AlphaCycle — Stock Prediction Framework
Python XGBoost Pipeline Walk-forward
Context
How to build a reproducible, config-driven research stack for systematic equity signals?
Data & Modeling
Built a 3-layer framework (data ingestion → feature store → model training → reporting) with walk-forward validation and regime labels.
Results
Clean separation of data/model/report layers; config-driven pipelines; evaluation with robust metrics across multiple horizons.
Takeaways
Add live paper-trading integration and expand to alternative data sources.
Evaluation
- Walk-forward validation (2015–2024), 12-month rolling refit
- Metrics: directional accuracy, IC, Sharpe ratio
- Feature importance tracked via SHAP across horizons