How we measure — and what we failed to prove
Every indicator you have ever been sold came with a backtest that worked. That is not evidence: it is the natural result of trying rules until one fits, and publishing only that one.
So we did the opposite. We wrote down how we would measure, before we looked, and we are publishing everything the measurement produced — including the parts that killed our own ideas.
This page is long on purpose. Skim the tables if you like; the point is that the work exists and you can check it.
The rule we set before looking
Four commitments, fixed in writing before any number was computed:
- One observation per day. An earlier audit of our own work found the sample inflated ~6× by counting overlapping intraday rows. Every statistic here counts each session once.
- The benchmark is not a coin. It is the trivial IB rule: the day closes on the side where the middle of the first hour sat relative to the open. Anyone can compute it in ten seconds with free data. Over our sample it is right 71% of the time (95% CI 66–77). That is the floor. A result only counts as ours if it beats that, not 50%.
- Predicted direction stated in advance. Where a hypothesis implies an ordering, we write the ordering down first. A result that matches a prediction made blind is worth far more than the same number found afterwards.
- One run. If a rule fails, it is not tuned and retried. It is reported as failed.
Everything below is measured on NQ and ES futures, roughly 260 sessions each, with the path walked bar by bar — not on closing prices.
What held up
The size of the gap predicts whether it closes
The auction logic is direct: a small gap is session noise and the market fills it back in; a large gap is genuine repricing — something happened, and reversing it would mean undoing that revaluation. So we predicted the ordering small > medium > large, and then measured it.
| Gap size (vs daily ATR) | NQ | ES |
|---|---|---|
| Small (< 0.25 × ATR) | 71% (95% CI 63–79) · n=126 | 79% (95% CI 72–87) · n=126 |
| Medium (0.25–0.5) | 55% · n=77 | 59% · n=68 |
| Large (≥ 0.5 × ATR) | 43% (95% CI 26–60) · n=35 | 41% (95% CI 26–59) · n=34 |
| Small − Large | +28.6pp (95% CI +10.3 to +46.7) | +38.2pp (95% CI +19.8 to +56.1) |
The predicted ordering held exactly, in both markets, and both confidence intervals exclude zero. It survives correction for the number of comparisons we ran.
How the first 30 minutes behave predicts the rest
When a gap closes, it closes fast — median one to two 15-minute bars. So we asked whether the first half hour tells you if the premise is still alive.
The naive version of this question is circular, and we discarded it before running: comparing the whole session's adverse move between days that filled and days that did not is just the same observation twice — a day that did not fill moved away by definition. Instead we measured only the first 30 minutes, and evaluated only days that had not already filled inside that window. The measurement window and the outcome window do not overlap.
| Adverse move in the first 30 min | NQ | ES |
|---|---|---|
| Small (< 0.10 × ATR) | 64% · n=28 | 63% · n=35 |
| Medium (0.10–0.25) | 54% · n=57 | 59% · n=56 |
| Large (≥ 0.25 × ATR) | 24% · n=70 | 37% · n=59 |
| Small − Large | +40.0pp (95% CI +19.3 to +60.0) | +25.6pp (95% CI +5.1 to +45.5) |
Again the predicted ordering held in both markets, and both intervals exclude zero. This one is a live read: at 10:00 it tells you whether to keep waiting or let the day go.
A number the industry repeats that is simply false
You will find it in courses, blogs and forums: "price returns to the previous day's POC about 80% of the time."
We measured it on our own data: 57% on NQ, 55% on ES. Two markets, consistent, and nowhere near 80%. We do not use that number, and neither should anyone quoting it to you.
What we failed to prove
This is the part nobody publishes.
| What we tested | Result |
|---|---|
| Predicting the day's direction | Never beat the trivial IB rule, in any segment |
| Overnight inventory imbalance gating the gap | No discrimination — measured three times; on the third it came out reversed |
| Where the open sits vs prior value, conditioning the gap | +4.0pp (95% CI −12.3 to +20.4). Predicted ordering did not hold |
| An early read from location alone | −3.6pp (95% CI −15.5 to +7.9), and it degraded out of sample (22% → 12%) |
| An early rotational read | +15.4pp (95% CI −0.4 to +31.0) — failed by 0.4pp against a criterion signed in advance |
| Finer chart granularity | Coverage did not move at all |
| Filtering trades by reward-to-risk | Measured before building it: the win rate falls almost exactly in proportion as the payoff rises. The market prices the asymmetry fairly |
The inventory result changed the product: the indicator used to require an imbalance before issuing the gap plan. That filter was systematically discarding the better days. We removed it.
The early rotational read is the one that stings. It missed by four tenths of a percentage point on a criterion we had signed before exporting the data. We could have called it "essentially positive." We did not, and we are not going to.
What we do not claim
- We do not claim the product is profitable. No measure of expectancy we ran cleared zero.
- We do not publish a win rate for trades. We do not have one we can defend.
- The frequencies above are historical rates, not probabilities for today. They describe how that kind of gap behaved in our sample. Tomorrow is not in the sample.
- Everything above is one instrument family — US equity index futures — over roughly one year.
What the product does is narrower and, we think, more useful: it builds the day's auction map the same way every session, marks the two moments where a decision exists, and shows you the destination, the invalidation and the reward-to-risk so you can decide with the arithmetic done.
Check it yourself
The indicator writes its own telemetry. Turn on the backtest log, let it rebuild your history, and you get three CSV files with the raw path — prices, bars, timestamps — and no verdict columns. Nothing that says "correct" or "hit": the judgement is computed in the analysis, so changing the question never requires re-exporting the data.
That means the numbers on this page are reproducible from your own chart, on your own instrument. If your data disagrees with ours, we want to know.