This project implements a complete quantitative research pipeline using machine learning for alpha factor generation and stock return prediction. It features a comprehensive factor library with systematic feature selection reducing 100+ candidates to 85 high-quality factors, multiple deep learning architectures (LSTM, ResNet, Transformer), automated hyperparameter tuning via Optuna, and a backtesting framework. The best-performing LSTM+ResNet model achieved a 72% performance improvement over linear regression baselines.