LLM-Assisted Gold Trading: Evaluating Reasoning-Based Strategies Against Technical Indicator Baselines

Sammanfattning

The rapid development of Large Language Models (LLMs) has opened new possibilities for applying AI-driven reasoning to financial decision-making. While LLMs have already shown promising results in general financial prediction tasks, limited research has evaluated their effectiveness in trading for specific commodity markets such as gold, especially when measured against established benchmarks. This thesis investigates whether LLM-assisted trading strategies, developed using prompt engineering and a multi-agent architecture, can outperform technical indicator-based strategies when applied to gold market data, and how different temperature settings influence the performance of the LLM-assisted strategy. To address this, a backtesting pipeline is implemented using Python that evaluates the different trading strategies under identical market conditions, primarily focusing on historical gold price data obtained via Yahoo Finance. The indicator-based strategies are constructed using technical indicators, including RSI, MACD, and Bollinger Bands. The LLM-assisted strategies implement a multi-agent architecture where specialized agents handle different roles, such as market data collection, risk assessment, and final trade decisions. The agents are supported by prompt engineering techniques for optimizing the generated output. For the LLM-assisted strategy, the temperature parameter is varied across predefined settings in order to analyze how changes in output randomness affect trading performance. Both strategies are evaluated using standard trading performance metrics, including win rate, profit factor, maximum drawdown, and the Sharpe ratio. The findings show that every LLM-based strategy outperforms every technical indicator baseline on total return, with LLM returns ranging from $33.1\%$ to $61.2\%$ compared to the highest baseline return of $17.5\%$ achieved by MACD. All LLM-based strategies achieve a Sharpe ratio above $1.0$, while no baseline strategy exceeds this threshold. Among the evaluated temperature configurations, $t = 0.5$ yielded the best overall performance across most metrics. These findings suggest that LLM-assisted trading strategies can outperform single traditional indicator-based strategies on historical gold market data. Temperature was found to be a meaningful parameter for the LLM-assisted approach, with a moderate value of $0.5$ topping the podium across most performance metrics. However, due to this study being limited to a single asset and a relatively short evaluation period, the results should be interpreted with caution.