Why the next 15 minutes, and not the next 5 or the next 120

Every chart-reading question has a horizon buried in it, and it is usually treated as a setting. Ours is fifteen minutes: you see a real intraday session frozen at a decision point, and you call where price is fifteen minutes later.

Change that number and you have not adjusted the question. You have asked a different one. Across 140,319 decision points in 22 US equities and ETFs, the five-minute answer and the fifteen-minute answer point in opposite directions 31% of the time.

So the horizon had to be chosen, not defaulted into. Here is what we measured.

What we were choosing between#

A question is worth asking only if two things are true at once. There has to be something to call — a move large enough that it is not just the price wobbling. And it has to be implied by what is on the screen, rather than by news that arrives afterwards.

Those two pull in opposite directions as the horizon grows, which is the whole problem.

To compare horizons on equal terms we scanned every decision point in the cache with no magnitude filter at all — the published bank is deliberately filtered, so measuring horizons on it would answer a question we had already assumed. What follows is the raw sweep.

Why not five minutes#

At a five-minute horizon, 33% of decision points move less than a quarter of the symbol's own ATR. A third of the time, nothing has happened. The reader is not being asked to read a chart; they are being asked to guess which way a stationary price will twitch.

Put the other way round: only 11% of five-minute outcomes land in the band where a move is big enough to have been implied by something and small enough not to be a shock.

There is a second reason, and it comes from measurement rather than from us. Below roughly five minutes, price data stops being a clean record of what the market did and starts carrying the mechanics of trading itself — the bid-ask bounce, the discreteness of the tick. The reference point here is Liu, Patton and Sheppard's 2015 comparison of over 400 volatility estimators across 31 assets and 11 years, which found "little evidence that it is outperformed by any other measures" when five-minute realized variance was used as the benchmark.

That paper is about measuring volatility, not about forecasting direction, and it would be sloppy to stretch it into a claim about horizons. What it does support is the bar: five minutes is where the profession reads intraday prices, because below it you are increasingly reading the plumbing.

Why not two hours#

The long end fails differently, and the reason is boring and structural: a horizon has to fit inside the trading day.

A question needs its outcome bars to exist. The median regular session is 78 five-minute bars, and after reserving the trailing window the ATR is computed from, the horizon eats the rest:

A two-hour horizon costs more than a third of every session — but the number that decided it is the right-hand column. The last question you could ask would be written at 14:00. The closing hour would not exist in the bank at all, and neither would most of the afternoon. Those are not interchangeable minutes: the hour into the close is one of the most distinctive stretches of the session, and a practice bank that has never shown it to you has a hole in it exactly where a lot of real decisions happen.

The part we had wrong#

We expected the naive rule — fade the last candle — to work best at the shortest horizon, where mean reversion should be strongest. It does not.

Unfiltered, across all 140,319 decision points, fading the last candle is correct 50.1% of the time at five minutes, 50.0% at fifteen, 49.9% at thirty, 49.8% at sixty and 50.0% at two hours. Five horizons, and every one of them is a coin.

That is worth sitting with, because the same rule scores 60.7% inside the candidate pool the bank is drawn from.

⚠️ This paragraph used to blame the magnitude band on its own, and that is only half of it. The band applied to every eligible bar leaves the rule at 49.3%. It reaches 63% only once the generator's spacing rule is switched on as well — the one that skips forward after each accepted question so outcome windows stay disjoint, and which as a side effect keeps the first bar of every qualifying run. The other article now carries the stage-by-stage decomposition.

What survives of the original point is the part that matters here: the reversal tendency is not in the tape at any horizon. It is a property of how the question set is assembled, which is a different and less comfortable claim than the one this section used to make.

So why fifteen#

Thirty minutes is defensible, and by the readable-band measure it is better: 43% against 31%. We chose fifteen anyway, for two reasons we can state plainly.

The first is coverage, above — thirty costs another five percent of decision points and pulls the last question of the day back to 15:30.

The second is that for the setups the bank actually keeps, waiting the extra fifteen minutes barely changes the answer. Across the 52,768 questions in the current bank, the thirty-minute direction agrees with the fifteen-minute one 96% of the time. You would be asking very nearly the same question and paying session coverage for the privilege.

⚠️ That 96% is conditional and must not be read as a fact about the market. The bank's questions were selected on their fifteen-minute outcome, so they are precisely the setups where fifteen minutes produced a decisive move — and a decisive move tends to still be pointing the same way a quarter of an hour later. Measured on unfiltered decision points, thirty minutes agrees with fifteen 75% of the time, not 96%. The honest version of the argument is therefore narrower: for the kind of setup this bank asks about, the extra fifteen minutes is mostly redundant. Whether that holds for setups we do not select is not something this measurement can say.

So: fifteen minutes is the shortest horizon at which the outcome is usually not noise, and the longest one that costs almost nothing in coverage. It is a corner, not an optimum.

What is fixed, and when#

Two design decisions follow, and both are in the generator rather than in a document.

The horizon is written into the question when it is generated, never selected afterwards. A horizon chosen after the outcome is known is the purest form of look-ahead bias there is: with five horizons stored, picking the flattering one per question would let us manufacture any accuracy we wanted.

We do store all three of 5, 15 and 30 minutes for every question. That is deliberate and it is not a contradiction — the primary answer is always the fifteen-minute one, and the others exist so that a future mode asking "which way, and how long would you hold?" can be scored without regenerating the bank. The extra checkpoints are never an input to the answer you are graded on.

What this does not buy#

Fifteen minutes is a defensible choice inside a range of defensible choices. It is not a discovery about markets, and nothing here says the next fifteen minutes are knowable — the sweep above says the opposite about naive rules, at every horizon we tested.

What the choice does buy is a question that is worth answering: long enough that something usually happened, short enough that it is still the chart in front of you that implied it, and fixed early enough that nobody — including us — can go back and choose a horizon that flatters the answer. The rest of the parameters are on our methodology page.