Stock Market Prediction Using Deep Learning

Abstract

Accurate stock market prediction remains a complex research challenge due to the market’s irregular, non-linear, and highly dynamic multivariate nature. Traditional statistical methods often struggle to capture the volatile patterns, and deep dependencies present in stock time-series data. This thesis addresses the problem by proposing a hybrid deep learning-based fusion model designed to improve the accuracy of stock price trend forecasting. The proposed solution adopts a late fusion approach, integrating the complementary strengths of four models: Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) to capture sequential dependencies, the Temporal Fusion Transformer (TFT) for modeling complex temporal relationships through attention mechanisms, and the Multi-Layer Perceptron (MLP) for nonlinear feature interactions in time-series data. This architecture forms a robust forecasting system, trained on multivariate time-series data from the NASDAQ-100, which includes both raw stock metrics and derived technical indicators. The fusion model achieved an R² score of 0.99196 and an MSE of 0.00033, clearly outperforming standalone LSTM, GRU, and stacked LSTM-GRU baselines. The predictions closely follow actual market movements with minimal lag, capturing both bullish and bearish trends effectively. This research presents a high-accuracy predictive framework that offers meaningful contributions to the field of stock forecasting, helping investors make informed buy/sell/hold decisions, thereby reducing risk and improving investment strategies.