Image-Driven Stock Price Prediction with LLaMA: A Prompt-Based Approach

Abstract

This study proposes a novel framework for predicting stock price movements using a prompt-based approach with the LLaMA model, where candlestick charts serve as the primary input. Unlike traditional deep learning models that process images through convolutional or transformer-based architectures, the proposed method leverages LLaMA’s prompt-driven reasoning to interpret financial chart patterns. In addition, a teacher-student model incorporating LLaMA and Qwen is explored. To assess the effectiveness of this prompt-based LLM approach, its performance is compared with established models, including CNN, ResNet, and Vision Transformer. Experimental results demonstrate that the proposed method consistently outperforms these deep learning models, highlighting the potential of prompt-based LLM techniques for financial time series forecasting using visual inputs.