Triple Barrier Labeling

Triple-barrier labeling

A fixed-horizon label tells the model that a trade which fell through its stop and recovered was a winner. Advances in Financial Machine Learning (Lopez de Prado 2018), chapter 3. The label is the target the model optimises; if the label is about a path you could not have held, the model learns to predict something you cannot trade, and no amount of cross-validation catches it because the labels and the backtest agree with each other.

Every number below is printed by scripts/triple_barrier.py (numpy only, seed 0, 1.2 s). The test series is 6,000 bars of clustered volatility (log-vol AR(1), phi = 0.99) with zero drift, so every asymmetry below is geometry, not alpha. Bars are integer indices; the book keys everything by DatetimeIndex, which changes none of the arithmetic.

⚠️ Snippet and page numbers below are as cited in mlfinpy 0.1.2's own docstrings; what is verified here is the code, not the book.

1. The three barriers, and why two of them must be volatility-scaled

Per event t0: a profit-taking barrier at +pt * trgt, a stop-loss barrier at -sl * trgt, and a vertical barrier vert bars later. The label is which one is touched first.

✅ Source-verified against mlfinpy 0.1.2 (mlfinpy/labeling/labeling.py, mlfinpy/util/volatility.py):

• triple_barriers (AFML Snippet 3.2, page 45) computes cum_returns = (closing_prices / close[loc] - 1) * events_.at[loc, "side"] — arithmetic returns, signed by the side — and takes cum_returns[cum_returns > profit_taking[loc]].index.min() for the touch. Strict inequality, first index, so a bar that lands exactly on the barrier does not count.
• get_daily_vol (Snippet 3.1, page 44) is (close / close.shift(1) - 1).ewm(span=lookback).std() with lookback = 100 — arithmetic returns, pandas' adjusted EWM, unbiased standard deviation. ✅ The script's numpy re-implementation matches pandas' ewm(span=100).std() to 9.7e-17.
• get_events (Snippet 3.6, page 50) applies target = target[target > min_ret] before anything else: 🚨 min_ret silently deletes every event whose target volatility is below it. That is a sample filter that removes calm periods, not a label parameter.

Why the target must be the local volatility. ✅ Measured, 1,136 events, pt_sl = (4, 4), 20-bar vertical barrier, split into volatility terciles by the EWMA target at the event:

🚨 With one fixed barrier width, the meaning of the label changes with the regime: in the quiet third of the sample 71 % of events time out and in the noisy third only 22 % do. The model learns "class 0 means low volatility", which is a fact about the labeller, not about the future. The realised per-bar volatility here spans 40x (0.0018 to 0.0717), which is not extreme for real data.

2. What the labels look like

✅ Same 1,136 events, same 20-bar horizon:

The sign-only version is the one most people write, and it has no neutral class at all: every event is a directional bet, including the 41 % that the barrier method says never went anywhere.

3. 🚨 How often the horizon label is wrong about the path

The disagreement is not a rounding error. ✅ Measured, sign-only fixed-horizon labels against triple-barrier labels over the identical 20-bar horizon and the identical barriers:

• the two labels disagree on 535 of 1,136 events (47.1 %);
• 33 events (2.9 %) are called +1 by the horizon although the 4-sigma stop was crossed on the way; 35 (3.1 %) are called -1 although the profit target was crossed;
• 68 events (6.0 %) describe a path you could not have held.

🚨 And 6.0 % is the number for a very loose stop. Tighten it and the horizon labels fall apart:

At a 1-sigma stop — a perfectly ordinary risk limit — 44 % of the training labels describe trades that no longer existed by the horizon.

What that costs in the backtest. Take the ceiling: a perfect model of the fixed-horizon labels, which is the best any model trained on them can be.

That 21 % is not a modelling error, a cost assumption or a slippage estimate. It is profit the backtest booked from drawdowns the risk limits would not have let you hold.

4. 🚨 The drawdown the horizon label assumes you sat through

✅ Among the 608 events the sign-only label calls +1, the worst drawdown before the horizon:

44.4 % of these "winners" went through a 1-sigma stop first, and 5.4 % through a 4-sigma stop. The median winner spent part of its life 0.8 sigma under water. Every one of those is a positive training example.

5. 🚨 pt_sl sets the class balance, and the class balance is not evidence

✅ Same events, same zero-drift series, only the barrier multiples change:

🚨 A wide stop with a tight target produces 70 % positive labels on a series with no drift. If you report "the model predicts up moves 70 % of the time", you are reporting pt_sl. Class balance, base rate and any accuracy figure derived from them are properties of the barrier geometry. pt_sl and vert are researcher degrees of freedom: record them in ../../../fin-core/skills/backtest-validation/scripts/trial_ledger.py.

6. 🚨 Snippet 3.2 touches on arithmetic returns, Snippet 3.9 relabels on log returns

✅ Source-verified in mlfinpy 0.1.2. triple_barriers decides the touch with the arithmetic return against pt * trgt. barrier_touched (Snippet 3.9, page 55), called from get_bins, then re-derives the label from out_df["ret"], which get_bins builds as a log return (np.log(prices.loc[t1]) - np.log(prices.loc[t0])), and tests it against np.log(1 + target) * events.loc[date_time, "pt"].

Those two thresholds are equal only at pt = 1. ✅ Measured on the same events:

At target 2 % and pt = 2, the touch fires at r > 0.0400 and the relabelling needs r > 0.0404; a touch inside that band is recorded as a vertical-barrier label. Small, silent, and it grows with the multiple. If you use pt_sl above 1, decide which convention you want and apply it in both places.

7. Traps

• 🚨 Fixed-horizon labels. Sections 3-4. If you must use them, compute the maximum adverse excursion alongside and drop or reweight the ones that breached your stop.
• 🚨 A fixed barrier width across regimes. Section 1. The label's meaning drifts with volatility, and the model learns the regime.
• 🚨 min_ret is a sample filter. get_events drops events, it does not just skip barriers. Whatever it removes is missing from the training set and from every count you report.
• 🚨 t1 is the touch time, not the vertical barrier. Everything downstream depends on it: purging, embargoing, uniqueness and weights all need the bar where the label actually resolved. Passing the vertical barrier instead over-purges; passing t0 + 1 under-purges and silently reintroduces the leak — ../../../fin-libraries/skills/lib-purgedcv/SKILL.md owns that, and evaluation_times is where the touch time goes.
• 🚨 These labels overlap, so the samples are not i.i.d. A 20-bar horizon sampled every 5 bars means four labels share every bar. Weight them by uniqueness before fitting (../sample-weights-and-uniqueness/SKILL.md) and cross-validate them with purging and an embargo (../../../fin-libraries/skills/lib-purgedcv/SKILL.md). Neither is optional.
• 🚨 The barrier is checked on the close, but you would have been stopped intrabar. This script — and the book's snippet — scan closes. A real stop is a resting order against the low or the high, so the true touch is earlier and worse than what closes report. ../../../fin-core/skills/intraday-microstructure/SKILL.md owns the bar-versus-tick question.
• ⚠️ Accuracy on triple-barrier labels is still not profit. The classes are unbalanced by construction (section 5), and each event's return is a different size. Score with the return-weighted metrics and the barrier returns, not with accuracy_score.
• 🚨 The event sampler is part of the label. Sampling every 5 bars, as here, is a placeholder; the book samples on a CUSUM filter so that events cluster where something happened (../structural-breaks/SKILL.md). Changing the sampler changes the label distribution and the effective sample size.
• 🔴 pip install mlfinlab cannot work — ✅ checked 2026-09-09, the PyPI JSON API returns 404 and the simple index lists zero files. The maintained fork is mlfinpy 0.1.2 (MIT), which ⚠️ pins numpy<1.27. The snippets above were read from its sdist, not from an installed copy.

8. Scripts

• scripts/triple_barrier.py — the causal EWMA target volatility checked against pandas, the first-touch scan, triple-barrier and meta-labels, fixed-horizon labels, the volatility-tercile table, the path-conflict counts and the stop-tightness sweep, maximum adverse excursion, the pt_sl class-balance table, and the Snippet 3.2 / 3.9 return-convention mismatch. numpy only (pandas is used once, inside a function, purely to check the EWMA). Seed 0, 1.2 s.

9. Where this sits

../sample-weights-and-uniqueness/SKILL.md takes the t1 touch times from here and turns them into concurrency, uniqueness and sample weights. ../meta-labeling/SKILL.md takes the side argument of get_events and adds the secondary model. ../structural-breaks/SKILL.md owns the CUSUM filter that should choose the events. ../fractional-differentiation/SKILL.md is the feature side of the same pipeline. ../../../fin-libraries/skills/lib-purgedcv/SKILL.md owns purged and embargoed cross-validation for the overlapping labels this produces — never re-implement it. ../../../fin-core/skills/backtest-validation/SKILL.md owns the trial ledger for pt_sl and vert. ../../../fin-core/skills/signal-construction/SKILL.md owns the causality check on the features that go with these labels.