LBJLincoln26/nba-feature-cache

[NBA] v2 notebook — load full-odds-historical-2017-2026.json (11023 keys) so engine market features fire on ALL 9 training seasons (was current-season only). Target: holdout Brier < 0.21139.

[NBA] add full-odds-historical-2017-2026.json (11023 keys, 9 seasons) — unblocks engine market features for 2017-2025 training rows

[colab] FULL REWRITE: pull-9-seasons + walk-forward holdout (overfit-safe winner picking) + numpy-only (no LGBM warning) + asserts + clean cells

[colab] pull ALL 9 seasons (11532 games), unlocks time-series features. Was 1257 games -> ~3800 alive; expect ~4400-4500 alive on full multi-season.

[feature_data] +8 historical seasons (2017-18 ... 2024-25) — full 11,532-game multi-season build for time-series feature activation

[colab] cell-3 fix: full-odds keys are YYYY-MM-DD_AWAY@HOME, parse correctly to (date,home,away) → was odds=0

[engine] Cat 84-88: per100 + combos + composite_prob + temporal_decay + ctx. 7152->7246 (+94 cols). Total 4210/7246=58.1% alive.

[engine] Cat 79-83: ref_x_team + depth_metrics + injury_heavy + shot_mix + vegas_dev. 7058->7152 (+94 cols, 92% alive). Total 4203/7152=58.8%

[engine] Cat 76-78: matchup-specific + form trajectory + odds-derived. 6995->7058 (+63 cols, ~93% alive). Total 4125/7058=58.4% alive on probe.

[engine] Cat 72-75: ratios+log/sq+balance+env_inter. 6880->6995 (+115 cols, ~93% alive)

[engine] Cat 68-71: REAL aug + top5 aggregates + h/a interactions + context interactions. 6452->6880 (+428 cols, 392 alive 92%)

[engine] Cat 68 REAL-DATA AUGMENTED (+248 cols / +216 alive on probe). Reads streaks/clutch/misc/pace/shot_zones/top5_players/team_est/franchise_real. Engine ceiling 6452 -> 6700.

[feature_data] +3 REAL API sources (player_top/team_est/franchise_real); replaces curated franchise estimates with NBA stats API truth

[feature_data] +4 sources (streaks/clutch/misc_team/shot_zones); player_data_merged 119->X fields/key

[feature_data] add travel + franchise + vegas + coach_ext + star_meta; player_data_merged now 119 fields/key (was 82)

[colab] notebook v2 — XGBoost+LightGBM on ALL ~3,800 alive features + TabICL on top-186, pick best winner

[feature_data] tracking_data.json — engine-compatible keys (rim_rate, mid_rate, three_rate, contested_shots, deflections, drives, drive_fg_pct, drive_tov_pct, drive_pts_pct, avg_speed, loose_balls)

[colab] tabicl_oracle_train.ipynb — pull all enriched feature data + run engine.build inside Colab + train TabICL on T4

[feature_data] add full-odds-2025-26.json (802 games × 249 cats)

[feature_data] 6 files: enriched ref+player_merged+quarter+polymarket+engine

[FRANKENSTEIN] NBA feature cache 6509x6452, used by Colab oracle-pipeline.ipynb to predict 2025-26 games via Kaggle Oracle pkl

initial commit