ensemble.py

import sqlite3
import os
import numpy as np

DB_FILE = "alphatrade_v31_dino.db" # 🦖 Base de données V30/V33

def get_dynamic_weights(symbol, timeframe, regime):
    # 1. Poids de base selon le timeframe
    if timeframe in ["1m", "5m"]:
        w_time, w_ml, w_dino, w_sent = 0.50, 0.25, 0.15, 0.10
    else:
        w_time, w_ml, w_dino, w_sent = 0.30, 0.30, 0.25, 0.15

    # 2. Ajustement contextuel (Régime de marché)
    if regime == 0:   w_dino += 0.05; w_ml -= 0.05 
    elif regime == 1: w_time += 0.05; w_dino -= 0.05 
    elif regime == 2: w_ml += 0.05; w_time -= 0.05   

    if not os.path.exists(DB_FILE): 
        total = w_time + w_ml + w_dino + w_sent
        return w_time/total, w_ml/total, w_dino/total, w_sent/total
        
    try:
        with sqlite3.connect(DB_FILE, timeout=2) as conn:
            cursor = conn.cursor()
            # On cherche les 10 derniers trades clôturés (vrai résultat)
            cursor.execute('''SELECT direction, status, prob_time, prob_ml, prob_lstm 
                              FROM signals WHERE symbol=? AND timeframe=? AND (status LIKE '%GAGNÉ%' OR status LIKE '%PERDU%')
                              ORDER BY id DESC LIMIT 10''', (symbol, timeframe))
            trades = cursor.fetchall()

        # 3. AUTO-LEARNING (Le juge des cerveaux)
        if len(trades) >= 3:
            time_score, ml_score, dino_score = 0, 0, 0
            for t in trades:
                direction, status, p_time, p_ml, p_dino = t
                # Le marché a-t-il vraiment pump ou dump ?
                actual_up = ('GAGNÉ' in status and direction == 'HAUSSIER') or ('PERDU' in status and direction == 'BAISSIER')
                
                # Distribution des bons points
                if (p_time > 0.5) == actual_up: time_score += 1
                if (p_ml > 0.5) == actual_up: ml_score += 1
                if (p_dino > 0.5) == actual_up: dino_score += 1

            # 4. REWARD : On booste le meilleur cerveau
            if time_score > ml_score + 2 and time_score > dino_score + 2: 
                w_time += 0.10; w_ml -= 0.05; w_dino -= 0.05
            elif dino_score > time_score + 2 and dino_score > ml_score + 2: 
                w_dino += 0.10; w_time -= 0.05; w_ml -= 0.05
            elif ml_score > dino_score + 2 and ml_score > time_score + 2: 
                w_ml += 0.10; w_dino -= 0.05; w_time -= 0.05
                
    except Exception as e: 
        print(f"⚠️ Erreur Auto-Learning Ensemble : {e}") 
    
    # 🛡️ SÉCURITÉ V33 : On empêche les poids de tomber à 0 ou en négatif
    w_time, w_ml, w_dino, w_sent = max(0.05, w_time), max(0.05, w_ml), max(0.05, w_dino), max(0.05, w_sent)
    
    # Normalisation stricte pour que la somme fasse exactement 1.0
    total = w_time + w_ml + w_dino + w_sent
    return w_time/total, w_ml/total, w_dino/total, w_sent/total

def combine_scores(symbol, timeframe, t, m, dino, sent, r):
    # Récupération des poids intelligents
    w_t, w_m, w_d, w_s = get_dynamic_weights(symbol, timeframe, r)
    
    # Si le sentiment est parfaitement neutre (0.5), on l'exclut et on redistribue
    if sent == 0.5:
        w_s = 0.0
        total = w_t + w_m + w_d
        w_t, w_m, w_d = w_t/total, w_m/total, w_d/total
        
    # L'équation finale de l'Ensemble
    final_p = (t * w_t) + (m * w_m) + (dino * w_d) + (sent * w_s)

    # 🛡️ PROTECTION ANTI-CRASH V33
    if np.isnan(final_p) or np.isinf(final_p):
        final_p = 0.5

    # 🛑 GARDE-FOU TENDANCE : Empêche l'IA de shorter un Bull Run ou longer un Bear Run
    if r == 0 and final_p < 0.45: final_p = 0.45 
    elif r == 1 and final_p > 0.55: final_p = 0.55 
        
    return final_p, w_t, w_m, w_d, w_s