LLM classifier that operates either on raw executions (timestamps, direction, size, and profit or loss) or on concise, behavior centric summaries (risk–return profiles, equity linked sizing, and outcome responsive scaling). On realistic forex datasets, ForexAgent attains macro F1 up to 94.6%, surpassing rule based and heuristic baselines, and remains data efficient in zero and few shot settings. Ablation studies show that focused, compact prompts, or even raw logs alone, often outperform feature heavy descriptions, suggesting attention dilution from verbose inputs. By recognizing pattern regularities (for example, loss responsive size escalation versus equity proportional growth) rather than brittle thresholds, ForexAgent generalizes to noisy, variant, and mixed behaviors (for example, Martingale scalers in the range of 1.8× to 2.3× with occasional rule breaks). The framework offers a practical path for brokers and regulators to flag high risk styles without strategy specific retraining, and its prompt based architecture provides interpretable rationales as well as a foundation for extension to broader markets and hybrid style taxonomies.