Research · Intraday
The short answer
Shorter bars make the close an artifact of the clock, costs a larger share of each move, and the still-forming bar a trap that has already invalidated one study here.
Candlestick formations were described on daily bars, where a session has an open, a close and a natural boundary. Applied intraday, three things change at once, and all three work against the reader.
The specific trap intraday: acting on the bar in progress. A formation is defined by its close, and mid-bar the close does not exist. A hammer at 14:30 on an hourly bar is not a hammer — it is a shape that may or may not be one at 15:00.
This is not a beginner's error. The measured run this site was preparing to publish included the forming bar in its candlestick rows, and it was withdrawn for exactly that reason rather than published with a caveat.
Where this goes wrong
If you act on a formation before its bar closes, you are trading a different rule from the one that was tested — and you have no idea what that rule's properties are, because nobody has measured it.
The chart formations here confirm later than tradition suggests: a rule fires on the bar where the last required swing is confirmed, not where a trader would enter. A swing at bar *k* is only known once enough further bars have printed without exceeding it.
That lag is not a limitation of the implementation, it is the price of refusing look-ahead. Any detector that identifies swings faster is using information the market had not yet produced — which is why its backtest looks better and its live results do not.
If you trade intraday, the parts of this site that carry over cleanly are the risk arithmetic, not the formations: the position size formula, expectancy, and the fact that costs scale with your trade count while edge does not.
On the formations themselves, wait for the bar to close, and treat every intraday result as needing a higher bar of evidence than a daily one — because the noise is larger and the costs are proportionally worse.
An H1 bar at 03:00 London time and one at 14:30 are not comparable observations. Liquidity, spread and participation differ by an order of magnitude, and a formation detected in a thin session is mostly detecting thinness.
| Session | What a formation is likely to be measuring |
|---|---|
| Asian | Range behaviour and low participation |
| London open | Genuine repricing, widest ranges |
| London/NY overlap | The deepest liquidity of the day |
| Late NY | Position squaring and thinning books |
A study pooling all sessions reports an average of several different markets. Splitting by session is more honest and immediately multiplies the number of tests, which then has to be corrected for.
Intraday data is abundant, and the abundance is misleading. Ten years of M5 bars is around 750,000 observations, which sounds decisive and is not: they are heavily autocorrelated, they cover the same few thousand daily market states, and the effective sample is far smaller than the count.
Meanwhile the costs scale with the trade count while the edge does not. A rule generating twenty signals a day pays the spread twenty times for the same underlying move a daily rule captures once.
Where this goes wrong
The honest test for an intraday rule is a higher bar than for a daily one: more data, more corrections, and a cost model that reflects the frequency rather than a nominal spread.
None of this makes intraday trading unworkable. It changes what the evidence has to clear before a rule is worth acting on.