Crash detection

plots/BTC/crash/best_threshold.txt

@@ -0,0 +1 @@
+0.06525423728813559

tools/sklearn/crash_detection.py

@@ -0,0 +1,36 @@
+import pandas as pd
+from sklearn.ensemble import RandomForestClassifier
+from sklearn.model_selection import train_test_split
+
+# ================================
+# 1. PREPARE DATA
+# ================================
+df = pd.read_feather("/home/jerome/Perso/freqtradeDocker/user_data/data/binance/BTC_USDC-1h.feather")
+
+# features
+df['returns'] = df['close'].pct_change()
+df['atr'] = (df['high'] - df['low']).rolling(14).mean()
+df['slope'] = df['close'].rolling(20).mean().diff()
+
+df['drawdown'] = (df['close'] - df['close'].rolling(48).max()) / df['close'].rolling(48).max()
+
+# label : crash si -12% dans les 48h
+future = df['close'].shift(-48)
+df['future_dd'] = (future - df['close']) / df['close']
+df['crash'] = (df['future_dd'] < -0.12).astype(int)
+
+df = df.dropna()
+
+X = df[['returns','atr','slope','drawdown']]
+y = df['crash']
+
+# ================================
+# 2. TRAIN MODEL
+# ================================
+X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, shuffle=False)
+
+model = RandomForestClassifier(n_estimators=200)
+model.fit(X_train, y_train)
+
+print("Accuracy:", model.score(X_test, y_test))
+print("Feature importance:", model.feature_importances_)

tools/statistique/crash_risk.py

@@ -0,0 +1,163 @@
+import json
+import pandas as pd
+import numpy as np
+import matplotlib.pyplot as plt
+
+DATA_FILE = "/home/jerome/Perso/freqtradeDocker/user_data/data/binance/BTC_USDC-4h.feather"
+
+
+def compute_crash_risk_index(df):
+    # VOLATILITÉ
+    df['H-L'] = df['high'] - df['low']
+    df['ATR'] = df['H-L'].rolling(14).mean()
+    df['atr_norm'] = df['ATR'] / df['close']
+
+    # DRAWDOWN (critique)
+    df['rolling_max'] = df['close'].rolling(48).max()
+    df['drawdown'] = (df['close'] - df['rolling_max']) / df['rolling_max']
+    df['dd_score'] = np.clip(-df['drawdown'] / 0.10, 0, 1)
+
+    # TENDANCE (slope)
+    df['MA7'] = df['close'].rolling(7).mean()
+    df['MA14'] = df['close'].rolling(14).mean()
+    df['slope'] = df['MA7'] - df['MA14']
+    df['slope_score'] = np.clip(1 - (df['slope'] / df['close']), 0, 1)
+
+    # NÉGATIVE STREAK
+    df['neg_streak'] = df['close'].pct_change().apply(lambda x: min(x, 0)).rolling(24).sum()
+    df['neg_score'] = np.clip(-df['neg_streak'] / 0.05, 0, 1)
+
+    # COMPOSANTS COURT TERME
+    df['pct_change_3'] = df['close'].pct_change(3)
+    df['pct_change_3_smooth'] = df['pct_change_3'].rolling(6).mean()
+    df['crash_score'] = np.clip(1 + (df['pct_change_3_smooth'] / 0.05), 0, 1)
+
+    df['speed'] = df['close'].diff().rolling(6).mean()
+    df['accel'] = df['speed'].diff().rolling(6).mean()
+    df['STD20'] = df['close'].rolling(20).std()
+    df['accel_score'] = np.clip(1 + (df['accel'] / (df['STD20'] + 1e-9)), 0, 1)
+
+    # INDEX FINAL
+    df['crash_raw'] = (
+            0.35 * df['dd_score'] +  # le plus important pour crash lent
+            0.25 * df['neg_score'] +
+            0.20 * df['slope_score'] +
+            0.10 * df['crash_score'] +
+            0.10 * df['accel_score']
+    )
+
+    # LISSAGE SIMPLE
+    df['crash_risk_index'] = df['crash_raw'].ewm(span=24).mean()
+
+    return df
+
+
+#
+# def compute_crash_risk_index(df):
+#     # -- volatilité
+#     df['H-L'] = df['high'] - df['low']
+#     df['ATR'] = df['H-L'].rolling(14).mean()
+#     df['atr_norm'] = df['ATR'] / df['close']
+#
+#     # -- variations lissées pour éviter les spikes
+#     df['pct_change_3'] = df['close'].pct_change(3)
+#     df['pct_change_3_smooth'] = df['pct_change_3'].rolling(6).mean()
+#
+#     # -- speed/accel : on SMOOTH sinon c'est incontrôlable
+#     df['speed'] = df['close'].diff().rolling(6).mean()
+#     df['accel'] = df['speed'].diff().rolling(6).mean()
+#
+#     # -- Bollinger
+#     df['MA20'] = df['close'].rolling(20).mean()
+#     df['STD20'] = df['close'].rolling(20).std()
+#     df['BB_lower'] = df['MA20'] - 2 * df['STD20']
+#
+#     # -------- Scores normalisés & STABLES --------
+#     df['crash_score'] = np.clip(1 + (df['pct_change_3_smooth'] / 0.05), 0, 1)
+#
+#     df['accel_score'] = np.clip(
+#         1 + (df['accel'] / (df['STD20'] + 1e-9)),
+#         0, 1
+#     )
+#
+#     df['boll_score'] = np.clip(
+#         (df['close'] - df['BB_lower']) / (3 * df['STD20']),  # + 3σ au lieu de 2σ
+#         0, 1
+#     )
+#
+#     df['atr_score']  = np.clip(
+#         1 - (df['atr_norm'] / 0.06),  # tolérance plus large
+#         0, 1
+#     )
+#
+#     # -------- COMBINAISON + DOUBLE SMOOTHING --------
+#     df['crash_raw'] = (
+#         0.30 * df['crash_score'] +
+#         0.25 * df['accel_score'] +
+#         0.20 * df['boll_score'] +
+#         0.15 * df['atr_score']
+#     )
+#
+#     # Lissage ultra important pour éviter le 0.3 → 0.9
+#     df['crash_risk_index'] = (
+#         df['crash_raw'].ewm(span=48).mean()   # 2 jours en 1h
+#     ).rolling(24).mean()                      # filtre final
+#
+#     return df
+
+# def compute_crash_risk_index(df):
+#     df['H-L'] = df['high'] - df['low']
+#     df['ATR'] = df['H-L'].rolling(14).mean()
+#     df['atr_norm'] = df['ATR'] / df['close']
+#
+#     df['pct_change_3'] = df['close'].pct_change(3)
+#
+#     df['speed'] = df['close'].diff()
+#     df['accel'] = df['speed'].diff()
+#
+#     df['MA20'] = df['close'].rolling(20).mean()
+#     df['STD20'] = df['close'].rolling(20).std()
+#     df['BB_lower'] = df['MA20'] - 2 * df['STD20']
+#
+#     df['crash_score'] = np.clip(1 + (df['pct_change_3'] / 0.05), 0, 1)
+#     df['accel_score'] = np.clip(1 + (df['accel'] / df['STD20']), 0, 1)
+#     df['boll_score'] = np.clip((df['close'] - df['BB_lower']) / (2 * df['STD20']), 0, 1)
+#     df['atr_score']  = np.clip(1 - (df['atr_norm'] / 0.04), 0, 1)
+#
+#     df['crash_raw'] = (
+#         0.30 * df['crash_score'] +
+#         0.25 * df['accel_score'] +
+#         0.20 * df['boll_score'] +
+#         0.15 * df['atr_score']
+#     )
+#
+#     # Lissage ultra important pour éviter le 0.3 → 0.9
+#     df['crash_risk_index'] = (
+#         df['crash_raw'].ewm(span=4).mean()   # 2 jours en 1h
+#     ).rolling(2).mean()                      # filtre final
+#
+#     return df
+
+# Load Freqtrade OHLCV JSON
+# with open(DATA_FILE, "r") as f:
+#     raw = json.load(f)
+
+df = pd.read_feather(DATA_FILE)
+print(df.head())
+
+
+# df = pd.DataFrame(raw)
+df['date'] = pd.to_datetime(df['date'])
+df = df.set_index('date')
+
+# compute risk
+df = compute_crash_risk_index(df)
+
+# Only keep Nov. 2025
+df = df["2025-10-01":"2025-12-10"]
+
+plt.figure(figsize=(12, 6))
+plt.plot(df.index, df['crash_risk_index'])
+plt.title("Crash Risk Index – Novembre 2025")
+plt.grid(True)
+plt.show()