Triple Barrier Method

Every observation is assigned a label based on the first barrier touched:

• Upper Barrier (Profit-Taking): A horizontal line at a fixed return or volatility-scaled threshold above the entry price. Touching this generates a +1 label.
• Lower Barrier (Stop-Loss): A horizontal line at a fixed return or volatility-scaled threshold below the entry price. Touching this generates a -1 label.
• Vertical Barrier (Time Horizon): A fixed number of bars or calendar days after entry. If neither price barrier is hit first, the return at this expiration determines a 0 label or a continuous value.

The method captures path dependency—it matters how price reached its final level, not just where it ended.

Static dollar or percentage thresholds fail across different volatility regimes. The standard approach scales barriers using a rolling exponential moving standard deviation:

• Compute the daily volatility estimate from recent returns.
• Multiply by a constant factor (e.g., 2x) to set the upper and lower barrier distances.
• This ensures labels are regime-adaptive—a 1% move in a calm market may be as significant as a 3% move in a turbulent one.

The result is a labeling schema that normalizes for time-varying market conditions, preventing the model from learning spurious patterns tied to volatility clusters.

The algorithm proceeds observation-by-observation through historical data:

1. Define entry point at time t with price P_t.
2. Set upper barrier at P_t × (1 + threshold) and lower barrier at P_t × (1 - threshold).
3. Set vertical barrier at t + h bars.
4. Monitor forward from t+1 to t+h: if the high touches the upper barrier first, label = +1; if the low touches the lower barrier first, label = -1; if the vertical barrier expires first, label = sign of return at t+h or 0.

This produces a three-class categorical label or a continuous return value, depending on the modeling objective.

The vertical barrier prevents infinite holding periods and defines a maximum trade duration. Its treatment varies by use case:

• Binary Classification: Observations that hit the vertical barrier first are often discarded to avoid ambiguous labels, or assigned a 0 class.
• Regression: The return at the vertical barrier is used directly as a continuous target variable.
• Meta-Labeling: A secondary model predicts the probability of a +1 given that the primary model already entered a position.

The choice impacts sample size and label purity. Discarding time-barrier observations reduces noise but shrinks the training set.

Traditional fixed-horizon labeling assigns a class based solely on the return at time t+h, ignoring the path taken. The Triple Barrier Method offers distinct advantages:

• Captures stop-loss logic: Real trading strategies exit at predetermined loss levels, not arbitrary future dates.
• Reduces label noise: A position that briefly spikes to a profit target and then reverses is correctly labeled +1, whereas a fixed-horizon label might call it -1.
• Aligns with execution: The labels mirror how a strategy would actually be traded, improving the signal-to-noise ratio in training data.
• Enables asymmetric payoffs: The barrier distances can be set asymmetrically (e.g., 2% profit target, 1% stop-loss) to reflect desired risk-reward ratios.

The Triple Barrier Method is a foundational component of meta-labeling, a two-stage modeling approach:

• Primary Model: Generates a trading signal (long/short) based on any strategy.
• Secondary Model (Meta-Labeler): Trained on Triple Barrier labels to predict the probability of a successful outcome given the primary model's entry.

This separation allows the meta-labeler to act as a position-sizing filter—reducing size or skipping trades when the predicted success probability is low. It decouples the what to trade decision from the whether to trade decision, improving risk-adjusted returns.