Advanced Feature

Walk-Forward ML Strategy

Train, re-train, and validate Machine Learning models on rolling historical windows.

The Retraining Cycle

Unlike static backtests, Walk-Forward ML simulates the reality of live trading where models are retrained as new data arrives. RLX automates this cycle by partitioning data into rolling segments where each test window is preceded by a fresh training phase.

1. TRAIN
Fit model on historical N bars
2. TEST
Trade on next M bars (OOS)
3. ROLL
Shift both windows forward

Building ML Strategies

Inherit from BaseMLStrategy and implement the train and generate_signals methods.

Python (sklearn example)
from rlxbt.ml import BaseMLStrategy, WalkForwardML
from sklearn.ensemble import RandomForestRegressor

class MyMLStrategy(BaseMLStrategy):
    def __init__(self):
        super().__init__(model=RandomForestRegressor())

    def train(self, data):
        # data is your training slice
        X, y = self.prepare_features(data)
        self.model.fit(X, y)

    def generate_signals(self, data):
        X, _ = self.prepare_features(data)
        preds = self.model.predict(X)
        # Return signal DataFrame (-1, 0, 1)
        return preds_to_signals

Running the Trainer

The WalkForwardML trainer orchestrates the entire process and consolidates results into a single report.

Python
bt = Backtester(initial_capital=100000)
wf_trainer = WalkForwardML(backtester=bt)

results = wf_trainer.run(
    strategy_class=MyMLStrategy,
    data=df,
    train_size=5000, # Retrain window
    test_size=500,   # Trading window
    step_size=500    # Shift size
)

print(f"Consolidated Return: {results['total_return']:.2%}")

Automated Hyperparameter Tuning

Instead of fixed parameters, you can use Optuna to find the optimal settings for each walk-forward window automatically.

Hyperparameter Tuning with Optuna
optuna_config = {
    "n_trials": 20,
    "param_space": {
        "n_estimators": ("int", 50, 200),
        "max_depth": ("int", 3, 10),
        "lr": ("float", 1e-4, 1e-1, True) # Log scale
    }
}

results = wf_trainer.run(
    strategy_class=MyMLStrategy,
    data=df,
    optuna_config=optuna_config,
    # ... other window params ...
)
Stability Analysis Reinforcement Learning