FinAI Contest Submission – LARSA: LLM-Augmented Regime-Switching Agent

Overview

This submission introduces LARSA, a hybrid RL + LLM-based trading system that uses real-time sentiment and risk signals extracted from Bitcoin news to adaptively switch trading regimes.

Our agent combines:

• LLM-Engineered Signals (via DeepSeek V3)
• Recurrent Feature Aggregation (Alpha101 + RNN forecasting)
• DQN-Based Trading Agents
• Ensemble Coordination with signal-weighted decision logic

All components are built to respect the evaluation environment and generalize across unseen data.

---

Key Innovations

• Custom Sentiment & Risk Signal Extraction: We use DeepSeek V3 to process BTC news headlines and article bodies into structured JSON signals, including confidence tracking and fallback handling.
• Confidence Filtering & Checkpointing: Sentiment extraction is fault-tolerant with automatic retries and logging. Only high-confidence samples are fed into agents.
• Recurrent Factor Miner: RNN model trained on Alpha101 + news sentiment to generate strong predictive features.
• Modular Ensemble Design: Each agent votes with its own signal alignment; ensemble strategy adapts based on regime (volatility, trend).
• Production-Ready Signal Pipeline: Fully restartable with checkpointing, CLI-configurable, and compatible with unseen data at test time.

---

Folder Structure

trained\_models/               # .pth files from agent training
factor\_mining/
└── deepseek\_signals.py     # API signal extractor
└── seq\_run.py              # trains RNN on factors
task1\_ensemble.py             # combines model outputs
task1\_eval.py                 # runs backtest on test data
trade\_simulator.py            # only changed if modified
requirements.txt              # pip packages

---

Dependencies

# Setup virtual environment (optional but recommended)
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows

# Install dependencies
pip install -r requirements.txt

---

Running the System

1. Extract News Signals

python factor_mining/deepseek_signals.py --input ./data/news_train.csv --output ./data/news_with_signals.csv

2. Generate Factors

python factor_mining/seq_run.py

3. Train Agent Ensemble

python task1_ensemble.py

4. Evaluate on Unseen Data

python task1_eval.py

---

Notes on Environment

• We did not modify transaction costs, reward functions, or market structure in trade_simulator.py.
• However, we support regime-switching mechanisms via signals passed to the ensemble.
• If you'd like to re-run with modifications, please update the args.input and args.output paths in deepseek_signals.py.

---

Metrics on Validation Set

---

Contact

Author: Matthew C. Busel Email: mattbusel@gmail.com GitHub: github.com/mattbusel

---

---