LLM-Powered Multi-Agent Frameworks for Algorithmic Trading

A production-ready multi-agent system that combines LLM reasoning with Reinforcement Learning for auditable, high-performance algorithmic trading. A hierarchical pipeline of Analyst, Decision, Risk, and Execution agents processes multi-modal market data and communicates via structured messages to generate and execute trades with full XAI transparency.

---

Table of Contents

• Overview
• Architecture
• Repository Structure
• Quick Start
• Results
• Evaluation
• License

---

Overview

Feature	Description
Hierarchical Multi-Agent System	Orchestrated pipeline of Analyst, Decision, Risk, and Execution agents for sequential, auditable decision-making
Hybrid LLM + RL Intelligence	GPT-4 for high-level market interpretation combined with PPO/DQN for low-latency order execution
Multi-Modal Data Pipeline	OHLCV data, real-time news sentiment via FinBERT, and macroeconomic indicators via FRED
Explainable Trading	Natural language trade justifications with LLM Attention Visualization and SHAP values for RL policies
Backtesting Engine	Vectorized backtesting with slippage, transaction costs, Sharpe Ratio, MDD, and Bootstrap CI
Prompt Engineering Framework	YAML-based prompt registry with version control and A/B testing for agent instructions
Broker Integration	Abstracted broker layer supporting Alpaca and Interactive Brokers for paper and live trading

---

Architecture

The MultiAgentOrchestrator (trading/agents/orchestrator.py) coordinates five specialized agents.

Agent	Responsibility	Function
Analyst	Market Interpretation	Processes technical, sentiment, and macro data into a market insight report
Decision	Strategy Formulation	Consumes the Analyst report and formulates a BUY/SELL/HOLD action and position size
Risk	Constraint Validation	Validates proposed trades against max drawdown, position limits, and volatility rules
Execution	Trade Implementation	Executes approved trades using RL-optimized order placement to minimize market impact
Explainability	Audit Trail	Logs agent communication flow and generates natural language trade justifications

---

Repository Structure

Path	Description
trading/agents/	Orchestrator and agent communication logic
trading/backtest/	Vectorized backtesting engine
trading/brokers/	Alpaca and IBKR broker connectors
trading/costs/	Transaction cost modeling (fixed, variable, market impact)
trading/data/	OHLCV, news, and macro data ingestion and feature engineering
trading/explainability/	LLM attention visualization and RL policy SHAP interpretation
trading/models/	Unified LLM wrapper (GPT-4, Claude, etc.)
trading/prompts/	Version-controlled YAML prompt registry per agent
trading/risk/	Portfolio-level VaR and drawdown control
trading/rl/	Gymnasium trading environment and PPO/DQN trainer
trading/strategies/	Quantitative baselines (Momentum, Mean-Reversion, Pairs)
trading/reporting/	Trade reports and performance metrics
figures/	Generated equity curves and XAI visualizations
results/	Experiment metrics, trade logs, and model checkpoints
tests/	Unit and integration tests

---

Quick Start

Docker (Recommended)

git clone https://github.com/quantsingularity/LLM-Powered-Multi-Agent-Frameworks-for-Algorithmic-Trading.git
cd LLM-Powered-Multi-Agent-Frameworks-for-Algorithmic-Trading

export OPENAI_API_KEY="sk-..."

docker-compose build && docker-compose up -d
docker-compose run llm-trading python trading/run_experiment.py

Local Installation

python3.10 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

python trading/run_experiment.py

Results are saved to results/ and figures to figures/.

---

Output

Console output stays readable: benign third-party warnings and library progress bars are suppressed, so only meaningful log lines remain. Each run finishes with a clean, aligned summary block reporting per-ticker total return, Sharpe ratio, max drawdown and final portfolio value.

Results

Pilot experiment on AAPL comparing the hybrid system against traditional and single-model baselines.

Metric	Buy and Hold	RL Only (PPO)	LLM Only	Hybrid LLM + RL
Total Return	-2.15%	-5.23%	-4.89%	-4.04%
Sharpe Ratio	-1.12	-2.45	-2.10	-1.80
Max Drawdown	6.21%	7.84%	5.92%	4.66%
Win Rate	N/A	38.2%	40.5%	42.4%
Agent Decision Time	N/A	N/A	N/A	2.1s (+-0.4s)

---

Evaluation

Component	Command
Unit tests	pytest tests/test_simple.py
Integration tests	pytest tests/test_integration.py
Quantitative baselines	trading/strategies/baselines.py
Ablation studies	trading/prompts/experiments/
Trade reports	trading/reporting/trade_report.py

---

License

Licensed under the MIT License. See LICENSE for details.