Enhanced LSTM with ELU for Stock Market Forecasting

Undergraduate thesis implementing an enhanced LSTM (Long Short-Term Memory) neural network with ELU activation for stock market time-series forecasting. The project compares performance across different activation functions and evaluates forecasting accuracy.

Project Overview

This thesis demonstrates the implementation of LSTM neural networks enhanced with ELU (Exponential Linear Unit) activation for predicting stock market prices. The ELU activation helps mitigate the vanishing gradient problem and improves convergence compared to traditional ReLU-based architectures.

Key Features

• LSTM model architecture with ELU activation
• Time series data preprocessing and normalization
• Comparative analysis across activation functions (ReLU, ELU, Tanh, Sigmoid)
• Evaluation metrics: MAE, RMSE, R²
• Visualization of predictions vs actual data
• Hyperparameter tuning for optimal performance
• TensorFlow and PyTorch implementations

Technologies Used

• Language: Python
• ML Frameworks: TensorFlow/Keras, PyTorch
• Data Processing: NumPy, Pandas
• Visualization: Matplotlib, Seaborn
• Development: Jupyter Notebooks