bluegreencity-ki

ML Risk Assessment

This section collects the project documentation for the AutoGluon-based ML risk branch.

ml_methodology_story.md - Entry point: the continuous methodology thread (data → target → features → model → validation → calibration → thresholds → operational products), with per-chapter Grafana/doc navigation and FAQ
ml_risk_scoring_guide.md - End-to-end guide for labels, training, inference, thresholds, and ML output columns
automl_features.md - Feature documentation for the current October 2026 operational model (54 features); the operative feature set is defined in config/config.yaml (feature_columns)
../validation/validation.md - Validation workflow, standard OOF metrics, grouped CV, and baseline comparison
../workflow.md - Overall pipeline workflow including the optional ML branch

reports/README.md - Index of finalized ML analysis reports
reports/treelife_overlap_report_2026-10-09.md - Current status: external validation of the 2026_10 retraining against real LHH Treelife interventions (issue #257); soft monitoring signal, not an acceptance gate
reports/operational_panel_grouped_cv_report_2026-09-16.md - Latest published robustness report: grouped 5-fold panel-aware check for future_drought_damage_severe_t1 (September 2026 retraining round: v6 panel, 35 features, 3600 s per fold). The October 2026 round (2026_10, 54 features) has its grouped-CV results in the model directory (data/models/autogluon/annual_validation/2026_10/.../grouped_cv/), the successor report is pending
reports/operational_panel_result_report_2026-04-08.md - Historical April 2026 reference: operational panel model vs. baseline and minimal, including 2024→2025 backtest validation
archive/operational_panel_grouped_cv_report_2026-06-08.md - Historical June 2026 reference: grouped 5-fold robustness check before the September 2026 retraining round (v4-era panel, 1800 s per fold)
reports/stable_ablation_report_2026-04-03.md - Historical reference: Stable AutoGluon benchmark and feature-ablation comparison
reports/high_risk_feature_combinations_2026-04-07.md - Analysis of the ML feature combinations behind the highest predicted risk