AI Models
XGBoost loss scoring, LightGBM premium pricing, forecasting, and ML-assisted fraud risk signals.
Model Portfolio
| Objective | Model | Output used by |
|---|---|---|
| Loss / disruption risk | XGBoost | computes Lf (composite loss fraction) |
| Dynamic weekly pricing transform | LightGBM | calibrates premium signal |
| Ahead-of-time disruption forecasting | time-series regression | supports forward-looking discounts/updates |
| Fraud anomaly detection | IsolationForest + GBDT | ML probability for suspicious claims |
Model types are already listed above: XGBoost (risk/Lf), LightGBM (pricing), forecasting (near-term risk), and fraud gating (IsolationForest + GBDT).
Risk Assessment Pipeline (Placeholder)
Fraud ML Scoring (Placeholder)
Fraud ML gate (IsolationForest + GBDT) is executed during claim automation (see Adversarial Defense and End-to-End Workflow).
Operational Strategy
Aegis starts deterministic and transparent:
- trigger thresholds and eligibility are deterministic
- ML is layered gradually behind observability and fallback logic
- the fraud gate is built to minimize impact on honest riders