diff --git a/api_v1/__init__.py b/api_v1/__init__.py
index 268fd62..2e2504a 100644
--- a/api_v1/__init__.py
+++ b/api_v1/__init__.py
@@ -22,6 +22,8 @@ Registers sub-blueprints:
   /api/v1/history     — historique préd. ML (Free:7j, Premium:90j, Pro:illimité)
   /api/v1/org/        — organisations Pro (multi-compte, max 5 users)
   /api/v1/docs        — Swagger UI (via flasgger, registered on app)
+  /api/v1/ml/feedback/run  — trigger feedback loop ML (admin)
+  /api/v1/ml/feedback/stats — stats par stratégie (premium+)
 """
 
 from flask import Blueprint
@@ -38,6 +40,7 @@ from .routes.user import user_bp
 from .routes.user_tokens import user_tokens_bp
 from .routes.history import history_bp
 from .routes.org import org_bp
+from .routes.ml_feedback import ml_feedback_bp
 
 # Master blueprint that aggregates all sub-routes under /api/v1
 api_v1_bp = Blueprint("api_v1", __name__, url_prefix="/api/v1")
@@ -57,3 +60,4 @@ def register_api_v1(app):
     app.register_blueprint(user_tokens_bp)
     app.register_blueprint(history_bp)
     app.register_blueprint(org_bp)
+    app.register_blueprint(ml_feedback_bp)
diff --git a/api_v1/routes/ml_feedback.py b/api_v1/routes/ml_feedback.py
new file mode 100644
index 0000000..7c4523f
--- /dev/null
+++ b/api_v1/routes/ml_feedback.py
@@ -0,0 +1,191 @@
+#!/usr/bin/env python3
+"""
+ml_feedback.py — API routes pour le feedback loop ML (turf_saas).
+
+Routes:
+  POST /api/v1/ml/feedback/run   — Déclenche le feedback loop (admin uniquement)
+  GET  /api/v1/ml/feedback/stats — Stats performances par stratégie
+
+Sécurité admin : token via variable d'environnement ML_ADMIN_TOKEN
+ou plan "pro" en fallback pour les stats.
+"""
+
+import os
+import sys
+from datetime import datetime
+
+from flask import Blueprint, jsonify, request, g
+
+# Ajoute le répertoire parent de api_v1 dans le path pour importer ml_feedback_saas
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.dirname(__file__))))
+
+from api_v1.utils import get_db, internal_error, bad_request
+from auth import jwt_required_middleware, plan_required
+
+ml_feedback_bp = Blueprint("v1_ml_feedback", __name__, url_prefix="/api/v1/ml/feedback")
+
+# Token admin interne — configurable via variable d'environnement
+ML_ADMIN_TOKEN = os.environ.get("ML_ADMIN_TOKEN", "")
+
+
+def _check_admin(req):
+    """Vérifie le token admin via header X-Admin-Token ou Authorization Bearer (plan pro)."""
+    # 1. Token interne (scheduler/cron)
+    admin_token = req.headers.get("X-Admin-Token", "").strip()
+    if ML_ADMIN_TOKEN and admin_token == ML_ADMIN_TOKEN:
+        return True, None
+
+    # 2. Pas de token admin configuré → autoriser les utilisateurs "pro" authentifiés
+    user = getattr(g, "current_user", None)
+    if user and user.get("plan") == "pro":
+        return True, None
+
+    return False, jsonify({"error": "Accès admin requis", "code": 403}), 403
+
+
+@ml_feedback_bp.route("/run", methods=["POST"])
+@jwt_required_middleware
+def feedback_run():
+    """
+    Déclenche le feedback loop ML pour une date donnée.
+    ---
+    tags:
+      - ML Feedback
+    summary: Déclenche le feedback loop XGBoost (admin only)
+    security:
+      - Bearer: []
+      - AdminToken: []
+    parameters:
+      - name: body
+        in: body
+        schema:
+          type: object
+          properties:
+            date:
+              type: string
+              description: Date YYYY-MM-DD (défaut aujourd'hui)
+              example: "2026-04-25"
+            mode:
+              type: string
+              description: "run (défaut) ou backfill"
+              enum: [run, backfill]
+              example: run
+    responses:
+      200:
+        description: Feedback loop exécuté avec succès
+      400:
+        description: Paramètre invalide
+      403:
+        description: Accès refusé
+      500:
+        description: Erreur interne
+    """
+    # Vérification admin
+    user = getattr(g, "current_user", None)
+    admin_token = request.headers.get("X-Admin-Token", "").strip()
+    is_admin = (ML_ADMIN_TOKEN and admin_token == ML_ADMIN_TOKEN) or (
+        user and user.get("plan") == "pro"
+    )
+    if not is_admin:
+        return jsonify({"error": "Accès admin requis", "code": 403}), 403
+
+    body = request.get_json(silent=True) or {}
+    date_str = body.get("date") or datetime.now().strftime("%Y-%m-%d")
+    mode = body.get("mode", "run")
+
+    # Validation date
+    try:
+        datetime.strptime(date_str, "%Y-%m-%d")
+    except ValueError:
+        return bad_request(f"Date invalide : {date_str}. Format attendu : YYYY-MM-DD")
+
+    if mode not in ("run", "backfill"):
+        return bad_request("mode doit être 'run' ou 'backfill'")
+
+    try:
+        import ml_feedback_saas
+
+        if mode == "backfill":
+            inseres, maj = ml_feedback_saas.backfill(date_str)
+            total_inseres = inseres
+        else:
+            result = ml_feedback_saas.run(date_str)
+            total_inseres = sum(result["inseres"].values())
+            maj = result["maj"]
+
+        return jsonify(
+            {
+                "status": "ok",
+                "date": date_str,
+                "mode": mode,
+                "paris_inseres": total_inseres,
+                "paris_mis_a_jour": maj,
+            }
+        ), 200
+
+    except Exception as e:
+        return internal_error(str(e))
+
+
+@ml_feedback_bp.route("/stats", methods=["GET"])
+@jwt_required_middleware
+@plan_required("premium", "pro")
+def feedback_stats():
+    """
+    Stats performances ML par stratégie.
+    ---
+    tags:
+      - ML Feedback
+    summary: Stats paris ML par stratégie (premium+)
+    security:
+      - Bearer: []
+    parameters:
+      - name: date_debut
+        in: query
+        type: string
+        description: Date de début YYYY-MM-DD
+      - name: date_fin
+        in: query
+        type: string
+        description: Date de fin YYYY-MM-DD
+    responses:
+      200:
+        description: Stats par stratégie
+      401:
+        description: Token invalide
+      403:
+        description: Plan insuffisant (premium ou pro requis)
+    """
+    date_debut = request.args.get("date_debut")
+    date_fin = request.args.get("date_fin")
+
+    # Validation optionnelle des dates
+    for d_str, label in [(date_debut, "date_debut"), (date_fin, "date_fin")]:
+        if d_str:
+            try:
+                datetime.strptime(d_str, "%Y-%m-%d")
+            except ValueError:
+                return bad_request(f"{label} invalide : {d_str}. Format : YYYY-MM-DD")
+
+    conn = get_db()
+    try:
+        import ml_feedback_saas
+
+        stats = ml_feedback_saas.get_feedback_stats(conn, date_debut, date_fin)
+
+        return jsonify(
+            {
+                "status": "ok",
+                "strategies": stats,
+                "filters": {
+                    "date_debut": date_debut,
+                    "date_fin": date_fin,
+                },
+                "total_strategies": len(stats),
+            }
+        ), 200
+
+    except Exception as e:
+        return internal_error(str(e))
+    finally:
+        conn.close()
diff --git a/ml_feedback_saas.py b/ml_feedback_saas.py
new file mode 100644
index 0000000..2d2c0be
--- /dev/null
+++ b/ml_feedback_saas.py
@@ -0,0 +1,600 @@
+#!/usr/bin/env python3
+"""
+ml_feedback_saas.py — Feedback loop ML pour turf_saas.
+Enregistre les paris virtuels XGBoost depuis ml_predictions_cache
+et met à jour les résultats/dividendes depuis pmu_partants + pmu_rapports.
+
+DB cible : /home/h3r7/turf_saas/turf_saas.db
+
+Stratégies :
+  A) xgboost_sg    : simple_gagnant — top1 ML par course, ml_score >= 70, mise 1€
+  B) xgboost_value : simple_gagnant — is_value_bet = 1, mise 1€
+  C) xgboost_sp    : simple_place   — top1 ML par course, ml_score >= 50, mise 1€
+  D) xgboost_2sur4 : deux_sur_quatre — top4 ML par course, 6 combos x 1€ = mise 6€
+
+Usage :
+  python3 ml_feedback_saas.py                      # Traite aujourd'hui
+  python3 ml_feedback_saas.py --backfill 2026-04-25
+  python3 ml_feedback_saas.py --date 2026-04-25
+"""
+
+import sqlite3
+import sys
+import logging
+import os
+from datetime import datetime, timedelta
+
+DB_PATH = "/home/h3r7/turf_saas/turf_saas.db"
+
+os.makedirs("/home/h3r7/turf_saas/logs", exist_ok=True)
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [ml_feedback_saas] %(levelname)s %(message)s",
+    handlers=[
+        logging.FileHandler("/home/h3r7/turf_saas/logs/ml_feedback_saas.log"),
+        logging.StreamHandler(),
+    ],
+)
+log = logging.getLogger(__name__)
+
+
+# ─────────────────────────────────────────────────────────
+# UTILITAIRES
+# ─────────────────────────────────────────────────────────
+
+
+def get_db():
+    conn = sqlite3.connect(DB_PATH)
+    conn.row_factory = sqlite3.Row
+    return conn
+
+
+def pari_existe(cursor, date, num_reunion, num_course, numero1, type_pari, source_reco):
+    """Vérifie si un pari identique existe déjà (idempotence)."""
+    cursor.execute(
+        """
+        SELECT id FROM paris
+        WHERE date_course = ? AND source_reco = ?
+          AND type_pari = ? AND numero1 = ?
+          AND race_label = ?
+        """,
+        (date, source_reco, type_pari, numero1, f"R{num_reunion}C{num_course}"),
+    )
+    return cursor.fetchone() is not None
+
+
+def pari_2sur4_existe(cursor, date, num_reunion, num_course, source_reco):
+    """Vérifie si un pari 2sur4 existe déjà pour cette course."""
+    cursor.execute(
+        """
+        SELECT id FROM paris
+        WHERE date_course = ? AND source_reco = ?
+          AND race_label = ?
+        """,
+        (date, source_reco, f"R{num_reunion}C{num_course}"),
+    )
+    return cursor.fetchone() is not None
+
+
+def get_top_ml_par_course(cursor, date, n=4, min_score=0):
+    """Retourne les n meilleurs chevaux ML par course pour une date."""
+    cursor.execute(
+        """
+        SELECT num_reunion, num_course, horse_name, horse_number,
+               ml_score, odds, recommendation, is_value_bet,
+               race_label, race_name, hippodrome, heure,
+               discipline, distance
+        FROM ml_predictions_cache
+        WHERE date = ?
+          AND ml_score >= ?
+        ORDER BY num_reunion, num_course, ml_score DESC
+        """,
+        (date, min_score),
+    )
+    rows = cursor.fetchall()
+
+    courses = {}
+    for r in rows:
+        key = (r["num_reunion"], r["num_course"])
+        if key not in courses:
+            courses[key] = []
+        if len(courses[key]) < n:
+            courses[key].append(dict(r))
+    return courses
+
+
+# ─────────────────────────────────────────────────────────
+# STRATÉGIE A — Simple Gagnant top1 ML (score >= 70)
+# ─────────────────────────────────────────────────────────
+
+
+def save_ml_paris_sg(conn, date):
+    """Insère 1 pari simple_gagnant par course : top1 ML avec ml_score >= 70."""
+    cursor = conn.cursor()
+    courses = get_top_ml_par_course(cursor, date, n=1, min_score=70)
+    inseres = 0
+
+    for (num_reunion, num_course), chevaux in courses.items():
+        cheval = chevaux[0]
+        if pari_existe(
+            cursor,
+            date,
+            num_reunion,
+            num_course,
+            cheval["horse_number"],
+            "simple_gagnant",
+            "xgboost_sg",
+        ):
+            continue
+
+        cursor.execute(
+            """
+            INSERT INTO paris
+              (date_pari, date_course, race_name, race_label, hippodrome,
+               type_pari, chevaux, cheval1, numero1, cote, mise,
+               statut, gain, source_reco, model_source)
+            VALUES (?, ?, ?, ?, ?, 'simple_gagnant', ?, ?, ?, ?, 1.0,
+                    'EN_ATTENTE', 0.0, 'xgboost_sg', 'xgboost_v1')
+            """,
+            (
+                date,
+                date,
+                cheval.get("race_name") or "",
+                f"R{num_reunion}C{num_course}",
+                cheval.get("hippodrome") or "",
+                cheval["horse_name"],
+                cheval["horse_name"],
+                cheval["horse_number"],
+                cheval["odds"],
+            ),
+        )
+        inseres += 1
+
+    conn.commit()
+    log.info(f"[SG]     {date} → {inseres} paris simple_gagnant insérés (score>=70)")
+    return inseres
+
+
+# ─────────────────────────────────────────────────────────
+# STRATÉGIE B — Value Bet (is_value_bet = 1)
+# ─────────────────────────────────────────────────────────
+
+
+def save_ml_paris_value(conn, date):
+    """Insère 1 pari simple_gagnant pour chaque cheval is_value_bet=1."""
+    cursor = conn.cursor()
+    cursor.execute(
+        """
+        SELECT num_reunion, num_course, horse_name, horse_number,
+               ml_score, odds, race_label, race_name, hippodrome
+        FROM ml_predictions_cache
+        WHERE date = ? AND is_value_bet = 1
+        ORDER BY num_reunion, num_course, ml_score DESC
+        """,
+        (date,),
+    )
+    rows = [dict(r) for r in cursor.fetchall()]
+    inseres = 0
+
+    for r in rows:
+        if pari_existe(
+            cursor,
+            date,
+            r["num_reunion"],
+            r["num_course"],
+            r["horse_number"],
+            "simple_gagnant",
+            "xgboost_value",
+        ):
+            continue
+
+        cursor.execute(
+            """
+            INSERT INTO paris
+              (date_pari, date_course, race_name, race_label, hippodrome,
+               type_pari, chevaux, cheval1, numero1, cote, mise,
+               statut, gain, source_reco, model_source)
+            VALUES (?, ?, ?, ?, ?, 'simple_gagnant', ?, ?, ?, ?, 1.0,
+                    'EN_ATTENTE', 0.0, 'xgboost_value', 'xgboost_v1')
+            """,
+            (
+                date,
+                date,
+                r.get("race_name") or "",
+                r.get("race_label") or f"R{r['num_reunion']}C{r['num_course']}",
+                r.get("hippodrome") or "",
+                r["horse_name"],
+                r["horse_name"],
+                r["horse_number"],
+                r["odds"],
+            ),
+        )
+        inseres += 1
+
+    conn.commit()
+    log.info(f"[VALUE]  {date} → {inseres} paris value_bet insérés")
+    return inseres
+
+
+# ─────────────────────────────────────────────────────────
+# STRATÉGIE C — Simple Placé top1 ML (score >= 50)
+# ─────────────────────────────────────────────────────────
+
+
+def save_ml_paris_sp(conn, date):
+    """Insère 1 pari simple_place par course : top1 ML avec ml_score >= 50."""
+    cursor = conn.cursor()
+    courses = get_top_ml_par_course(cursor, date, n=1, min_score=50)
+    inseres = 0
+
+    for (num_reunion, num_course), chevaux in courses.items():
+        cheval = chevaux[0]
+        if pari_existe(
+            cursor,
+            date,
+            num_reunion,
+            num_course,
+            cheval["horse_number"],
+            "simple_place",
+            "xgboost_sp",
+        ):
+            continue
+
+        cursor.execute(
+            """
+            INSERT INTO paris
+              (date_pari, date_course, race_name, race_label, hippodrome,
+               type_pari, chevaux, cheval1, numero1, cote, mise,
+               statut, gain, source_reco, model_source)
+            VALUES (?, ?, ?, ?, ?, 'simple_place', ?, ?, ?, ?, 1.0,
+                    'EN_ATTENTE', 0.0, 'xgboost_sp', 'xgboost_v1')
+            """,
+            (
+                date,
+                date,
+                cheval.get("race_name") or "",
+                f"R{num_reunion}C{num_course}",
+                cheval.get("hippodrome") or "",
+                cheval["horse_name"],
+                cheval["horse_name"],
+                cheval["horse_number"],
+                cheval["odds"],
+            ),
+        )
+        inseres += 1
+
+    conn.commit()
+    log.info(f"[SP]     {date} → {inseres} paris simple_place insérés (score>=50)")
+    return inseres
+
+
+# ─────────────────────────────────────────────────────────
+# STRATÉGIE D — 2sur4 top4 ML (6 combinaisons x 1€)
+# ─────────────────────────────────────────────────────────
+
+
+def save_ml_paris_2sur4(conn, date):
+    """Insère 1 pari deux_sur_quatre par course : top4 ML, mise 6€."""
+    cursor = conn.cursor()
+    courses = get_top_ml_par_course(cursor, date, n=4, min_score=0)
+    inseres = 0
+
+    for (num_reunion, num_course), chevaux in courses.items():
+        if len(chevaux) < 4:
+            continue
+        if pari_2sur4_existe(cursor, date, num_reunion, num_course, "xgboost_2sur4"):
+            continue
+
+        top4 = chevaux[:4]
+        nums = [str(c["horse_number"]) for c in top4]
+        noms = [c["horse_name"] for c in top4]
+        chevaux_str = "/".join(noms)
+
+        cursor.execute(
+            """
+            INSERT INTO paris
+              (date_pari, date_course, race_name, race_label, hippodrome,
+               type_pari, chevaux, cheval1, numero1, cote, mise,
+               statut, gain, source_reco, model_source, commentaire)
+            VALUES (?, ?, ?, ?, ?, 'deux_sur_quatre', ?, ?, ?, 0.0, 6.0,
+                    'EN_ATTENTE', 0.0, 'xgboost_2sur4', 'xgboost_v1', ?)
+            """,
+            (
+                date,
+                date,
+                top4[0].get("race_name") or "",
+                f"R{num_reunion}C{num_course}",
+                top4[0].get("hippodrome") or "",
+                chevaux_str,
+                top4[0]["horse_name"],
+                top4[0]["horse_number"],
+                f"top4 ML: {'/'.join(nums)}",
+            ),
+        )
+        inseres += 1
+
+    conn.commit()
+    log.info(f"[2S4]    {date} → {inseres} paris deux_sur_quatre insérés")
+    return inseres
+
+
+# ─────────────────────────────────────────────────────────
+# UPDATE RÉSULTATS + DIVIDENDES
+# ─────────────────────────────────────────────────────────
+
+
+def update_ml_paris_results(conn, date):
+    """
+    Met à jour statut + gain (dividende PMU réel) pour tous les paris ML EN_ATTENTE.
+    Sources: pmu_partants (ordre_arrivee) + pmu_rapports (dividende_euro).
+    """
+    cursor = conn.cursor()
+
+    cursor.execute(
+        """
+        SELECT id, race_label, type_pari, numero1, chevaux, mise, source_reco, commentaire
+        FROM paris
+        WHERE date_course = ? AND statut = 'EN_ATTENTE'
+          AND source_reco LIKE 'xgboost%'
+        """,
+        (date,),
+    )
+    paris = [dict(r) for r in cursor.fetchall()]
+
+    if not paris:
+        log.info(f"[UPDATE] {date} → aucun pari ML EN_ATTENTE")
+        return 0
+
+    maj = 0
+    for pari in paris:
+        pari_id = pari["id"]
+        race_label = pari["race_label"] or ""
+        type_pari = pari["type_pari"]
+        numero1 = pari["numero1"]
+        mise = pari["mise"]
+
+        # Extraire num_reunion / num_course depuis le race_label "R{r}C{c}"
+        try:
+            parts = race_label.replace("R", "").split("C")
+            num_reunion = int(parts[0])
+            num_course = int(parts[1])
+        except Exception:
+            log.warning(f"[UPDATE] race_label invalide : {race_label}")
+            continue
+
+        if type_pari == "simple_gagnant":
+            cursor.execute(
+                """
+                SELECT ordre_arrivee FROM pmu_partants
+                WHERE date_programme = ? AND num_reunion = ?
+                  AND num_course = ? AND num_pmu = ?
+                """,
+                (date, num_reunion, num_course, numero1),
+            )
+            row = cursor.fetchone()
+            if not row or row["ordre_arrivee"] is None or row["ordre_arrivee"] == 0:
+                continue
+
+            gagne = row["ordre_arrivee"] == 1
+            gain = 0.0
+            if gagne:
+                cursor.execute(
+                    """
+                    SELECT dividende_euro FROM pmu_rapports
+                    WHERE date_programme = ? AND num_reunion = ?
+                      AND num_course = ? AND type_pari = 'SIMPLE_GAGNANT'
+                      AND CAST(combinaison AS INTEGER) = ?
+                      AND libelle NOT LIKE '%NP%'
+                    """,
+                    (date, num_reunion, num_course, numero1),
+                )
+                div = cursor.fetchone()
+                gain = div["dividende_euro"] if div and div["dividende_euro"] else 0.0
+
+            cursor.execute(
+                "UPDATE paris SET statut=?, gain=? WHERE id=?",
+                ("GAGNE" if gagne else "PERDU", gain, pari_id),
+            )
+            maj += 1
+
+        elif type_pari == "simple_place":
+            cursor.execute(
+                """
+                SELECT ordre_arrivee FROM pmu_partants
+                WHERE date_programme = ? AND num_reunion = ?
+                  AND num_course = ? AND num_pmu = ?
+                """,
+                (date, num_reunion, num_course, numero1),
+            )
+            row = cursor.fetchone()
+            if not row or not row["ordre_arrivee"]:
+                continue
+
+            gagne = 1 <= row["ordre_arrivee"] <= 3
+            gain = 0.0
+            if gagne:
+                cursor.execute(
+                    """
+                    SELECT dividende_euro FROM pmu_rapports
+                    WHERE date_programme = ? AND num_reunion = ?
+                      AND num_course = ? AND type_pari = 'SIMPLE_PLACE'
+                      AND CAST(combinaison AS INTEGER) = ?
+                      AND libelle NOT LIKE '%NP%'
+                    """,
+                    (date, num_reunion, num_course, numero1),
+                )
+                div = cursor.fetchone()
+                gain = div["dividende_euro"] if div and div["dividende_euro"] else 0.0
+
+            cursor.execute(
+                "UPDATE paris SET statut=?, gain=? WHERE id=?",
+                ("GAGNE" if gagne else "PERDU", gain, pari_id),
+            )
+            maj += 1
+
+        elif type_pari == "deux_sur_quatre":
+            # Récupère les 4 numéros depuis commentaire "top4 ML: n1/n2/n3/n4"
+            try:
+                nums_str = (
+                    pari["commentaire"].split(": ")[1]
+                    if pari.get("commentaire")
+                    else ""
+                )
+                nums_top4 = [int(n) for n in nums_str.split("/") if n.strip().isdigit()]
+            except Exception:
+                nums_top4 = []
+
+            if len(nums_top4) < 4:
+                # Fallback : reconstituer top4 depuis ml_predictions_cache
+                cursor.execute(
+                    """
+                    SELECT horse_number FROM ml_predictions_cache
+                    WHERE date = ? AND num_reunion = ? AND num_course = ?
+                    ORDER BY ml_score DESC LIMIT 4
+                    """,
+                    (date, num_reunion, num_course),
+                )
+                nums_top4 = [r["horse_number"] for r in cursor.fetchall()]
+
+            if len(nums_top4) < 2:
+                continue
+
+            cursor.execute(
+                """
+                SELECT combinaison, dividende_euro FROM pmu_rapports
+                WHERE date_programme = ? AND num_reunion = ?
+                  AND num_course = ? AND type_pari = 'DEUX_SUR_QUATRE'
+                  AND libelle NOT LIKE '%NP%'
+                """,
+                (date, num_reunion, num_course),
+            )
+            rapports = [dict(r) for r in cursor.fetchall()]
+            gain_total = 0.0
+
+            for rap in rapports:
+                try:
+                    n1, n2 = [int(x) for x in rap["combinaison"].split("-")]
+                except Exception:
+                    continue
+                if n1 in nums_top4 and n2 in nums_top4:
+                    gain_total += rap["dividende_euro"]
+
+            gagne = gain_total > 0
+            cursor.execute(
+                "UPDATE paris SET statut=?, gain=? WHERE id=?",
+                ("GAGNE" if gagne else "PERDU", round(gain_total, 2), pari_id),
+            )
+            maj += 1
+
+    conn.commit()
+    log.info(f"[UPDATE] {date} → {maj}/{len(paris)} paris ML mis à jour")
+    return maj
+
+
+# ─────────────────────────────────────────────────────────
+# STATS PAR STRATÉGIE
+# ─────────────────────────────────────────────────────────
+
+
+def get_feedback_stats(conn, date_debut=None, date_fin=None):
+    """Stats performances ML par stratégie (source_reco)."""
+    cursor = conn.cursor()
+    cursor.execute(
+        """
+        SELECT source_reco,
+               COUNT(*) as n_paris,
+               SUM(CASE WHEN statut='GAGNE' THEN 1 ELSE 0 END) as n_gagne,
+               SUM(CASE WHEN statut='PERDU' THEN 1 ELSE 0 END) as n_perdu,
+               SUM(CASE WHEN statut='EN_ATTENTE' THEN 1 ELSE 0 END) as n_attente,
+               ROUND(100.0 * SUM(CASE WHEN statut='GAGNE' THEN 1 ELSE 0 END)
+                     / NULLIF(SUM(CASE WHEN statut IN ('GAGNE','PERDU') THEN 1 ELSE 0 END), 0), 1) as win_rate_pct,
+               ROUND(SUM(gain), 2) as gain_total,
+               ROUND(SUM(mise), 2) as mise_totale,
+               ROUND(SUM(gain) - SUM(mise), 2) as roi_net
+        FROM paris
+        WHERE source_reco LIKE 'xgboost%'
+          AND (:debut IS NULL OR date_course >= :debut)
+          AND (:fin IS NULL OR date_course <= :fin)
+        GROUP BY source_reco
+        ORDER BY source_reco
+        """,
+        {"debut": date_debut, "fin": date_fin},
+    )
+    return [dict(r) for r in cursor.fetchall()]
+
+
+# ─────────────────────────────────────────────────────────
+# PIPELINE COMPLET
+# ─────────────────────────────────────────────────────────
+
+
+def run(date):
+    """Enregistre les paris ML du jour + met à jour les résultats de J-1."""
+    conn = get_db()
+    log.info(f"=== ml_feedback_saas.run({date}) ===")
+
+    # 1. Enregistre les paris ML pour la date (depuis le cache du jour)
+    sg = save_ml_paris_sg(conn, date)
+    vb = save_ml_paris_value(conn, date)
+    sp = save_ml_paris_sp(conn, date)
+    s4 = save_ml_paris_2sur4(conn, date)
+    log.info(f"[SAVE]   {date} → total insérés : SG={sg} VALUE={vb} SP={sp} 2S4={s4}")
+
+    # 2. Met à jour les résultats de J-1 (résultats PMU disponibles)
+    yesterday = (datetime.strptime(date, "%Y-%m-%d") - timedelta(days=1)).strftime(
+        "%Y-%m-%d"
+    )
+    maj = update_ml_paris_results(conn, yesterday)
+    log.info(f"[UPDATE] {yesterday} → {maj} paris mis à jour")
+
+    conn.close()
+    return {"inseres": {"sg": sg, "value": vb, "sp": sp, "2sur4": s4}, "maj": maj}
+
+
+def backfill(date):
+    """Backfill : insère ET met à jour les résultats pour une date passée."""
+    conn = get_db()
+    log.info(f"=== ml_feedback_saas.backfill({date}) ===")
+
+    sg = save_ml_paris_sg(conn, date)
+    vb = save_ml_paris_value(conn, date)
+    sp = save_ml_paris_sp(conn, date)
+    s4 = save_ml_paris_2sur4(conn, date)
+    log.info(f"[SAVE]   {date} → SG={sg} VALUE={vb} SP={sp} 2S4={s4}")
+
+    maj = update_ml_paris_results(conn, date)
+    log.info(f"[UPDATE] {date} → {maj} paris mis à jour")
+
+    conn.close()
+    return sg + vb + sp + s4, maj
+
+
+# ─────────────────────────────────────────────────────────
+# MAIN
+# ─────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    if "--backfill" in sys.argv:
+        idx = sys.argv.index("--backfill")
+        date = sys.argv[idx + 1] if idx + 1 < len(sys.argv) else None
+        if not date:
+            print("Usage: python3 ml_feedback_saas.py --backfill YYYY-MM-DD")
+            sys.exit(1)
+        inseres, maj = backfill(date)
+        print(f"Backfill {date} : {inseres} paris insérés, {maj} mis à jour")
+
+    elif "--date" in sys.argv:
+        idx = sys.argv.index("--date")
+        date = sys.argv[idx + 1] if idx + 1 < len(sys.argv) else None
+        if not date:
+            print("Usage: python3 ml_feedback_saas.py --date YYYY-MM-DD")
+            sys.exit(1)
+        result = run(date)
+        total = sum(result["inseres"].values())
+        print(f"Run {date} : {total} paris insérés, {result['maj']} mis à jour")
+
+    else:
+        result = run(datetime.now().strftime("%Y-%m-%d"))
+        total = sum(result["inseres"].values())
+        print(f"Run today : {total} paris insérés, {result['maj']} mis à jour")