In market prediction, overfitting happens when a model memorizes historical sequences instead of learning patterns that generalize. High accuracy on historical data can therefore exaggerate what the model will achieve on new data. section 10.5.1
Treat impressive accuracy claims cautiously: look-ahead bias, cherry-picked periods, and class imbalance can inflate results independently of a model’s predictive value. Define the model’s intended operating environment and compare performance with a meaningful baseline. section 10.5.1
For a random forest trained on noisy financial data, constrain complexity to reduce the chance of fitting noise:
- Limit tree depth and require more samples per leaf and split to discourage narrow, noise-driven rules. section 10.5.5
- Restrict feature sampling to encourage model diversity. section 10.5.5
- Evaluate predictions on out-of-sample time-series folds; regard the result as predictive evidence, not proof of trading profitability. section 10.5.5
Listing 10.9 Final model training with conservative parameters
from sklearn.ensemble import RandomForestClassifier
print("=== FINAL MODEL TRAINING ===")
# Parameters selected through grid search (code omitted for space constraints)
best_params = {
'n_estimators': 150, #1
'max_depth': 6, #2
'min_samples_leaf': 20, #3
'max_features': 0.2, #4
'min_samples_split': 30 #5
}
print(f"Selected parameters: {best_params}")
# Collect predictions across all time-series folds
all_predictions = []
all_actuals = []
for fold, split in enumerate(splits):
X_train = features.iloc[split['train_idx']]
X_test = features.iloc[split['test_idx']]
y_train = target.iloc[split['train_idx']]
y_test = target.iloc[split['test_idx']]
# Create fresh model for each fold to ensure independence
model = RandomForestClassifier( #6
**best_params,
random_state=42 + fold, #7
class_weight='balanced',
n_jobs=-1
)
model.fit(X_train, y_train) #8
predictions = model.predict(X_test)
all_predictions.extend(predictions) #9
all_actuals.extend(y_test)
print(f"Fold {fold+1}: {len(predictions)} predictions generated")
print(f"\nTotal predictions collected: {len(all_predictions)}")
The final random-forest training example shows conservative parameter settings and predictions collected across time-series folds. section 10.5.5
A modest signal still needs scrutiny
A final market-direction model reports 52.2% directional accuracy on qualifying predictions. The result suggests a modest edge, but does not account for transaction costs, slippage, or capital constraints. section 10.5.5
