LSTM Stocks Lab

Context & Goals

Personal research lab for equities forecasting with a focus on reproducibility. Provides a workflow where data scientists can prototype models, run CI-backed evaluations, and visualize results in dashboards.

Architecture

• Data ingestion: Batch jobs pull price data, fundamentals, and sentiment, storing them in Parquet with metadata catalogs.
• Model training: TensorFlow LSTM models, hyperparameter sweeps via Python scripts, and Dockerized workloads ready for cloud scaling.
• Vector search: Redis Stack stores embeddings for fundamental similarity, powering peer analysis.
• Frontend: Next.js dashboards for equity watchlists, performance attribution, and error analysis.

Highlights

• GitHub Actions orchestrate nightly training/backtesting with result publishing to dashboards.
• Experiment registry built with SQLite + Prisma for fast comparisons.
• Docker Compose environment mirrors CI runs, making local iteration trivial.

What’s next

• Port inference layer to ONNX Runtime for faster scoring.
• Integrate with broker APIs for paper trading.