The Problem with Fixed Time Horizons

Most backtesting frameworks suffer from a fundamental flaw: they assume we hold positions for a fixed time horizon. Buy on Monday, sell on Friday. Label the return as positive or negative. Repeat.

This approach ignores how traders actually operate. We don’t set a calendar reminder to exit positions. We set barriers:

  • Profit target: “If this hits +10%, I’m out”
  • Stop loss: “If this drops -5%, I cut it”
  • Time limit: “If nothing happens in 20 days, I reallocate”

The first barrier touched determines the outcome. This is the Triple-Barrier Method, introduced by Marcos Lopez de Prado in Advances in Financial Machine Learning (2018).

The Mathematics of Barriers

Consider a position entered at price $P_0$ with daily volatility $\sigma$. The three barriers are:

Upper barrier (profit taking): \(P_{upper} = P_0 \times (1 + \tau_{profit} \cdot \sigma)\)

Lower barrier (stop loss): \(P_{lower} = P_0 \times (1 - \tau_{stop} \cdot \sigma)\)

Vertical barrier (time limit): \(t_{max} = t_0 + \Delta t\)

Where:

  • $\tau_{profit}$ and $\tau_{stop}$ are volatility multipliers (typically 1.5-3.0)
  • $\sigma$ is the rolling daily volatility
  • $\Delta t$ is the maximum holding period

The label is determined by which barrier is touched first:

  • Upper touched first: Label = 1 (win)
  • Lower touched first: Label = -1 (loss)
  • Vertical touched first: Label = 0 (neutral)

Why This Matters

The Fixed-Horizon Bias

Imagine two trades:

  1. Trade A: Enters at $100, rallies to $110 by day 3, falls back to $105 by day 20
  2. Trade B: Enters at $100, drops to $95 by day 3, recovers to $105 by day 20

With a fixed 20-day horizon, both have +5% returns. Same label. Same “success.”

But this is nonsense. Trade A hit +10% — a disciplined trader would have taken profits. Trade B hit -5% — a risk-managed trader would have been stopped out.

The triple-barrier method captures this nuance:

  • Trade A touches upper barrier at day 3 → Label 1, +10% return
  • Trade B touches lower barrier at day 3 → Label -1, -5% return

Volatility-Adjusted Barriers

Static barriers ($P_0 \pm 5\%$) fail in changing regimes. In high volatility, they’re too tight. In low volatility, too loose.

The solution: dynamic barriers scaled by realized volatility.

\[\text{Barrier Width} = \tau \cdot \sigma_{20d}\]

This means:

  • In calm markets ($\sigma = 0.5\%$ daily): Barriers at $\pm 1\%$ (tight)
  • In volatile markets ($\sigma = 2\%$ daily): Barriers at $\pm 4\%$ (wide)

The barrier width adapts to market conditions, preventing false stops in turbulence and capturing meaningful moves in calm periods.

Implementation

Configuration Presets

Different strategies need different barrier configurations:

Conservative (tight stops, quick exits):

  • $\tau_{profit} = 1.5$
  • $\tau_{stop} = 1.0$
  • $\Delta t = 10$ bars

Symmetric (balanced):

  • $\tau_{profit} = 2.0$
  • $\tau_{stop} = 2.0$
  • $\Delta t = 20$ bars

Aggressive (wide stops, patient):

  • $\tau_{profit} = 3.0$
  • $\tau_{stop} = 2.5$
  • $\Delta t = 30$ bars

The Algorithm

def apply_triple_barrier(price_series, entry_idx, daily_vol, config):
    entry_price = price_series[entry_idx]
    
    # Calculate barriers
    upper = entry_price * (1 + config.profit_take_std * daily_vol)
    lower = entry_price * (1 - config.stop_loss_std * daily_vol)
    vertical_idx = entry_idx + config.max_holding
    
    # Walk forward until a barrier is touched
    for i in range(entry_idx + 1, vertical_idx + 1):
        price = price_series[i]
        
        if price >= upper:
            return Label(barrier=UPPER, return=(price/entry_price - 1))
        
        if price <= lower:
            return Label(barrier=LOWER, return=(price/entry_price - 1))
    
    # Vertical barrier touched
    exit_price = price_series[vertical_idx]
    return Label(barrier=VERTICAL, return=(exit_price/entry_price - 1))

Results on Real Data

I applied the triple-barrier method to mean-reversion signals on SPY, QQQ, and GLD over 6 months:

Configuration Win Rate Avg Return Stop-Loss Rate
Conservative 62% +0.8% 18%
Symmetric 55% +1.2% 25%
Aggressive 48% +1.8% 31%

Insight: Conservative configs have higher win rates but smaller average wins. Aggressive configs win less often but capture larger moves. The “optimal” configuration depends on your utility function — are you optimizing for Sharpe ratio, win rate, or total return?

Barrier Distribution

For the symmetric configuration on SPY:

  • Upper touches: 35% (profit targets hit)
  • Lower touches: 25% (stop losses hit)
  • Vertical touches: 40% (time expiration)

The high vertical percentage (40%) suggests that many signals don’t generate strong directional moves within the holding period. This is valuable information — it tells us the signal’s “half-life” and helps calibrate position sizing.

Applications Beyond Labeling

Meta-Labeling

Lopez de Prado’s meta-labeling framework uses the triple-barrier method as a first step:

  1. Primary model: Predicts direction (long/short) using features $X$
  2. Secondary model: Predicts whether to take the trade using the triple-barrier outcome as label

The secondary model learns when the primary model’s predictions are reliable. This filters out low-confidence trades and improves risk-adjusted returns.

Position Sizing

The barrier distribution informs position sizing:

  • If stop-loss rate > 30%: Reduce position size or tighten filters
  • If vertical rate > 50%: Consider shorter holding periods or momentum filters
  • If win rate < 45%: Re-evaluate signal quality

Strategy Comparison

The triple-barrier method enables fair comparison of strategies with different holding periods. A scalping strategy (avg hold: 2 hours) and a trend-following strategy (avg hold: 20 days) can be compared on equal footing — by their barrier-touch distributions, not just raw returns.

Limitations and Caveats

Path Dependency

The triple-barrier method captures the first barrier touched, but what about the path? A position that oscillates wildly before hitting a barrier has different risk characteristics than one that moves monotonically.

Mitigation: Track path statistics (max adverse excursion, time in drawdown) alongside barrier labels.

Volatility Estimation

The method depends on accurate volatility estimates. In rapidly changing regimes, realized volatility may lag implied volatility.

Mitigation: Use shorter volatility windows (10-15 days) or blend realized and implied vol where available.

Look-Ahead Bias

When labeling historical data, we must use volatility estimates available at the entry time, not future volatility.

Critical: Always calculate barriers using rolling volatility from data strictly before the entry point.

The Deeper Principle

The triple-barrier method embodies a fundamental truth about trading: outcomes are path-dependent and multi-dimensional.

Fixed-time labels reduce a complex path through price-time space to a single number. The triple-barrier method preserves more information:

  • Which barrier was touched (profit vs loss vs time)
  • When it was touched (speed of move)
  • How the barriers were configured (risk tolerance)

This richer representation enables better machine learning models, more realistic backtests, and ultimately, more robust strategies.

Implementation in Almost Surely Profitable

I’ve integrated the triple-barrier method into our backtesting framework:

from backtest.triple_barrier import BarrierConfig, label_events

# Define configuration
config = BarrierConfig.symmetric()

# Label events
labels = label_events(prices, event_times, config=config)

# Analyze results
stats = analyze_barrier_distribution(labels)

The module supports:

  • Volatility-adjusted barriers
  • Multiple preset configurations
  • Barrier distribution analysis
  • Integration with existing data pipelines

See src/backtest/triple_barrier.py for the full implementation.

Conclusion

Backtesting is only as good as the labels we assign. Fixed-time horizons create unrealistic assumptions and biased performance metrics.

The triple-barrier method provides a more realistic labeling framework that aligns with how traders actually operate: setting profit targets, stop losses, and time limits — and reacting to whichever comes first.

In probability terms, the triple-barrier method captures the first passage time distribution of price processes through dynamic boundaries. This is a richer, more accurate representation than point estimates at fixed times.

Almost surely, better labels lead to better models.

Implementation: almost-surely-profitable
Reference: Lopez de Prado, M. (2018). Advances in Financial Machine Learning.