tika/trader-ml
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: MLStrategyModel — encodage labels [-1,0,1] → [0,1,2] pour XGBoost ≥ 2.x

Browse Source 
XGBoost v2+ exige des classes dans [0, n_classes-1].
- train() : y_enc = y + 1 avant fit()
- predict() : décodage pred_enc - 1, prob_map ajusté (0=SHORT, 1=NEUTRAL, 2=LONG)
- _walk_forward_eval() : mask NEUTRAL corrigé (1 au lieu de 0 en espace encodé)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Tika
2026-03-10 19:47:00 +00:00
parent d245d7d8f4
commit 7af7248b4d
1 changed files with 16 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/ml/ml_strategy_model.py
View File

@@ -156,19 +156,23 @@ class MLStrategyModel:
logger.info(f" {len(X)} échantillons, {len(self.feature_names)} features") logger.info(f" {len(X)} échantillons, {len(self.feature_names)} features")
logger.info(f" Distribution : LONG={( y==1).sum()}, SHORT={(y==-1).sum()}, NEUTRAL={(y==0).sum()}") logger.info(f" Distribution : LONG={( y==1).sum()}, SHORT={(y==-1).sum()}, NEUTRAL={(y==0).sum()}")
# Encodage labels [-1,0,1] → [0,1,2] (requis par XGBoost ≥ 2.x)
y_enc = y + 1
# 3. Walk-forward cross-validation (3 folds temporels) # 3. Walk-forward cross-validation (3 folds temporels)
wf_metrics = self._walk_forward_eval(X, y, n_splits=3) wf_metrics = self._walk_forward_eval(X, y_enc, n_splits=3)
# 4. Entraînement sur la totalité des données # 4. Entraînement sur la totalité des données
self.scaler = StandardScaler() self.scaler = StandardScaler()
X_scaled = self.scaler.fit_transform(X) X_scaled = self.scaler.fit_transform(X)
self.model = self._build_model() self.model = self._build_model()
self.model.fit(X_scaled, y) self.model.fit(X_scaled, y_enc)
self.is_trained = True self.is_trained = True
# 5. Évaluation finale (in-sample — indicative) # 5. Évaluation finale (in-sample — indicative)
y_pred = self.model.predict(X_scaled) y_pred_enc = self.model.predict(X_scaled)
y_pred = y_pred_enc - 1 # décodage [0,1,2] → [-1,0,1]
report = classification_report(y, y_pred, labels=[-1, 0, 1], report = classification_report(y, y_pred, labels=[-1, 0, 1],
target_names=['SHORT', 'NEUTRAL', 'LONG'], target_names=['SHORT', 'NEUTRAL', 'LONG'],
output_dict=True, zero_division=0) output_dict=True, zero_division=0)
@@ -235,24 +239,25 @@ class MLStrategyModel:
last = last[self.feature_names].fillna(0) last = last[self.feature_names].fillna(0)
X_scaled = self.scaler.transform(last) X_scaled = self.scaler.transform(last)
pred = self.model.predict(X_scaled)[0] pred_enc = self.model.predict(X_scaled)[0]
pred = int(pred_enc) - 1 # décodage [0,1,2] → [-1,0,1]
# Probabilités si disponibles # Probabilités si disponibles
probas = {'long': 0.0, 'short': 0.0, 'neutral': 1.0} probas = {'long': 0.0, 'short': 0.0, 'neutral': 1.0}
confidence = 0.0 confidence = 0.0
if hasattr(self.model, 'predict_proba'): if hasattr(self.model, 'predict_proba'):
proba_arr = self.model.predict_proba(X_scaled)[0] proba_arr = self.model.predict_proba(X_scaled)[0]
classes = list(self.model.classes_) classes = list(self.model.classes_) # [0, 1, 2] encodés
prob_map = {c: p for c, p in zip(classes, proba_arr)} prob_map = {c: p for c, p in zip(classes, proba_arr)}
probas = { probas = {
'long': float(prob_map.get(1, 0.0)), 'long': float(prob_map.get(2, 0.0)), # encodé 2 = LONG (1)
'short': float(prob_map.get(-1, 0.0)), 'short': float(prob_map.get(0, 0.0)), # encodé 0 = SHORT (-1)
'neutral': float(prob_map.get(0, 1.0)), 'neutral': float(prob_map.get(1, 1.0)), # encodé 1 = NEUTRAL (0)
} }
confidence = float(max(probas['long'], probas['short'])) confidence = float(max(probas['long'], probas['short']))
return { return {
'signal': int(pred), 'signal': pred,
'confidence': confidence, 'confidence': confidence,
'probas': probas, 'probas': probas,
'tradeable': confidence >= self.min_confidence and pred != 0, 'tradeable': confidence >= self.min_confidence and pred != 0,
@@ -383,7 +388,8 @@ class MLStrategyModel:
acc = (y_pred == y_te.values).mean() acc = (y_pred == y_te.values).mean()
# Précision/Recall sur les signaux directionnels uniquement # Précision/Recall sur les signaux directionnels uniquement
mask = (y_te != 0) | (y_pred != 0) # y encodé : 0=SHORT, 1=NEUTRAL, 2=LONG → NEUTRAL=1
mask = (y_te != 1) | (y_pred != 1)
prec, rec, _, _ = precision_recall_fscore_support( prec, rec, _, _ = precision_recall_fscore_support(
y_te[mask], y_pred[mask], average='macro', zero_division=0 y_te[mask], y_pred[mask], average='macro', zero_division=0
) if mask.sum() > 0 else (0, 0, 0, 0) ) if mask.sum() > 0 else (0, 0, 0, 0)