Using Large Language Models to Predict Crude Oil Price Based on Financial Sentiment Analysis

Financial sentiment analysis is essential for interpreting market dynamics and making informed strategic trading decisions. Although recent years have seen the adoption of advanced deep learning architectures and financial language models, this study advances the field by exploring the capability of large language model, specifically Gemini 2.5 pro offered by Google, for financial sentiment prediction, with a particular focus on the crude oil market. By monitoring news headlines and news content for the year 2025 and using a few-shot prompt technique, we evaluate multiple Gemini 2.5 prompt formulations on a carefully curated dataset of crude oil news headlines and text. The resulting prediction model is evaluated using the mean absolute error and the mean square error of the sentiment labels and the prediction prices. Incorporating news based textual information significantly improved forecasting performance. The RMSE decreased from 1.57 when using only historical price data to 1.14 when combining price information with news headlines and full article text through few-shot prompting. Similarly, the MAE improved from 1.32 to 0.78 under the same enhanced input configuration, demonstrating that incorporating news information substantially reduces the prediction error. These results suggest that the sentiments derived from the news provide valuable incremental information for short-term crude oil price prediction.