An everything-bagel of 2019 AI, applied to one stock ticker

What it does

This is a monolithic tutorial notebook, not a library. It builds a complete pipeline to predict daily price movements of Goldman Sachs stock by fusing historical prices, technical indicators, news sentiment, correlated assets, and options data. The centerpiece is a Generative Adversarial Network—an LSTM generator paired with a 1D CNN discriminator—trained in MXNet/Gluon.

The interesting bit

The author treats the project as a kitchen-sink demonstration of circa-2019 machine learning: BERT for NLP sentiment, Fourier transforms for trend decomposition, stacked autoencoders for hidden feature extraction, ARIMA, PCA eigen-portfolios, and self-organized maps for options anomaly detection. Hyperparameters are tuned via both Bayesian optimization with Gaussian processes and reinforcement learning using Rainbow and PPO, which is overkill in a charmingly academic way.

Key highlights

• GAN architecture with LSTM generator and CNN discriminator for time-series synthesis
• Hyperparameter search via Bayesian optimization and RL agents (Rainbow, PPO)
• Feature engineering spans BERT sentiment, Fourier transforms, ARIMA, stacked autoencoders, and PCA
• Statistical sanity checks (heteroskedasticity, multicollinearity, serial correlation) and XGBoost feature importance
• Built on MXNet/Gluon and designed for multi-GPU training

Caveats

• It is a single narrative notebook from January 2019, not a maintained package or reusable framework
• The repository is tagged as JavaScript, though the work is clearly Python-based (Jupyter/MXNet)
• The README is long on methodology but truncated before showing concrete results or performance metrics

Verdict

Worth a skim if you want to see how pre-Transformer-hype deep learning pieces fit together in finance; avoid if you need a robust backtesting framework or a library to import.