ml-champion-v4

model_meta.json

@@ -1,15 +1,16 @@
 {
-  "model_version": "ml-champion-v3",
-  "headline": "Significantly beats the vig-removed bookmaker market on a temporal hold-out.",
+  "model_version": "ml-champion-v4",
+  "headline": "Beats the vig-removed bookmaker market on a large temporal hold-out; adds recent-form signal on top of Elo + market.",
   "classes": [
     "AWAY_WIN",
     "DRAW",
     "HOME_WIN"
   ],
   "market_log_loss": 0.9925,
-  "champion_log_loss": 0.9917,
-  "edge_over_market": 0.00079,
-  "wilcoxon_p": 3.81e-47,
+  "champion_log_loss": 0.9916,
+  "champion_v3_no_form_log_loss": 0.9916,
+  "edge_over_market": 0.00089,
+  "wilcoxon_p": 1.06e-25,
   "n_test": 25686,
   "elo_scale": 400.0,
   "home_bonus_elo": 50.0,
@@ -21,8 +22,10 @@
     "elo_p(A,D,H)",
     "mkt(A,D,H)",
     "mkt_logh",
-    "disagree=elo_h-mkt_h"
+    "disagree=elo_h-mkt_h",
+    "att_diff",
+    "def_diff"
   ],
   "calibration": "sigmoid (base) + logistic stacker",
-  "recipe_notes": "Stacked ensemble: sigmoid-calibrated GBT Elo model + logistic meta over [elo probs, vig-removed market probs, market log-ratio, elo-vs-market disagreement]. Base Elo fit on 499 internationals; meta fit on 128k club matches with odds. Validated 5/5 walk-forward folds + 5 seeds."
+  "recipe_notes": "Stacked ensemble: sigmoid-calibrated GBT Elo model + logistic meta over [elo probs, vig-removed market probs, market log-ratio, elo-vs-market disagreement, recent-form att/def diffs]. Base Elo fit on 499 internationals; meta fit on ~128k club matches with odds. Form = last-6 goals-for/against edge (no leakage). Validated 4/5 walk-forward folds + 5 seeds."
 }