What actually changes with timeframe?
Four things change at once, which is why the question cannot be settled by preference: cost per unit of movement, available sample size, the noise-to-signal ratio, and how much of the day you have to be present for.
Cost. Roughly fixed per trade, so it falls as a fraction of the move as the horizon lengthens. This alone pushes marginal strategies toward higher timeframes.
Sample size. A strategy on M15 accumulates decisions roughly sixteen times faster than the same rule on H4. One year of M15 can be more informative than four years of H4, which matters enormously for validation.
Noise. Lower timeframes contain more movement that is not attributable to anything a rule can catch. Whether that hurts depends on the strategy — a mean-reversion rule may feed on it while a structural rule drowns in it.
Your attendance. A rule that fires at 3am and requires manual execution is a rule you will trade inconsistently, which is a real cost even though it appears nowhere in the backtest.
How do you compare timeframes fairly?
Hold the IDEA constant and express every parameter in units that scale — ATR rather than pips, bars rather than hours — or you are comparing five different strategies rather than one strategy at five resolutions.
A 20-pip stop is a tight stop on H4 and an enormous one on M5. Testing the same pip value across timeframes tests a different risk profile at each, and whichever timeframe happens to suit that fixed distance will win for reasons unrelated to the idea.
Expressing the stop as 1.5 ATR of the operating timeframe keeps the risk proportional to that timeframe's typical movement, which is what makes the comparison about the signal rather than about the units.
Then hold the cost model constant in absolute terms, because it genuinely is. The same 1.2-pip spread applies on every timeframe, and its shrinking relative weight as the horizon lengthens is a real effect the comparison should show rather than normalise away.
Doesn't testing five timeframes count as a search?
Yes, and it has to be counted as one — five timeframes is five trials, and the best of five looks better than the truth by an amount that grows with how similar they are.
The defence is the same as for any parameter: fix the choice on one part of the history and grade it once on a part you have not examined. If the winner on the development period is still the winner out of sample, that is meaningful; if the ranking reshuffles, the original ranking was noise.
Watch for a specific artefact while doing this. Lower timeframes generate more trades, so their results are less noisy, which can make them look more consistent without being better. Compare expectancy per trade rather than total return, and note the trade counts side by side.
And treat a result that appears on exactly one timeframe with suspicion. A real structural effect usually degrades gradually across neighbouring resolutions rather than existing at H1 and vanishing at M30 and H4.
Is there a reasonable default?
For most retail strategies the range from H1 to daily is the practical sweet spot, because costs are a small fraction of the moves while the trade count still accumulates fast enough to validate within a reasonable period.
Below H1 the cost model becomes the dominant term and the data requirements escalate — tick data, per-session spreads, latency assumptions — which is a real project rather than a setting.
Above daily, the sample-size arithmetic turns hostile. A daily-bar strategy taking two trades a month accumulates 24 a year, and the several hundred needed for confidence is a decade. That is not a reason to avoid the horizon; it is a reason to test across many instruments simultaneously to accumulate the sample.
The genuine answer remains that it depends on the idea, and the useful part is that finding out costs one afternoon and a properly scaled parameter set.