Explainable zero-shot trading using multi-agent LLM architecture: A backtested approach for Bitcoin price

Abstract

This study introduces a zero-shot, reasoning-based multi-agent trading framework utilizing large language models (LLMs) to integrate heterogeneous signals for Bitcoin trading over a 1400-day period. The framework combines specialized agents, each dedicated to a modality such as technical indicators, on-chain metrics, macroeconomic signals, and textual sentiment, with a metaagent that synthesizes their rationales into coherent trading decisions without task-specific fine-tuning. Empirical evaluations using Bitcoin market data reveal that the proposed framework outperforms conventional time-series models over a short-horizon (three-day) period, achieving a 21.75 % total return (29.30 % annualized) and a Sharpe ratio of 1.08, surpassing the Long Short-Term Memory (LSTM) baseline by 1.70 percentage points in total return and 0.003 in Sharpe ratio. Ablation results reveal that Reddit-based sentiment enhances profitability (23.30 % total return), while news-based sentiment introduces semantic noise that degrades performance. All strategies are rigorously evaluated under realistic backtesting conditions, explicitly considering slippage and transaction costs to ensure reproducibility and fair comparisons. Beyond raw returns, systematic evaluation through an LLM-based evaluation protocol (G-EVAL) validates the consistency and interpretability of agent rationales, reinforcing model transparency. The proposed framework's modularity, interpretability, and robust empirical performance highlight its potential as an interpretable, scalable, and transparent approach to financial decision-making, aligning with the broader goals of explainable artificial intelligence in risk-sensitive financial systems.