ml_engine/monte_carlo.py

# ml_engine/monte_carlo.py
import numpy as np
import pandas as pd
from arch import arch_model
import lightgbm as lgb
import traceback 
import json 

try:
    import pandas_ta as ta
except ImportError:
    print("⚠️ مكتبة pandas_ta غير موجودة، سيتم استخدام حسابات يدوية للمؤشرات.")
    ta = None

# 🔴 --- START OF CHANGE (FIX NumPy 2.0 Crash) --- 🔴
def _sanitize_results_for_json(results_dict):
    """
    Recursively converts numpy types (ndarray, np.float64, etc.)
    in a dictionary to standard Python types (list, float)
    to make it JSON serializable.
    """
    if isinstance(results_dict, dict):
        return {k: _sanitize_results_for_json(v) for k, v in results_dict.items()}
    elif isinstance(results_dict, list):
        return [_sanitize_results_for_json(v) for v in results_dict]
    elif isinstance(results_dict, np.ndarray):
        return results_dict.tolist()
    # (Removed np.float_ which is deprecated in NumPy 2.0)
    elif isinstance(results_dict, (np.float64, np.float32)):
        return float(results_dict)
    # (Removed np.int_ which is deprecated in NumPy 2.0)
    elif isinstance(results_dict, (np.int64, np.int32)):
        return int(results_dict)
    else:
        return results_dict
# 🔴 --- END OF CHANGE --- 🔴


class MonteCarloAnalyzer:
    def __init__(self):
        self.simulation_results = {}
    
    async def generate_1h_price_distribution(self, ohlcv_data, target_profit_percent=0.005):
        """
        (المرحلة 1 - سريعة)
        """
        try:
            if not ohlcv_data or '1h' not in ohlcv_data or len(ohlcv_data['1h']) < 30:
                if '15m' in ohlcv_data and len(ohlcv_data['15m']) >= 50:
                    closes = np.array([candle[4] for candle in ohlcv_data['15m']])
                else:
                    self.simulation_results = {'error': 'Insufficient OHLCV data (< 30 candles 1h)'}
                    return None
            else:
                all_closes = [candle[4] for candle in ohlcv_data['1h']]
                if '15m' in ohlcv_data and len(ohlcv_data['15m']) >= 16:
                    all_closes.extend([candle[4] for candle in ohlcv_data['15m'][-16:]])
                closes = np.array(all_closes)

            if len(closes) < 30:
                self.simulation_results = {'error': 'Insufficient combined OHLCV data (< 30 candles)'}
                return None
            
            current_price = closes[-1]
            if current_price <= 0:
                self.simulation_results = {'error': 'Invalid current price <= 0'}
                return None

            log_returns = np.log(closes[1:] / closes[:-1])
            log_returns = log_returns[~np.isnan(log_returns) & ~np.isinf(log_returns)]
            
            if len(log_returns) < 20:
                self.simulation_results = {'error': 'Insufficient log returns (< 20)'}
                return None

            mean_return = np.mean(log_returns)
            std_return = np.std(log_returns)
            
            num_simulations = 5000
            t_df = 10
            jump_lambda = 0.05
            jump_mean = 0.0
            jump_std = std_return * 3.0

            drift = (mean_return - 0.5 * std_return**2)
            diffusion = std_return * np.random.standard_t(df=t_df, size=num_simulations)
            jump_mask = np.random.rand(num_simulations) < jump_lambda
            jump_sizes = np.random.normal(jump_mean, jump_std, size=num_simulations)
            jump_component = np.zeros(num_simulations)
            jump_component[jump_mask] = jump_sizes[jump_mask]
            simulated_log_returns = drift + diffusion + jump_component
            simulated_prices = current_price * np.exp(simulated_log_returns)

            mean_price = np.mean(simulated_prices)
            median_price = np.median(simulated_prices)
            percentiles = np.percentile(simulated_prices, [2.5, 5, 25, 50, 75, 95, 97.5])
            pi_95 = [percentiles[0], percentiles[-1]]
            pi_90 = [percentiles[1], percentiles[-2]]
            pi_50 = [percentiles[2], percentiles[4]]
            VaR_95_price = percentiles[1]
            VaR_95_value = current_price - VaR_95_price
            losses_beyond_var = simulated_prices[simulated_prices <= VaR_95_price]
            CVR_95_price = np.mean(losses_beyond_var) if len(losses_beyond_var) > 0 else VaR_95_price
            CVaR_95_value = current_price - CVR_95_price
            target_price = current_price * (1 + target_profit_percent)
            probability_of_gain = np.mean(simulated_prices >= target_price)
            
            self.simulation_results = {
                'simulation_model': 'Phase1_Student-t_JumpDiffusion',
                'num_simulations': num_simulations,
                'current_price': current_price,
                'distribution_summary': {'mean_price': mean_price, 'median_price': median_price},
                'prediction_interval_50': pi_50,
                'prediction_interval_90': pi_90,
                'prediction_interval_95': pi_95,
                'risk_metrics': {
                    'VaR_95_price': VaR_95_price,
                    'VaR_95_value': VaR_95_value,
                    'CVaR_95_price': CVR_95_price,
                    'CVaR_95_value': CVaR_95_value,
                },
                'probability_of_gain': probability_of_gain,
                'raw_simulated_prices': simulated_prices[:100]
            }
            
            # (Sanitize Phase 1 results as well)
            return _sanitize_results_for_json(self.simulation_results)
            
        except Exception as e:
            print(f"❌ خطأ فادح في محاكاة مونت كارلو (Phase 1): {e}")
            traceback.print_exc()
            self.simulation_results = {'error': f'Phase 1 MC Error: {str(e)}'}
            return None

    async def generate_1h_distribution_advanced(self, ohlcv_data, target_profit_percent=0.005):
        """
        (المرحلة 2+3 - متقدمة)
        """
        try:
            if not ohlcv_data or '1h' not in ohlcv_data or len(ohlcv_data['1h']) < 50:
                self.simulation_results = {'error': 'Advanced MC requires 1h data (>= 50 candles)'}
                return await self.generate_1h_price_distribution(ohlcv_data, target_profit_percent)

            candles = ohlcv_data['1h']
            df = pd.DataFrame(candles, columns=['timestamp', 'open', 'high', 'low', 'close', 'volume'])
            df[['open', 'high', 'low', 'close', 'volume']] = df[['open', 'high', 'low', 'close', 'volume']].astype(float)
            df['timestamp'] = pd.to_datetime(df['timestamp'], unit='ms')
            df.set_index('timestamp', inplace=True)
            df.sort_index(inplace=True)

            if df.empty or len(df) < 50:
                raise ValueError("DataFrame creation failed or insufficient data after processing")
            
            current_price = df['close'].iloc[-1]
            df['log_returns'] = np.log(df['close'] / df['close'].shift(1)).fillna(0)
            log_returns_series = df['log_returns'].replace([np.inf, -np.inf], 0)

            # 3. (Phase 2) توقع التقلب باستخدام GARCH(1,1)
            try:
                # 🔴 --- START OF CHANGE (FIX GARCH Warning) --- 🔴
                # (Rescale by 100, and set rescale=False to stop GARCH from auto-scaling)
                garch_model = arch_model(log_returns_series * 100, vol='Garch', p=1, q=1, dist='t', rescale=False)
                res = garch_model.fit(update_freq=0, disp='off') 
                forecast = res.forecast(horizon=1)
                # (Divide by 100^2 = 10000)
                forecasted_var = forecast.variance.iloc[-1, 0] / (100**2)
                forecasted_std_return = np.sqrt(forecasted_var)
                # 🔴 --- END OF CHANGE --- 🔴
            except Exception as garch_err:
                forecasted_std_return = np.std(log_returns_series.iloc[-30:])
                print(f"⚠️ GARCH failed, using std: {garch_err}")


            # 4. (Phase 3) توقع الميل (Drift) باستخدام LightGBM
            try:
                if ta:
                    df['rsi'] = ta.rsi(df['close'], length=14)
                    macd = ta.macd(df['close'], fast=12, slow=26, signal=9)
                    df['macd_hist'] = macd['MACDh_12_26_9']
                else: 
                    delta = df['close'].diff()
                    gain = (delta.where(delta > 0, 0)).rolling(window=14).mean()
                    loss = (-delta.where(delta < 0, 0)).rolling(window=14).mean()
                    rs = gain / (loss + 1e-9) # (Added 1e-9 to prevent zero division)
                    df['rsi'] = 100 - (100 / (1 + rs))
                    df['macd_hist'] = df['close'].ewm(span=12).mean() - df['close'].ewm(span=26).mean()

                df['lag_1'] = df['log_returns'].shift(1)
                df['lag_2'] = df['log_returns'].shift(2)
                
                features = ['rsi', 'macd_hist', 'lag_1', 'lag_2']
                df.dropna(inplace=True)

                if df.empty or len(df) < 20:
                     raise ValueError("Insufficient data after feature engineering")

                df['target'] = df['log_returns'].shift(-1)
                df.dropna(inplace=True)

                X = df[features]
                y = df['target']
                
                X_train, y_train = X.iloc[:-1], y.iloc[:-1]
                X_predict = X.iloc[-1:]

                lgbm_model = lgb.LGBMRegressor(n_estimators=100, learning_rate=0.1, n_jobs=1, verbose=-1)
                lgbm_model.fit(X_train, y_train)
                
                forecasted_mean_return = lgbm_model.predict(X_predict)[0]
            
            except Exception as lgbm_err:
                forecasted_mean_return = np.mean(log_returns_series.iloc[-30:])
                print(f"⚠️ LGBM failed, using mean: {lgbm_err}")

            # 5. تشغيل المحاكاة بالقيم الديناميكية
            num_simulations = 5000
            t_df = 10
            jump_lambda = 0.05
            jump_mean = 0.0
            jump_std = forecasted_std_return * 3.0 
            
            mean_return = forecasted_mean_return
            std_return = forecasted_std_return

            drift = (mean_return - 0.5 * std_return**2)
            diffusion = std_return * np.random.standard_t(df=t_df, size=num_simulations)
            jump_mask = np.random.rand(num_simulations) < jump_lambda
            jump_sizes = np.random.normal(jump_mean, jump_std, size=num_simulations)
            jump_component = np.zeros(num_simulations)
            jump_component[jump_mask] = jump_sizes[jump_mask]
            
            simulated_log_returns = drift + diffusion + jump_component
            simulated_prices = current_price * np.exp(simulated_log_returns)

            # 6. حساب المخرجات والتوزيع
            mean_price = np.mean(simulated_prices)
            median_price = np.median(simulated_prices)
            percentiles = np.percentile(simulated_prices, [2.5, 5, 25, 50, 75, 95, 97.5])
            pi_95 = [percentiles[0], percentiles[-1]]
            pi_90 = [percentiles[1], percentiles[-2]]
            pi_50 = [percentiles[2], percentiles[4]]
            VaR_95_price = percentiles[1]
            VaR_95_value = current_price - VaR_95_price
            losses_beyond_var = simulated_prices[simulated_prices <= VaR_95_price]
            CVR_95_price = np.mean(losses_beyond_var) if len(losses_beyond_var) > 0 else VaR_95_price
            CVaR_95_value = current_price - CVR_95_price
            target_price = current_price * (1 + target_profit_percent)
            probability_of_gain = np.mean(simulated_prices >= target_price)
            
            self.simulation_results = {
                'simulation_model': 'Phase2_GARCH_LGBM',
                'num_simulations': num_simulations,
                'current_price': current_price,
                'forecasted_drift_lgbm': forecasted_mean_return,
                'forecasted_vol_garch': forecasted_std_return,
                'distribution_summary': {'mean_price': mean_price, 'median_price': median_price},
                'prediction_interval_50': pi_50,
                'prediction_interval_90': pi_90,
                'prediction_interval_95': pi_95,
                'risk_metrics': {
                    'VaR_95_price': VaR_95_price,
                    'VaR_95_value': VaR_95_value,
                    'CVaR_95_price': CVR_95_price,
                    'CVaR_95_value': CVaR_95_value,
                },
                'probability_of_gain': probability_of_gain,
                'raw_simulated_prices': simulated_prices[:100]
            }
            
            # (Sanitize the results before returning)
            return _sanitize_results_for_json(self.simulation_results)
            
        except Exception as e:
            print(f"❌ خطأ فادح في محاكاة مونت كارلو المتقدمة (GARCH/LGBM): {e}")
            traceback.print_exc()
            self.simulation_results = {'error': f'Advanced MC Error: {str(e)}'}
            # العودة إلى المرحلة 1 في حالة الفشل الفادح
            return await self.generate_1h_price_distribution(ohlcv_data, target_profit_percent)

    def _calculate_trend_adjustment(self, closes):
        """(غير مستخدمة حالياً)"""
        try:
            if len(closes) < 10: return 1.0
            recent_trend = (closes[-1] - closes[-10]) / closes[-10]
            if recent_trend > 0.02: return 1.2
            elif recent_trend > 0.01: return 1.1
            elif recent_trend < -0.02: return 0.8
            elif recent_trend < -0.01: return 0.9
            else: return 1.0
        except Exception: return 1.0

print("✅ ML Module: Advanced Monte Carlo Analyzer loaded (FIXED: NumPy 2.0 & GARCH Scale v2)")