What problem does it solve? Financial-services ML prototypes often fail because of label leakage, missing baselines, and predictions that downstream layers cannot consume. This Skill guides the full Layer 2 lifecycle of a Databricks FSI solution so a fraud or risk model ends as a versioned, explainable, queryable prediction rather than an untracked notebook. ## Core Features & Use Cases - Leakage-safe feature engineering: Enforces chronological splits, point-in-time rolling features, and training-only preprocessing fits. - FSI evaluation discipline: Requires PR-AUC, calibration, analyst-capacity metrics, slice analysis, and baseline comparisons before model selection. - MLflow operationalization: Logs parameters, metrics, schemas, and thresholds, registers the model, and persists scores with reason codes to a governed prediction table. - Use Case: A solutions engineer building a fraud-detection demo uses this Skill to train a classifier on synthetic transaction data, register it in MLflow, and hand off trace-keyed predictions with reason codes to the GenAI and presentation layers. ## Quick Start Use the databricks-fsi-predictive-ml skill to build a leakage-safe fraud classification model with MLflow tracking and persist scored predictions with reason codes.
