Stock Sense: Deep Learning based stock-market prediction tool

Computer Science and Engineering (Data Science)

In the past few years, many financial forecasting methods have employed deep learning techniques to learn more complex, non-linear temporal relationships in time-series data. This project involves predicting stock prices utilizing two sophisticated Recurrent Neural Network (RNN) structures: Long Short-Term Memory (LSTMs) and Bidirectional Long Short-Term Memory (Bi-LSTMs). This study uses historical stock market data provided by Yahoo Finance to create models that analyze stock market prices to project forward in time stock prices as well as determine the comparison between the two models. The four major modules are: data collection, data preprocessing, model training and evaluation, and results visualization. The data collection module collects and extracts structured stock data including the following Open, High, Low, Close, and Volume values. The data preprocessing module takes the raw data and prepares it to be cleaned and normalized and assigns agglomeration of new processed data with several technical indicators (derived from financial data), which may capture the underlying market trends. The financial time-series structure must stay intact and having sliding windows allows for creating datasets a model can accept. The core modelling phase involves creating and training LSTM and Bi-LSTM networks with TensorFlow/Keras. The measure of performance of LSTM and Bi-LSTM networks includes measuring Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R² Score. The Bi-LSTM exhibited superior accuracy (or perhaps better generalisation) than the uni-directional LSTM networks since the Bi LSTM requires the data to be shaped in both the forward and backward direction. The final module is the visualization phase in which actual vs predicted price comparisons, training-validation loss curves, and technical indicators have been overlaid graphically to provide intuitive insights into the performance of the model. The findings suggest that Bi-LSTM consistently outperformed LSTM in terms of temporal dependencies associated with financial data; in this context, bidirectional architecture was useful for improving predictive accuracy. This project shows the practical feasibility of applying deep learning models for predicting financial data and sets the stage for even more advanced hybrid models that include additional market indicators and sentiment analysis in future refinements.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.