Forecasting service for market volatility and credit risk that ships calibrated uncertainty, not just point predictions. Gradient-boosted models and ARIMA are wrapped with conformal prediction for distribution-free prediction intervals, validated by a walk-forward backtesting harness, and served behind a FastAPI risk-scorecard API.
Why it exists: a risk number without a confidence interval is dangerous. QuantRisk treats calibrated uncertainty and honest backtesting as first-class, the way a risk desk actually needs.
flowchart LR
D[Market & credit data<br/>prices, fundamentals] --> FE[Feature engineering<br/>returns, vol, lags, TA]
FE --> M{Model zoo}
M --> X[XGBoost]
M --> L[LightGBM]
M --> AR[ARIMA / GARCH]
X & L & AR --> C[Conformal layer<br/>split / Mondrian intervals]
C --> BT[Walk-forward<br/>backtest harness]
C --> ML[(MLflow<br/>experiments + registry)]
BT --> SC[Risk scorecards<br/>+ analyst notebooks]
C --> API[FastAPI<br/>forecast + interval + scorecard]
API --> DASH[Dashboard]
| Typical portfolio project | QuantRisk | |
|---|---|---|
| Output | Point prediction | Prediction + calibrated interval (conformal) |
| Validation | Random train/test split | Walk-forward backtest (no leakage) |
| Tracking | A notebook | MLflow experiments + model registry |
| Delivery | A plot | FastAPI risk scorecards + analyst reports |
| Models | One model | XGBoost / LightGBM / ARIMA-GARCH + TimesFM 2.5 foundation model |
Python 3.11 · XGBoost · LightGBM · statsmodels (ARIMA/GARCH) ·
TimesFM 2.5 (zero-shot foundation model) ·
conformal prediction (MAPIE-style) · MLflow · FastAPI · pandas · Docker
🚧 Built in public, in phases — see ROADMAP.md. Every phase ships
tested code + a design note in docs/.
pip install -e ".[dev]"
pytestMIT © Venkateswarlu Nagineni