MULTI-TASK STOCK DIRECTION AND PRICE PREDICTION WITH RULE-BASED SENTIMENT AND TRANSFORMER ARCHITECTURE

Abstract

Accurate stock price forecasting remains challenging due to market volatility and complex interdependencies. While Large Language Models (LLMs) show promise for sentiment extraction, API costs of Rs.2.50-10 per 1000 tokens and 3-8 second latencies render them impractical for real-time trading. This work introduces a unified multi-task Trans- former framework that concurrently predicts directional movement and closing prices through joint optimization, incorporating a zero-cost rule-based sentiment module as an LLM alternative. Evaluation across seven equities (NVIDIA, Apple, Alibaba, Adobe, Caterpillar, AbbVie, AMD) spanning January 2015-January 2025 demonstrates 74.82% directional accuracy and 1.88% RMSE—achieving competitive performance with state-of-the-art methods while eliminating operational costs and maintaining 78ms response times.