Volume-Enhanced Stock Price Forecasting: Multivariate Approach with Large Language Models

Document Type : Original Article

Authors
Sajad Mirzababaei
Moslem Habibi
Sharif University of Technology
10.22034/jcse.2025.225717
Abstract
Accurate forecasting of stock prices and trading volumes underpins informed investment decisions and robust risk management. We present GPTTS, a transformer-based framework that integrates a specialized UnifiedDiffTokenizer for return-based discretization, parallel T5-derived price and volume encoders with lightweight adaptation (LoRA and adapters), and a gated cross-attention fusion module for cross-modal reasoning. On a comprehensive 15-year NASDAQ dataset, GPTTS delivers improvements of approximately 3% in MAE and 25% in MAPE over the strongest classical and deep-learning baselines, while reducing trainable parameters. These gains demonstrate the value of combining financial-aware tokenization with efficient fine-tuning to achieve high-fidelity, resource-conscious forecasting for algorithmic trading and risk analytics.

Keywords