gg3

ai-engine/scripts/generate_daily_picks.py

@@ -0,0 +1,154 @@
+"""
+Generate Daily Picks — the serving picker for the validated favourite policy.
+============================================================================
+Loads the saved leak-free MS model (models/favorite_v1) and applies the
+favourite-band value policy to a set of matches, emitting the day's STAKED
+picks and logging them for forward paper-trade settlement.
+
+Train/serve consistency: features MUST come from the SAME extractor that built
+training_data_v27.csv. Production path = run the extractor nightly INCLUDING
+upcoming (status NS) matches, then point this script at that CSV. Demo path =
+use the tail of the training CSV as stand-in "today" matches (with the real
+result shown, since those are settled).
+
+Policy: bet the MS side with the biggest model_prob - implied edge, ONLY if
+odds in [--lo,--hi] and edge>--margin. Flat 1u. No longshots, no parlays.
+Non-MS markets are NOT staked (efficient -> model error). One bet per match.
+
+Usage:
+  python scripts/generate_daily_picks.py --demo --n 20          # see it work now
+  python scripts/generate_daily_picks.py --features today.csv    # production
+  python scripts/generate_daily_picks.py --settle                # settle paper log
+"""
+from __future__ import annotations
+import argparse, json, os, sys, datetime
+import numpy as np, pandas as pd, xgboost as xgb
+
+if sys.stdout and hasattr(sys.stdout, "reconfigure"):
+    try: sys.stdout.reconfigure(encoding="utf-8")
+    except Exception: pass
+
+AI_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+MODEL_DIR = os.path.join(AI_DIR, "models", "favorite_v1")
+TRAIN_CSV = os.path.join(AI_DIR, "data", "training_data_v27.csv")
+PAPER_LOG = os.path.join(AI_DIR, "data", "paper_trades.csv")
+MS_ODDS = ["odds_ms_h", "odds_ms_d", "odds_ms_a"]
+MS_PICKS = ["1", "X", "2"]
+
+
+def load_model():
+    bst = xgb.Booster(); bst.load_model(os.path.join(MODEL_DIR, "model.json"))
+    with open(os.path.join(MODEL_DIR, "feature_cols.json"), encoding="utf-8") as f:
+        feats = json.load(f)
+    with open(os.path.join(MODEL_DIR, "metadata.json"), encoding="utf-8") as f:
+        meta = json.load(f)
+    return bst, feats, meta
+
+
+def pick_for_rows(df, bst, feats, lo, hi, margin):
+    X = df.reindex(columns=feats).apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    P = bst.predict(xgb.DMatrix(X))                 # [n,3] home/draw/away
+    O = df[MS_ODDS].apply(pd.to_numeric, errors="coerce").fillna(0.0).values
+    implied = np.where(O > 1.0, 1.0/O, np.nan)
+    edge = np.where(np.isnan(implied), -9.0, P - implied)
+    out = []
+    for i in range(len(df)):
+        k = int(np.argmax(edge[i])); o = float(O[i, k]); e = float(edge[i, k])
+        staked = (e > margin) and (lo <= o < hi)
+        out.append({"idx": i, "pick": MS_PICKS[k], "odds": round(o, 2),
+                    "model_prob": round(float(P[i, k]), 4), "edge": round(e, 4),
+                    "staked": staked})
+    return out
+
+
+def settle():
+    if not os.path.exists(PAPER_LOG):
+        print("No paper_trades.csv yet."); return
+    pt = pd.read_csv(PAPER_LOG)
+    open_bets = pt[pt["result"].isna()] if "result" in pt.columns else pt
+    if open_bets.empty:
+        print("No open bets to settle.");
+    # settle from training CSV scores if present, else needs DB (left as note)
+    src = pd.read_csv(TRAIN_CSV, low_memory=False, usecols=["match_id","score_home","score_away"])
+    sc = src.set_index("match_id")
+    def res(row):
+        if not pd.isna(row.get("result")): return row["result"]
+        m = sc.index == row["match_id"]
+        if not m.any(): return np.nan
+        r = sc[m].iloc[0]; sh, sa = r["score_home"], r["score_away"]
+        if pd.isna(sh): return np.nan
+        outcome = "1" if sh > sa else ("X" if sh == sa else "2")
+        won = (str(row["pick"]) == outcome)
+        return "WON" if won else "LOST"
+    pt["result"] = pt.apply(res, axis=1)
+    pt["pnl"] = pt.apply(lambda r: (r["odds"]-1.0) if r["result"]=="WON"
+                         else (-1.0 if r["result"]=="LOST" else np.nan), axis=1)
+    pt.to_csv(PAPER_LOG, index=False)
+    s = pt.dropna(subset=["pnl"])
+    if len(s):
+        print(f"Settled {len(s)} bets: hit={100*(s['result']=='WON').mean():.1f}%  "
+              f"ROI={100*s['pnl'].sum()/len(s):+.2f}%  net={s['pnl'].sum():+.1f}u")
+    return
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--features", help="CSV of upcoming matches in training schema")
+    ap.add_argument("--demo", action="store_true", help="use tail of training CSV as 'today'")
+    ap.add_argument("--n", type=int, default=20)
+    ap.add_argument("--lo", type=float, default=1.5)
+    ap.add_argument("--hi", type=float, default=2.2)
+    ap.add_argument("--margin", type=float, default=0.03)
+    ap.add_argument("--settle", action="store_true")
+    ap.add_argument("--log", action="store_true", help="append staked picks to paper_trades.csv")
+    args = ap.parse_args()
+
+    if args.settle:
+        settle(); return
+
+    bst, feats, meta = load_model()
+    print(f"Model {meta['version']} (trained {meta['trained_at']}, holdout "
+          f"ROI {meta['holdout_eval']['roi_pct']}%)  band[{args.lo},{args.hi}] margin {args.margin}\n")
+
+    if args.features:
+        df = pd.read_csv(args.features, low_memory=False)
+        demo = False
+    else:
+        df = pd.read_csv(TRAIN_CSV, low_memory=False).sort_values("mst_utc").tail(args.n).reset_index(drop=True)
+        demo = True
+        print("(DEMO: last matches of training CSV as stand-in for today)\n")
+
+    picks = pick_for_rows(df, bst, feats, args.lo, args.hi, args.margin)
+    staked = [p for p in picks if p["staked"]]
+    print(f"{len(df)} matches scanned -> {len(staked)} STAKED MS picks\n")
+    print(f"  {'match_id':<28}{'pick':>5}{'odds':>7}{'model%':>8}{'edge%':>7}" + ("   result" if demo else ""))
+    print("  "+"-"*60)
+    log_rows = []
+    for p in picks:
+        if not p["staked"]: continue
+        r = df.iloc[p["idx"]]; mid = str(r["match_id"])
+        res = ""
+        if demo:
+            sh, sa = r.get("score_home"), r.get("score_away")
+            if pd.notna(sh):
+                out = "1" if sh>sa else ("X" if sh==sa else "2")
+                res = "  WON" if p["pick"]==out else "  lost"
+        print(f"  {mid:<28}{p['pick']:>5}{p['odds']:>7.2f}{100*p['model_prob']:>8.1f}{100*p['edge']:>+7.1f}{res}")
+        log_rows.append({"logged_at": datetime.datetime.now().isoformat(timespec="seconds"),
+                         "match_id": mid, "market": "MS", "pick": p["pick"], "odds": p["odds"],
+                         "model_prob": p["model_prob"], "edge": p["edge"], "stake": 1.0,
+                         "result": np.nan, "pnl": np.nan})
+    if args.log and log_rows and not demo:
+        new = pd.DataFrame(log_rows)
+        if os.path.exists(PAPER_LOG):
+            new = pd.concat([pd.read_csv(PAPER_LOG), new], ignore_index=True)
+        new.to_csv(PAPER_LOG, index=False)
+        print(f"\n  logged {len(log_rows)} picks -> {PAPER_LOG}")
+    elif args.log and demo:
+        print("\n  (--log ignored in --demo; only real upcoming picks are logged)")
+    print("\nReminder: paper-trade only. Stake real money after weeks of forward")
+    print("CLV>0 + ROI>0 (settle with --settle, check scoreboard/clv_report).")
+
+
+if __name__ == "__main__":
+    main()