This repository contains our solution to Task 1 of the FinRL Contest 2025: developing a high-performance crypto trading agent using reinforcement learning and LLM-derived market signals.

LARSA: LLM-Augmented Regime-Switching Agent

DISCLAIMER: LARSA is experimental research software. It is NOT financial advice. Crypto trading carries substantial risk of loss. Never deploy real capital without fully understanding the system, the risks, and the applicable laws in your jurisdiction. Always start with dry-run or paper trading mode. Past simulated performance does not guarantee future results.

What is LARSA?

LARSA (LLM-Augmented Regime-Switching Agent) is a hybrid reinforcement learning and large language model trading system for crypto assets. It extracts structured sentiment and risk signals from BTC news via the DeepSeek V3 API, mines predictive factors with a recurrent neural network trained on Alpha101 features, trains an ensemble of three DQN-family agents (D3QN, DoubleDQN, TwinD3QN) that vote on trade decisions, and dynamically shifts ensemble weights based on a detected market regime (bull, bear, sideways, volatile). The system supports single-asset BTC trading, multi-asset crypto portfolios, live paper trading on Binance price feeds, production-safe live trading via the Live Trading Bridge, per-trade explainability reports, and automated hyperparameter search.

Round 2 additions extend LARSA with a multi-agent LLM debate layer, alternative data feeds, a continuous online learning pipeline, and advanced portfolio optimisation strategies —