01Project
Multi-Agent Crypto Predictor.
An autonomous multi-agent system utilizing LLMs to predict cryptocurrency market movements by synthesizing Reddit sentiment and technical analysis.
Back to work
Overview
This project orchestrates a multi-agent system (MAS) using a Directed Acyclic Graph (DAG) architecture. Specialized agents—including sentiment analysts (with data from Reddit), technical analysts, adversarial strategists, and an executive trader—collaborate to evaluate market conditions for BTC, ETH, and SOL in isolated time windows.
Challenge
Deploying Large Language Models in financial markets traditionally risks severe hallucinations and contextual degradation (the 'Telephone Effect') as information passes between prompts. The system also needed to avoid cognitive biases, such as the sunk cost fallacy, from past trading decisions.
Approach
I replaced linear prompting with a shared state machine, enforcing deterministic, machine-readable JSON data contracts via Pydantic. To eliminate cognitive bias, I implemented a 'Tabula Rasa' re-entry methodology for the executive layer. The system also featured adversarial debate logic and a Strategy Reflection Agent for performance evaluation.
Outcome
Backtesting on the e-INFRA CZ HPC grid showed a +10.50% return with Qwen 3.5, compared with -37.53% for buy and hold over the same period.