Predicting Stock Price Movements Using Deep Learning Techniques

This project aimed to assess the performance of attention extended LSTM and Gated Recurrent Unit models in stock price movement forecasting. The traditional models suffer from the challenges due to noisiness of financial markets, leading to incomplete specification of time dependencies in financial time series. These limitations are addressed by integrating attention mechanisms in LSTM and GRU models to allow the models to focus on specific critical time steps and features. It is shown that these enhanced models attain much improved predictive performance over their traditional counterparts. Long term dependencies and the reduction of the noise in stock price data are captured with increased model efficiency using attention mechanisms. Considering the stock data from seven major tech companies and macroeconomic indicators, the research uses the CRISP-DM methodology for system development. The study develops and tests various deep learning models in the form of traditional LSTM, GRU and attention-based variants, using data cleaning and feature engineering to preprocess the data. It is revealed by the findings that attention enhanced models outperform traditional approaches and that attention LSTM and attention GRU models achieve better performance metrics such as lower RMSE values and higher R2 values. These results reveal the potential of attention mechanisms for stock price movement forecasting as well as their impact on model generalizability and practical implications in the financial market. Future research recommendations include exploring alternative data sources, increasing the explainability of the model, and broadening the scope of predictive applications from different markets.