This project leverages deep learning to predict high-frequency price changes of two US stocks. Utilizing a dataset devoid of temporal dependencies allows for the application of simpler models such as Feedforward Neural Networks (FNNs), enhancing the training process and statistical analysis efficiency.
- Data A: A 200,000×22 matrix, with the first column indicating midprice direction changes and the remaining columns representing features.
- Data B (No Labels): A 20,000×21 matrix used for validation, following the same structure as Data A but without labels.
- Standardization: Applied to normalize feature values to zero mean and unit variance, facilitating faster convergence and improved model performance.
- Validation Split: Used to evaluate model generalization and tune hyperparameters.
- Output Layer: Utilizes the sigmoid activation function, suitable for binary classification tasks.
- Hidden Layers: Employs ReLU (Rectified Linear Unit) to introduce non-linearity and avoid the saturation problem, despite potential issues with "Dead ReLU" syndrome.
- Architecture Comparison: Explores both two-layer and three-layer networks to balance complexity, efficiency, and risk of overfitting.
- Optimizer: Adam, chosen for its adaptive learning rate capabilities, proving effective for large datasets.
- Learning Rate: Fine-tuned to 0.001 after evaluating loss over training epochs to balance learning speed and model stability.
- Batch Size: Set to 500, balancing efficiency with generalization.
- Epochs: Limited to 5 to prevent underfitting or overfitting while maintaining computational efficiency.
The project demonstrates that a Feedforward Neural Network, with carefully selected hyperparameters and preprocessing techniques, can effectively predict stock price direction changes with an accuracy range of 73% to 75%.
- Bengio, Y. (2012). "Practical recommendations for gradient-based training of deep architectures."