NIFTY-50 Investment Intelligence
Stock forecasts, portfolio construction and risk for India's NIFTY-50, tested honestly on data the model never saw.
- When
- May – Jun 2026
- Crew
- Solo
- Domain
- Quant · AI
- Stack
- Python · LightGBM · SHAP · GARCH · FastAPI · React
- 18.7%
- CAGR out of sample
- 0.52
- Sharpe vs 0.39
- 49
- Stocks
The problem
Most stock-prediction projects look great because they quietly test on data the model has already seen. I wanted numbers I would actually trust.
What I built
LightGBM models that forecast 1, 5 and 21 days ahead across 49 stocks, validated with purged, embargoed walk-forward splits so nothing leaks from the future. On top of that: mean-variance and HRP portfolios, GARCH volatility, VaR and CVaR, SHAP explanations and anomaly detection, served through FastAPI and React.
The hard part
Being honest about the edge. The daily signal is small (rank IC around +0.02 to +0.04), and I report it that way instead of dressing it up.
Result
Over 518,000 out-of-sample predictions. A simple top-quintile strategy returned 18.7% a year with a Sharpe of 0.52, against 14.8% and 0.39 for an equal-weight benchmark, after 15 bps of costs.