Background: Triple Barrier Labeling Method

"Traditional up/down labels ignore a key problem: how long can you hold?"

The Problem with Traditional Labels

Suppose you use machine learning to predict whether AAPL will go up or down tomorrow:

Label = 1 if tomorrow's close > today's close else 0

What's the problem?

Traditional labels completely ignore volatility during the holding period and risk management.

Core Idea of Triple Barrier

Set three "barriers," and whichever is triggered first determines the label:

Three Barriers:

1. Upper Barrier: Price rises to take-profit level -> Label = 1
2. Lower Barrier: Price falls to stop-loss level -> Label = -1
3. Vertical Barrier (Time): Holding period ends -> Label = 0 or based on final direction

Detailed Calculation Example

Setup:

• Entry price: $185.00 (AAPL)
• Take-profit: 2% ($188.70)
• Stop-loss: 1% ($183.15)
• Max holding period: 5 days

Scenario 1: Day 2 hits $189.00

Result: Upper barrier triggered
Label: 1 (Profitable exit)
Actual return: +2%

Scenario 2: Day 3 drops to $182.50

Result: Lower barrier triggered
Label: -1 (Stop-loss exit)
Actual return: -1%

Scenario 3: 5 days oscillating between $184-$187

Result: Time barrier triggered
Label: 0 or based on final price
Actual return: Depends on exit price

Dynamic Barriers: Volatility-Based

A fixed 2% take-profit is unreasonable in different market environments:

Solution: Use ATR (Average True Range) or rolling standard deviation to define barriers

Upper Barrier = Entry Price x (1 + k x sigma)
Lower Barrier = Entry Price x (1 - k x sigma)

Where:
  sigma = Past 20 days return standard deviation
  k = Multiplier (typically 1-3)

Example:

• sigma = 1.5%
• k = 2
• Entry price = $185

Upper Barrier = $185 x (1 + 2 x 1.5%) = $190.55
Lower Barrier = $185 x (1 - 2 x 1.5%) = $179.45

Asymmetric Barriers

In real trading, take-profit and stop-loss are often asymmetric:

Trend Strategy Example:

Take-profit = 3 x 1.5% = 4.5%
Stop-loss = 1 x 1.5% = 1.5%

Win rate only needs >25% to be profitable:
Expected value = Win rate x 4.5% - (1-Win rate) x 1.5%
At Win rate = 30%: Expected value = 0.3% (positive return)

Importance of Label Distribution

Triple Barrier produces three types of labels; distribution affects model training:

Handling Class Imbalance:

• Oversample minority class (SMOTE)
• Undersample majority class
• Adjust class weights
• Adjust barrier widths for more uniform distribution

Comparison with Traditional Labels

Multi-Agent Perspective

In multi-agent architecture, Triple Barrier can be applied like this:

Signal Agent: Predict labels (1, 0, -1)
     |
Risk Agent: Dynamically adjust barrier widths
     |
     - High volatility period -> Wider barriers, avoid frequent stop-losses
     - Low volatility period -> Narrower barriers, improve capital efficiency
     |
Execution Agent: Monitor barrier triggers, execute exits

Common Misconceptions

Misconception 1: Wider barriers are always better

Wrong. Wider barriers reduce stop-loss triggers, but:

• Larger losses per trade
• Longer capital lockup
• May miss other opportunities

Misconception 2: Time barrier is not important

Very important. Without a time barrier, capital may be locked up long-term. If 5 days pass without triggering any barrier, the judgment may be wrong, and you should exit to reassess.

Misconception 3: Triple Barrier labels are always better than traditional labels

Not necessarily. If your strategy doesn't use take-profit/stop-loss (e.g., T+1 must sell next day), traditional labels may be more appropriate. Labeling method should match actual trading logic.

Practical Recommendations

1. Start Simple: Use fixed percentage barriers first, add volatility adjustment after validation
2. Check Label Distribution: Adjust when three label types have very uneven ratios
3. Backtest Consistently: If training uses Triple Barrier, backtest should simulate same take-profit/stop-loss logic
4. Consider Trading Costs: Too narrow barriers lead to frequent trading, costs eat returns

Summary