Dual-Task Financial Forecasting with HSAT: Integrating Market Trends and Closing Price Prediction

پوختە

Precise stock prediction is very important in risk management and investment optimization under volatile financial markets. Recent advances in deep learning provide new avenues for modeling such time series data. This paper introduces HSAT (Hybrid Stock Attention Transformer), a unified transformer-based architecture designed for two core financial forecasting tasks: trend direction classification (up/down) and closing price regression. Unlike traditional models that rely solely on raw historical price data, HSAT integrates a comprehensive set of engineered technical indicators—including Relative Strength Index (RSI), Moving Average (MA), Moving Average Convergence Divergence (MACD), momentum, daily price differences (Close_diff), and volume—to enrich the input representation. The model uses a patch-based temporal encoding system that lets it quickly find both local and global patterns in noisy, changing financial time series. HSAT was tested on a dataset of 25 high-volume stocks from five major international markets over a period of ten years. The architecture was independently trained and optimized for each forecasting task. Results show that HSAT achieves strong and robust performance: for trend classification, it consistently outperforms both traditional machine learning models and state-of-the-art transformer variants, reaching an average F1 score of 0.70 and maintaining reliable prediction accuracy across all tested stocks, without exhibiting class bias. HSAT shows competitive error metrics (RMSE, MAE) for price regression compared to baseline models. It makes stable predictions even when the market is volatile. The HSAT architecture is modular and flexible, which makes it good for future improvements in real-world financial applications, such as integrating data from multiple sources and modelling trends.