Stock Price Prediction Using Univariate and Multivariate Historical Data with Post-Interpretation via Large Language Models

Abstract

In this study, we propose a hybrid approach that utilizes both univariate and multivariate historical data from key variables and related factors across four distinct groups. We developed a hybrid approach combining Volume Features, Valuation Metrics, Technical Indicators, and Market Sentiment for stock price prediction and investigate several state-of-the-art models, including Artificial Neural Networks (ANN), Gated Recurrent Units (GRU), Bidirectional GRU (BI-GRU), and Transformer-based Time Series (TST) models, while experimenting with different lags of inputs to capture intricate temporal patterns in stock price movements. Our experiments, conducted on seven stocks from various sectors, allow us to evaluate the robustness and generalizability of the models across different industries. To enhance interpretability, we employ large language models (LLMs) in the post-prediction phase, which transform the predictive outputs into human-readable narratives explaining the factors driving stock price predictions. Empirical results demonstrate that our approach, incorporating advanced deep learning models like ANN, GRU, BI-GRU, and TST with varying input lags, significantly improves prediction accuracy over traditional methods while providing actionable insights for financial decision-making.