Triple-Barrier-Labeling

What is the triple-barrier method?

From Marcos López de Prado (Advances in Financial Machine Learning, 2018). A method for labeling the outcome of an event — i.e. the target variable for ML research. For each event timestamp (e.g. a breakout, a CUSUM event, an entry candidate) you set three barriers and check which one is touched first:

1. Upper barrier — take-profit, e.g. +k·ATR above the entry
2. Lower barrier — stop-loss, e.g. −k·ATR below the entry
3. Vertical barrier — time limit, after N bars it's over

for each event at t0:
    upper  = price[t0] * (1 + k_up  * atr_pct[t0])
    lower  = price[t0] * (1 - k_dn  * atr_pct[t0])
    window = prices[t0 .. t0+N]
    label = +1  if upper is hit first   (move realized)
            -1  if lower is hit first   (stopped out)
             0  if the time limit comes first (or sign(return) at the end)

Skinner (Delta Trend Trading) uses its little sister in his quant-process video — double-barrier (without the time limit) — to label the outcomes of his sweeps.

Why it beats 'return after N bars'

Botty's ml/labels.py currently labels with fixed-horizon forward returns. That has three weaknesses which triple-barrier fixes:

1. Path realism. Fixed-horizon ignores that a trade would long since have been stopped out along the way. Triple-barrier asks: "If I had actually traded this with a stop + target — what happens first?" → the labels become consistent with Botty's actual execution (we trade with stops/targets, not 'hold for exactly N bars').
2. Vol scaling. Barriers as ATR/vol multiples adapt to the regime. A fixed % threshold is too tight in high vol (hit immediately) and too wide in low vol (never hit). Consistent with the finding that 'the vol regime is the actual signal'.
3. Discriminative power. The return at an arbitrary horizon N is noisy; 'move realizes before the reversal — yes/no' is a much clearer label for whether a feature has tradable predictive power.

Where it sits in Botty

• ml/labels.py — as a new labeling function alongside the existing forward-return/sign/vol-norm labels. The ml/ module is pattern discovery (event → context features → outcome); triple-barrier is the better outcome.
• Paired with event-based sampling (CUSUM filter): first define events, then label each event via triple-barrier — exactly the 'event-defining + outcome features' pipeline from the quant methodology.
• In ml/experiments/ a feature's discriminative power is then measured against the triple-barrier label instead of the raw forward return.

Honest assessment

• Research-only: it improves how we measure whether something works — it does not directly change the live bot. Belongs to the live-readiness discipline.
• The eval moves with it: the label is categorical ({−1,0,+1}) instead of continuous → the Spearman IC does not carry over 1:1; use classification/quantile metrics or hit rate × payoff instead.
• Overlapping labels: simultaneous events create temporally overlapping label windows → clean ML requires López de Prado's uniqueness/concurrency weighting (otherwise hidden bias from multiply counted, correlated observations).
• Meta-labeling is the related follow-up technique: a second model decides whether to follow a primary signal (sizing/filter) after triple-barrier has supplied the direction.

Status in Botty: implemented in ml/labels.py (triple_barrier_labels + average_uniqueness for the concurrency weighting); pilot pipeline in ml/experiments/cusum_triple_barrier/ (vol-scaled barriers, intrabar via high/low, pilot label split ~49/51 +1/−1).