Basketball Knowledge

BKE metric & modeling summary

• BKE is a possession-level composite that synthesizes ORAPM/DRAPM-like model outputs, possession features, and decomposition layers. Inputs: data/features/, data/tracking/, official stats.
• Modeling notes: apply shrinkage/compression steps and layered decomposition (portable talent, utilization efficiency, archetype elevation, scheme stability). Validate via cross-season correlations and backtesting.

Validation & diagnostics (recommended)

• Cross-season Pearson correlations (BKE stability).
• Rank-recall / top-N predictive enrichment and mean_abs_rank_shift.
• RMSE / MAE of team/season predictions vs realized outcomes.
• Variance explained and compression diagnostics.
• Archetype prevalence by minute thresholds to verify rarity gates.

Forecast & simulation guidance

• Minutes projection: blend carry-anchor + age/salary/draft multipliers; use roster-adaptive normalization targets and replacement buffers.
• Scenarios: support end_of_season and preseason_snapshot mapping; keep forecast scoring leakage-safe (no clutch-minute leakage).
• For lineup projection, enforce archetype-minute eligibility and position constraints (require at least one true big in starter lineup when applicable).

Where to find canonical docs

• readme.md (pipeline order, position-band policy, high-level conventions)
• reference/bke_repo_context_summary_for_ai.md (repo-level summary and modeling plan)
• reference/archetypes/ (archetype variable definitions and gating notes)
• src/data_compute/compute_player_archetypes.py (centralized assignment logic)
• src/profile_aggregate/build_profile_aggregate.py (profile aggregation guidance)

Agent usage guidance

• When changing archetype or metric logic: (1) update readme.md pipeline order, (2) update loop/ notes, (3) add tests in tests/, (4) run backtest validation and publish a short summary in reports/.