How many trades a hit rate needs before it means anything

A hit rate is an estimate, and every estimate has a width. We computed the width for all seventeen channels we can score, and for most of them the number on the page settles almost nothing.

What is a hit rate an estimate of?

Of something you cannot observe: the rate at which a channel's calls would work over the long run. What you can observe is a sample, and the percentage printed on a sales page is that sample's result, not the underlying rate.

The distance between those two is what this article measures. It is not a philosophical point. It is a number, it can be computed for any channel from two figures, and for most channels in our index it turns out to be enormous.

The tool is a confidence interval: the band of underlying rates that could plausibly have produced the sample you saw. Wide band, weak evidence. That a percentage needs one at all is covered in Win rate: why 90% accuracy can still lose you money; this article is the arithmetic, applied to real channels.

How wide is the band for a real channel?

Wider than the differences people use it to decide between. Here are the seventeen accounts in our index that publish calls we can score, as of August 2026, with the 95% interval around each one's hit rate.

One channel has a band under ten points wide. Nine of the seventeen have bands wider than 25 points, and the last row spans from a coin flip's worse to nearly certain.

Three of the rows are the same publisher. The channels at 266 and 267 scored outcomes carry hit rates within a fifth of a point of each other, because they are one feed under three brands. Their intervals are identical for that reason, and the index of seventeen is therefore at most fifteen independent publishers. That count is in Only 14 of the 47 signal channels we index still post.

Which of these differences are real?

Fewer than the table appears to show, because overlapping bands cannot be ranked.

Take the row at 85.7% on 28 scored trades, band 68.5 to 94.3. And the row at 88.1% on 109 scored, band 80.7 to 92.9. The second channel advertises a higher number, and every rate between 80.7% and 92.9% is consistent with both samples. The two are not distinguishable by their percentages.

Now take the row at 85.0% on 20 trades, band 64.0 to 94.8, against the row at 74.7% on 91 trades, band 64.9 to 82.5. Ten points separate the advertised numbers. The bands overlap across almost the whole of the smaller channel's range.

The general rule that falls out: when a channel with few trades outranks a channel with many, the ranking is usually an artefact of the sample rather than a difference in the underlying rate.

How many trades would you actually need?

Several hundred, for the kind of precision people assume a percentage carries.

Read the bottom row against the previous table. At a rate near 60%, plus or minus five points needs about 366 scored trades, and no account in our index has that many. Exactly one gets under ten points anyway, on 337 scored outcomes drawn from 1,534 recorded ones, and it manages it only because its rate sits far from 50%: at 80% the same precision needs 245 trades rather than 366.

Two consequences follow. A new channel cannot have a meaningful hit rate, however good it is, because it has not had time to generate one. And a channel advertising a precise-sounding number after a few months is advertising something the arithmetic cannot support yet.

Why do the best-looking percentages sit on the fewest trades?

Because the same conditions produce both, and because of what happens when many channels compete.

The mechanical half: a wide stop that price rarely touches leaves most calls unresolved, so few outcomes get scored, and those that do resolve skew toward the target. The channel in our index advertising 91.3% has 23 scored outcomes behind it. The one at 66.7% has six.

The selection half is worse and it does not require anyone to cheat. Take a hundred channels with no skill at all, each publishing twenty calls. Some will finish above 80% by chance, and those are the ones with a screenshot to post and a subscription to sell. You are not looking at a random sample of channels; you are looking at the ones whose samples came out well.

That is why a high percentage on a small sample is evidence of a small sample more than of anything else.

What does the denominator hide?

Usually more than the percentage reveals, because scored outcomes are a minority of published calls.

Across our corpus, 3,167 recorded outcomes produce only 1,583 scored verdicts as of August 2026. The rest are calls whose entry was never reached, calls overtaken before the entry was available, and calls that never resolved because no stop was published. So a channel's sample size counts only the calls that ended in a way anyone can grade, which is a much smaller number than the calls it made.

The gap between those two is itself a channel property worth checking. The account with 337 scored outcomes recorded 1,534, so roughly four in five of its calls contribute nothing to its own percentage. Where those go is covered in Half the trades never happen.

What does a regulator make of small samples?

The same thing the arithmetic does, in the markets where somebody can be fined for it.

Australia's ASIC treats past performance drawn from "an inappropriately short time period" as potentially misleading, and for an operation running under a year says such a period "would usually be inappropriate and may be misleading" (ASIC Regulatory Guide 234, June 2026). The US National Futures Association, examining whether a disclaimer repairs a thin record, concluded that "the use of the mandated disclaimer has not prevented recurring abuses" (NFA Interpretive Notice 9025).

Neither body regulates a Telegram channel, and a channel owes you no disclosure at all. The point of citing them is that the problem is well enough understood to have rules written about it elsewhere.

What do we do about this in our own ratings?

We publish a rating from ten scored verdicts, and by the standard in this article that threshold is far too low to prove anything.

We keep it because ten is the point where a percentage stops being pure noise. It is nowhere near the point where the number becomes reliable. A channel sitting at fourteen scored trades has a band 46 points wide, and our page will still show it a number. That is a deliberate trade between coverage and precision, and this article is the disclosure that goes with the trade.

Read any of our percentages alongside the count printed next to it. What our replay does and does not establish is on the Methodology page, and the wider limits of replay as evidence are in What a backtest can and cannot prove.

How do you check this yourself?

Two numbers and a rule of thumb, and you do not need our table.

1. Find the count, not the percentage. Ask how many trades the number is computed from, and how many of those were scored rather than published. A channel that cannot answer has already answered.
2. Apply the rough band. At about 20 scored trades the band is roughly 40 points wide; at 100 it is roughly 19; at 300 it is roughly 11. Those are approximations from the table above, and they are close enough to decide with.
3. Compare bands, not numbers. If two channels' bands overlap, their percentages do not rank them, whatever the sales pages say.
4. Check the count is still growing. A record that stopped updating describes a channel that stopped trading.

Building the count yourself, from the feed rather than from the claim, takes about ten minutes and the method is in How to check a signal channel's track record in 10 minutes.

What this does not prove

A wide band is not evidence that a channel is bad. It is evidence that the percentage cannot tell you either way, which is a different and more useful conclusion. The channel at 91.3% on 23 trades may well be excellent; nothing here says otherwise, and nothing here says it is not.

The intervals above are Wilson intervals at 95%, computed on our scored outcomes. They assume each call is an independent trial, which is generous: signal calls cluster by market condition and direction, and correlated trials carry less information than independent ones. Real bands are therefore wider than the table shows.

They also inherit every assumption of the replay that produced the outcomes, including that a candle touching the published entry counts as a fill. Our figures describe our index as of August 2026: 17 publishing accounts of 50 indexed, at most 15 of them independent, 3,167 outcomes and 1,583 scored verdicts. Not a random sample of the market, and every figure moves as the replay catches up.

Nothing here is financial advice, and nothing here alleges misconduct by any channel.

The practical read

Three questions replace the percentage, and none of them is harder to ask.

How many scored trades is this number computed from? How wide is the band that count implies? And does that band overlap the band of the channel I am comparing it to?

Current ratings and the counts behind them are at Signal Providers.

Sources

• NIST/SEMATECH e-Handbook of Statistical Methods, section 7.2.4.1: confidence intervals for a proportion
• ASIC Regulatory Guide 234, June 2026
• NFA Interpretive Notice 9025: use of hypothetical performance results
• CFTC: Commodity Trading Systems Sold on the Internet