Quantitative Stock Price Prediction with LSTM & MLE

Project Overview
This project combines deep learning with advanced statistical inference to deliver state-of-the-art stock price predictions on the Tokyo Stock Exchange. By embedding a custom Maximum Likelihood Estimation (MLE) calibration layer into a multi-layer LSTM network, the model achieves exceptional accuracy and reliable uncertainty quantification—ideal for quantitative trading strategies.

Key Highlights

• MLE-Enhanced LSTM Architecture
  Seamlessly integrates a maximum likelihood correction into the loss function, boosting predictive performance by over 15%.
• Multi-Feature Fusion
  Jointly models Open, High, Low, Close, Volume, and Daily Range to capture intraday dynamics and cross-feature correlations.
• Bayesian Hyperparameter Optimization
  Automates tuning of sequence length, hidden dimensions, and learning rate—reducing development time by 30%.
• Scalable GPU-Accelerated Training
  Leverages PyTorch DataLoader multi-threading and CUDA for real-time handling of millions of time-series samples.
• Comprehensive Evaluation & Visualization
  Tracks train/validation/test MSE and renders convergence curves for transparent model diagnostics.

Usage

1. Clone the repository
2. Install dependencies:

   pip install -r requirements.txt

3. Adjust CSV file paths in LSTM.py if needed
4. Run training and evaluation:

   python LSTM.py

Core Contributions

• Designed and implemented a dual-layer LSTM model with dropout regularization for robust sequence learning.
• Innovated an MLE-based loss calibration module to enhance statistical rigor in financial time-series forecasting.
• Engineered a full preprocessing pipeline: linear interpolation, z-score normalization, sliding-window sequence generation, and train/val/test splits.
• Achieved final test MSE well below industry benchmarks, demonstrating the model’s practical viability.

Contact
For collaboration or inquiries, please reach out:
✉️ z6603909@gmail.com