The last candle reverses 60% of the time — in our questions, not in the market

When we built the SwipeTA question bank, we ran every dumb rule we could think of against it before we let a single question near a player. One of them did unreasonably well.

The rule is: look at the last 5-minute candle before the decision point; call the opposite way. Green candle, swipe down. Red candle, swipe up. No context, no levels, no volume — one candle.

Across the 1,200 questions that run selected, that rule is right 60.67% of the time.

We scored its mirror image at the same time. Following the last candle instead of fading it — green candle, swipe up — comes out at 51.58%, which is a coin with an opinion.

Side by side, that looks like one obvious rule beating the other by ten points. It is more complicated than that, and the next section is a correction to what this article originally said about why.

Correction: where the 60.67% actually comes from#

This section used to say the opposite of what follows, and it was wrong. It claimed the rate was "not an artifact of our selection — it is there in the raw candidate pool before any filtering". A later measurement contradicted that, nobody reconciled the two, and the sentence stood for two weeks. Here is the accurate version.

We rebuilt the generator's pipeline stage by stage on the same seven symbols and period, scoring the same rule after each step. The reconstruction lands within 0.3% of the published pool size and within a point of the published rate, so it is close enough to say which step is responsible.

Unfiltered, the rule is a coin. Across 366,606 decision points it is right 49.78% of the time. Whatever the 60.67% is, it is not a standing property of five-minute candles.

The magnitude band, on its own, does nothing. Applied to every eligible bar it moves the rule by half a point. That was the obvious suspect and it is innocent by itself.

The thirteen points appear at the next line, and that line is not a filter. After the generator accepts a decision point it skips forward by one horizon before it will accept another. Its purpose is stated plainly in the code: outcome windows must not overlap, or two questions are decided by the same bars. That reasoning is correct and we would make the same choice again.

But qualifying bars do not arrive alone. They arrive in runs — neighbouring bars share most of an outcome window, so if one qualifies its neighbours usually do too. Skipping forward after each acceptance therefore keeps the first bar of every run and discards the rest. And selecting the onset of a move rather than a random point inside it is worth thirteen percentage points to a contrarian rule.

So the honest reading of 60.67% is that it describes the question set, not the tape. A step introduced to remove a statistical dependence quietly introduced a directional one, and nobody measured the second thing because the first thing was correct.

We have not established why the onset of a qualifying run should favour a fade. That is a real open question, and having now been wrong about this twice, we would rather leave it open than guess a third time.

Why it is still a bad rule#

Two reasons, and the second is the one that costs money.

It fails four times in ten. A rule at 60% with no notion of when it does not apply is a rule you cannot size. You have no way of telling the setups where it holds from the setups where you are standing in front of a trend that is only just getting started — and those are exactly the ones that run furthest against you.

Knowing it makes you worse at the thing you were trying to learn. This is the part we did not expect. If you internalise "fade the last candle" you stop reading the chart, because a rule that works most of the time removes the reason to look at anything else. You are then very well prepared for the 60% of situations that were going to resolve themselves anyway, and completely unprepared for the 40% that were not.

That is a bad trade for a training tool. The whole point of practising is to build the judgement that tells those two cases apart.

What we do about it#

We could have removed every question the naive rule answers correctly. That would have been worse: a bank where fading the last candle always fails is just as solvable, in the other direction, and players would learn a rule that is wrong in the real world.

What we do instead:

• The competitive pool keeps the same proportion as the candidate pool it is drawn from. Publication of the question bank aborts outright if that share drifts by more than 5 percentage points. Note what that guard does and does not do: it stops us from quietly editing the rule's success rate between the pool and the published set. It does not make the pool representative of the market, and the correction above is why that distinction matters.
• The practice pool deliberately keeps a share of questions where the rule fails. A beginner pool of only readable charts would teach a rule that then breaks two times in five as soon as someone plays a ranked round. That is not building confidence, it is setting up a betrayal.

What to do with this if you are practising#

Do not use it as a rule. Use it as a null hypothesis.

When you look at a frozen chart and your instinct says "this pushes back", ask whether you are reading the chart or whether you are just doing the thing that is right 60% of the time. The setups where those two answers differ are the only ones that are teaching you anything.

The mechanics of how these questions are built, filtered and judged are written out in full on the methodology page.