ML Model Release v1.0.0 (XGBoost Failure Prediction)

This release includes the primary inference artifacts for the ApiCortex failure prediction engine, optimized for high-precision downtime detection.

Artifacts

• xgboost_failure_prediction_v1_clean.pkl: The trained XGBoost model utilized by the ml-service to predict API transaction failures.
• model_metadata_v1_clean.pkl: Associated feature mapping and normalization metadata required for consistent inference performance.

Model Performance Metrics

The model was evaluated using a recall-prioritized threshold (0.73) to maximize early detection of downtime events while maintaining a precision floor.

Metric	Score
ROC-AUC	0.9692
PR-AUC	0.8866
Precision	0.7979
Recall	0.8528
F1-Score	0.8244
F2-Score	0.8412

Feature Importance (SHAP Analysis)

The model's decisions are primarily driven by rolling latency windows and error rate variance. The top 5 influential features are:

1. p95_latency_roll_max_15: Maximum P95 latency over the last 15 intervals.
2. error_rate_ewm: Exponentially weighted moving average of the error rate.
3. latency_p95_zscore: Statistical deviation of P95 latency from the baseline.
4. p95_latency_roll_mean_15: Average P95 latency over the last 15 intervals.
5. error_rate: Raw error rate in the current window.

---

Important

This model utilizes the ml-service's RollingFeatureEngineer. Ensure that your data ingestion pipeline is active to provide the necessary 15-interval historical window for accurate predictions.

Contributors

• @0xarchit
• @vxrachit
• @synapticPush