How long a sample to use

Updated 2026-10-23 Español·Deutsch·Русский·Italiano·Français·Türkçe·العربية

Sample length is the number of candles used to describe the situation you want to compare with the past. Choose it to include a clearly defined event, not to obtain the most attractive subsequent returns. A longer sample is not automatically a better one.

Short window

Short windowCandle count: 3 [110, 102, 106, 100, 104, 99, 109, 105, 116]99116Time →
  • Candle count: 3

Long window

Long windowCandle count: 9. Window boundaries change which events enter the comparison. [110, 102, 106, 100, 104, 99, 109, 105, 116]99116Time →
  • Candle count: 9
  • Window boundaries change which events enter the comparison.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

What a shorter or longer window changes

A short window may isolate a recent move and leave out the context that explains it. Simple shapes can recur in very different situations. This does not make every short sample useless; it means the shape alone may not distinguish those situations.

A longer window includes more context, but it can also combine separate events or different market regimes. Requiring agreement over a longer stretch may reduce the match count under a particular comparison rule; it does not guarantee that the remaining matches are closer or more useful. The effect depends on the data, normalisation and selection criterion.

Choose the event before looking at outcomes

  1. State the question in observable terms, for example: a range, a break beyond its boundary and a return to that boundary. This describes an event, not a promise about the next move.
  2. Choose the timeframe and check the trading calendar and missing candles. Mark the last completed candle as the end of the sample.
  3. Set the starting point by a repeatable rule using only information available at that end point. Include the event's relevant beginning; do not choose an anchor because the later price path looks convincing.
  4. Check that the window contains the event you named without adding unrelated history. If you consider neighbouring lengths, judge event coverage first, not which one produces the best future outcome summary.
  5. Record the boundaries and selection rule. Then inspect the historical matches and evaluate subsequent outcomes on a separate chronological period, without retuning the rule to those outcomes.

Different analysts can ask different questions of the same chart. Making the rule explicit is more useful than claiming a single universally correct candle count.

Duration

DurationSame candle count does not mean the same duration.. A longer window does not prove causality. ['50 × 1h = 50h', '50 × 4h = 200h', '50 × 1d = 50d']50 × 1h = 50h50 × 4h = 200h50 × 1d = 50d
  • Same candle count does not mean the same duration.
  • A longer window does not prove causality.
Teaching diagrams with hypothetical numbers. They explain concepts, not the next price move.

Fifty daily candles are not always the same time span

On a continuously trading crypto market with no missing data, 50 daily candles cover 50 calendar days. On an exchange that closes at weekends and holidays, 50 daily candles represent 50 trading sessions, spread over a longer calendar period. Intraday candles also follow the instrument's trading hours.

A candle count therefore needs a timeframe and a calendar. Changing either can change the event described even when the count stays the same.

Keep the sample and outcome period distinct

The sample describes the shape up to its end; the observation period measures what happened afterwards. There is no general rule that one must be inversely proportional to the other. Define the outcome period before evaluating results and report it explicitly. The article on analogue statistics shows why an endpoint and the path to it can tell different stories.

When no suitable matches appear

An empty result means none met the stated rule in the available data. It does not prove that the situation has never occurred. Check event boundaries, missing data and context, including the limits of an EMA200 regime label. If you revise the rule, record that revision rather than adjusting repeatedly until an attractive result appears.

A defensible window makes a historical comparison clearer. It does not turn similarity into a forecast; see research on technical analysis.

Related questions

In depth

See it on a live chart

The same markup on a prepared example: the chart, the objects it found and what followed.

Open the example