We stopped shipping trade plans because we could not defend their quality. That leaves an obvious question: if the bot is not going to tell you what to do, what does it actually know about the events it reports?
This is the first answer. It is descriptive, not promotional. Nothing below is a win rate, a success rate or an accuracy figure, and none of it says a setup was tradeable.
What is being measured
Every event the bot delivers opens a virtual position in an internal book at the live price, with an invalidation level derived from the event's own structure. A cron resolves it against 1-minute candles. From that book we can read, per event type:
- Median move with the event - the median of the largest excursion in the direction the event pointed, before the event resolved.
- Upper quartile - the same measure at the 75th percentile: the better quarter of cases.
- Median move against - the median of the largest excursion the other way over the same window.
- Came back through the level - the share of events where the adverse excursion reached the event's own structural invalidation. In plain terms: price traded back through the level the event was built on.
- Median time to resolve - how long the event took to finish, one way or the other.
The numbers
Delivered events, 90 days to 22 August 2026. Only event types with at least 30 resolved events are shown - below that a median is noise, and publishing it would be the same sin we are trying to stop committing.
| Event type | Events | Median move with | Upper quartile | Median move against | Came back through the level | Median time to resolve |
|---|---|---|---|---|---|---|
| Fakeout (failed break) | 108 | 1.26% | 2.33% | 1.07% | 36% | 51 min |
| Market-structure break (5m) | 72 | 1.85% | 3.56% | 0.94% | 24% | 1.7 h |
| Level approach * | 110 | 7.91% | 15.55% | 4.13% | 21% | 24.3 h |
* Approach events are measured on a day-scale horizon; the other two resolve within minutes to hours. Read across a row, never down a column - a 7.91% median over a day is not "better" than a 1.85% median over 100 minutes, it is a different question.
What we think this says - and what it does not
The honest reading of the two minute-scale rows is that the move with the event and the move against it are close to the same size. A fakeout's median 1.26% in favour sits next to a median 1.07% against, and roughly a third of them trade back through the level that defined them. That is the shape of an event worth looking at, not the shape of a setup that carries itself.
That is also exactly why these alerts no longer carry an entry and a stop. A plan needs the favourable excursion to dominate the adverse one by enough to pay for costs and for being wrong; these distributions do not show that, and we are not going to imply it with a chart of cherry-picked wins.
What the numbers do give you is calibration. If you take a fakeout, you now know the typical room the market gives it before it decides, how often it walks back through the level, and roughly how long you will be sitting in it. That is a real input into your own decision, made with your own levels.
The caveats, stated plainly
- Everything here is conditional on the events that passed our gate. These are not all breaks or all fakeouts in the market; they are the ones our filters let through. Change the gate and the table changes.
- Ninety days at this sample size is a description, not a finding. Three months of a few hundred events cannot separate a real structural property from a regime that happened to be running.
- The excursion is measured to resolution, so an event type with a longer horizon mechanically shows larger excursions in both directions. That is the whole reason for the asterisk.
- Median is not typical for you. Half of the events did worse than the middle number, and the distribution has a long tail on both sides.
We will refresh this page as the sample grows. If a number moves against the story above, the number wins.