Al Brooks price action, tested against 3,946 trading days

Al Brooks wrote what is effectively the encyclopedia of price action trading — three volumes and a course that between them describe the intraday market bar by bar. What makes that body of work unusual is that it is full of numbers: 80% of trading-range breakouts fail, 90% of days have an extreme in by the 90-minute mark, 60% of pullbacks to the halfway point test the old high first. Those are falsifiable statements. As far as we can tell, nobody had ever checked them. So we did.

The seven verdicts

Every row covers the same 3,946 days except the two noted underneath. "Traditionally quoted" is the figure as it is taught; "measured" is what 2010–2026 of ES actually did.

The last row is measured within 5, 10 and 20 bars respectively — the claim's truth depends entirely on how long you are willing to wait, which the claim never says. Two rows do not cover all 3,946 days: the opening-range breakout can only be scored on the 3,578 days that produced a breakout at all, and the "80% rule" row counts range-breakout events rather than days, roughly 5,000 to 9,700 depending on the bar window.

What the pattern of errors tells you

Look down the table and the shape of the error is consistent enough to be useful on its own. Where the teaching quotes a big round number — 90%, 80% — the measured rate comes in below it by about seven to ten points. The two earliest checkpoints run the other way: an extreme gets pinned sooner than taught (26.3% after one bar against a quoted 20%), while the session goes on making new ones for longer than taught (81.2% at 90 minutes against a quoted 90%).

A workable calibration for the whole genre falls out of this: trust the direction, discount the round number by about a tenth. That is also, quietly, a compliment. These were people describing real market behaviour from memory, without a database, and they got the sign right in every case. The rules are not wrong. The confidence attached to them is.

Then we trained the framework into a model

Fact-checking claims is one thing. The harder question is whether the underlying framework — the idea that a session has a shape, and that the shape is legible early — survives being turned into something mechanical.

Partly, it does. A 5-class day-type classifier, evaluated only on held-out years, reaches 66% top-1 and 80% top-2 accuracy, with a calibration error under 10%. The number that matters alongside those is the baseline: always guessing the most common day type gets 37%. So the model is worth roughly 29 points over guessing — real, and a long way from clairvoyance.

It is also far weaker when you would most want it. At the 90-minute mark — the point the courses treat as decisive — top-1 accuracy is only 53.5% and top-2 is 68.6%. Most of the model's skill arrives late in the session, which is exactly the opposite of what a forecast is for. That is worth saying plainly rather than quoting the 66% on its own.

The one finding that survived everything

The strongest result in this program is not a claim from the books at all. It came from combining two things that are usually taught separately: using the day-type reading as a filter on a Weis Wave event rather than as a forecast in its own right.

On ES, the long trigger ran at 56.5% under an up-trending day-type read against 50.8% under a down-trending one — a 5.6-point spread, p=0.002. Unconditionally, with no gate at all, the same trigger runs at 51.8%. So most of the value is not in the trigger; it is in knowing when to ignore it.

The part that makes it interesting is what happened next. Applied zero-shot to instruments the day-type model had never seen, with nothing refitted, the split held and got wider: 56.1% against 45.3% on NQ (n=1,536 and 689, p=2e-6), and 54.3% against 44.3% on QQQ (n=1,200 and 592, p=7e-5).

Both numbers in each pair are the same event under opposite day-type readings — they are not a comparison against running the trigger ungated. Ungated is 51.8% on ES and 52.9% on NQ; the ungated QQQ figure did not pass its pre-registered threshold, so it is not quoted here.

Run as one mechanical rule over eight years of QQQ, that gate produced 784 trades at a 55.6% hit rate — a cumulative +61.3% on a $10,000 starting balance, with a maximum drawdown of 12.4% and only 13.4% of the time spent holding anything at all. It now runs in a live paper account with every trade published as it happens — including the losing ones, which is the entire point of publishing it.

A 55.6% hit rate in a backtest is a description of a fixed historical sample. It is not a prediction, it does not include your costs or slippage, and it is not a recommendation to trade. The live journal exists so the claim can fail in public.

Where it did not work

Three failures are worth more than the successes above, because they mark the edges of where any of this applies.

• Crypto has an inverted signature. Of 8 pre-registered signals tested on BTC, 0 confirmed. One of them reversed direction outright. Gold and ETH show the same inconsistency. Whatever these patterns are measuring, it is specific to index futures and does not carry across.
• Chart patterns did not survive at all. Over 121,000+ pattern detections in the Bulkowski style produced an AUC of 0.54 — close enough to a coin flip that the honest conclusion is that this particular formalisation carries no signal.
• The gate does not transfer to the Asian session. Pre-registered before measuring, the conclusion was NO_GO. It works when it works, and the hours it does not work in are named rather than quietly dropped.

Publishing these is not modesty. A study that only reports where a method works cannot tell you anything about where it will fail on you, which is the part you actually need.

What is still unmeasured

The seven rows above are the complete measured set — not the complete library. Of 92 structured claims, 19 are specific enough to test cleanly, and 4 of those have been measured so far (they expand into the seven rows because several claims bundle multiple checkpoints). That leaves 15 high-measurability claims structured and waiting, including some of the most actionable ones in the corpus — the 60% figure attached to entering at a halfway pullback, and the claim that opening breakouts reverse half the time no matter how strong they look.

They will be added to this table when they are measured, with the same rule: the definition is written down before the number is seen, and the result is published whichever way it lands.

Using these readings live

The same measurements run as an API. The price-action endpoint returns a rolling event reading for a symbol — which of five structural events is currently active, with the pre-registered historical rate for that event in that window — and the day-type block returns the 5-class reading described above during the regular session. The methodology page carries the full definitions, and the free tier returns complete responses without an account so you can check the shape of the output before deciding anything.

Common questions

Quant Data. "Al Brooks price action, tested: seven classic claims measured against 3,946 ES trading days." quantdata.uk, 2026-08-08. https://quantdata.uk/research/al-brooks-price-action-tested

@misc{quantdata-brooks-tested,
  author       = {Quant Data},
  title        = {Al Brooks price action, tested: seven classic claims measured against 3,946 ES trading days},
  year         = {2026},
  howpublished = {\url{https://quantdata.uk/research/al-brooks-price-action-tested}},
}

https://quantdata.uk/research/al-brooks-price-action-tested