Nearly every accuracy figure a signal channel advertises is 90% or better, and almost none of them say what was counted. We measured the claims in 25,932 archived messages, then measured the signals.
How accurate are crypto signals, claimed and measured?
Claimed: almost always 90% or higher, and almost always the same figure. Measured under one fixed rule across the seventeen channels with enough scored trades to rate: 55% to 91%, with a median of 68%. The two ranges are not comparable with each other, and why that is so is the subject of this article. We searched 25,932 messages archived from 42 Telegram channels for any claim of the form "X% accuracy", "X% win rate", "hit rate" or "success rate", and found 1,445 messages carrying one, spread across 14 of the channels. Of those claims, 1,423 sit at 90% or above.
Twenty-two claims fall below ninety - twenty of them between 68% and 88%, the lowest anywhere in the archive being 23% - and the other 1,423 cluster at the top. One value dominates outright: 95%, published in 1,299 separate messages by three channels. Sixty-five messages claim exactly 100%.
Two channels account for 1,298 of those 95% claims, so this is a measurement of what gets repeated more than a poll of the market. The caveat matters and it does not rescue the picture, because the shape survives when you count channels instead of messages: twelve of the fourteen channels that state any accuracy figure state one of 90% or better at least once. A number below ninety is close to unpublishable in this business.
Why is an advertised accuracy figure not a number?
Because "accuracy" names a question, and the answer depends on four decisions that nobody has to disclose. A channel can make all four honestly, publish the arithmetic, and still produce a figure that cannot be set beside any other channel's.
| Decision | Two defensible answers | What it does to the figure |
|---|---|---|
| How far the price must move to count as a win | The first rung of a six-target ladder, or the last rung reached | A ladder whose first target sits 0.5% away is hit constantly; one aiming for 5% is not |
| What happens to calls the market never let you enter | Dropped from the count, or carried as failures | Roughly half of everything in our own corpus falls in this bucket |
| What ends a trade with no published stop | A time limit, a fixed adverse move, or the author's judgement | Nothing closes the trade until somebody decides it is closed |
| Which stretch of history is counted | Since inception, or the last thirty days | Any operation has a good month somewhere in it |
None of those requires a false statement. Change the first decision alone and the same feed of signals produces two very different percentages, which our own replay demonstrates on 3,988 parsed signals. The reason a percentage stays incomplete even after all four are pinned down is separate arithmetic, worked through in Win rate: why 90% accuracy can still lose you money.
What happens to the same signals under two different rules?
They come out looking like two different businesses. We replay every parsed signal twice, under rule sets that disagree only about what closes the trade, and the two answers point in opposite directions.
As of 17 August 2026 we hold 3,988 parsed signals from the 22 accounts in our 47-channel index that publish anything a parser can read. Under our published one-take rule, which closes the whole position at a 2% move, 1,935 signals produced a scored verdict and 1,305 of them reached that take before the stop. That is a hit rate of 67.44%.
Under the second replay, which follows each channel's own ladder of targets and its own stop exactly as published, 1,892 outcomes carry a profit or loss figure. Their average is -0.54%, and 1,284 of the 1,892 finish at or below zero.
Nothing changed between those two sentences except the definition of an ending. One definition says two thirds of these calls worked; the other says two thirds ended flat or down. The channel did not get better or worse in between. A percentage is therefore a fact about the rule that produced it at least as much as a fact about the channel, and a channel that publishes one without naming the rule has published a mood.
Why do we score every channel against one fixed target?
To make the number comparable, which is the only thing a fixed target buys and the reason it is worth the loss of realism.
Our Methodology page states the rule set in full. The short version: enter at the published entry with the whole position and no leverage, close everything at a 2% move, and exit at the channel's own published stop. Each trade gets 30 days, resolved against real 1-minute candles using highs and lows instead of closing prices. Fees are charged. Signals whose entry the market never reached are excluded from the percentage entirely, because "we could not tell" and "it lost" are different statements.
The fixed target exists because target ladders are the easiest thing in this market to tune. As the methodology page puts it, numbers built on the channels' own targets are "trivially gameable by placing the first target closer". A channel that wants a better percentage under its own convention needs no new skill and no new signal. It moves TP1.
So the 2% take is a low bar on purpose. It asks every channel the same question and it answers only that question: how often did the market move 2% the way you said, before it hit your stop. Clearing that bar and losing money on the same trade is entirely possible, and it happens often enough to have its own count in You took the first target and still lost money.
Where does our own rule set flatter a channel?
At the fallback stop, and it flatters the channels that published the least.
Of the 3,988 signals we hold, 2,022 carry no stop at all. A replay still has to end those trades somewhere, so we apply a stop 10% from entry and disclose on each channel's page how many of its verdicts rest on that fallback. A stop that far away is rarely touched, which lifts the hit rate of every channel that never published one. The size of that effect is measured in What a backtest can and cannot prove about a channel, and why the stop goes missing so often is in Why so many signals come without a stop loss.
So our own figures carry the same warning this article aims at everybody else's. Comparing our number for a stopless channel with our number for a channel that publishes stops sets a lenient measurement beside a strict one.
What accuracy did the channels actually produce?
Between 55% and 91% under the one-take rule, with the middle of the range nowhere near the advertising.
Of the 47 accounts in the index as of 17 August 2026, 17 have enough scored verdicts for us to publish a hit rate at all. Across those seventeen the figures run from 55.07% to 91.3%, and the median sits at 68.18%. Five further accounts have between five and nine scored verdicts, below our minimum of ten, and are shown as a dash instead of a number.
The full outcome distribution behind those figures, including the largest category of all, is in 3,988 signals, replayed: what the numbers say.
One channel does clear 90 under our rule, at 91.3%, on 23 scored verdicts. Why the highest numbers in any such table sit on the thinnest evidence, and how little a count that size settles, is the subject of How many trades a hit rate needs before it means anything.
What happens when you put the claim beside the record?
Most of the claims have no record to sit beside. Here are all twelve channels in the archive that advertise 90% or more, ordered by how often they say it, with whatever our replay found for each.
| Claim messages at 90% or higher | Parsed signals we can score | Our one-take hit rate |
|---|---|---|
| 699 | none | no record exists |
| 599 | none | no record exists |
| 96 | none | no record exists |
| 7 | 9 scored | 77.8%, below our ten-verdict minimum |
| 6 | none | no record exists |
| 6 | 69 scored | 55.1% |
| 5 | 22 scored | 68.2% |
| 1 | 14 scored | 57.1% |
| 1 | 117 scored | 88.9% |
| 1 | 23 scored | 91.3% |
| 1 | 53 scored | 56.6% |
| 1 | 9 scored | 77.8%, below our ten-verdict minimum |
Eight of the twelve produce calls we can replay. Six of those clear the ten-verdict minimum, and one of the six reaches the advertised territory. The five rows at the bottom state the figure once each, and every one of them publishes calls we can score, which is close to the opposite of the pattern at the top.
Where do the loudest claims come from?
From the feeds with nothing in them to check. The four channels with no record at all supply 1,400 of the 1,423 claims of 90% or better between them, and not one of them publishes a call our parser can score.
Claim volume runs inversely to evidence here. The channels repeating a percentage hundreds of times are the ones whose posts contain no entry, no stop or no target to replay against, while the channels publishing complete calls mention a percentage once and move on. What a feed has to contain before any outsider can grade it is set out in What a channel must publish before anyone can rate it.
None of this establishes that those four channels are wrong about their own performance. It establishes only this: nobody, including them, can demonstrate the figure from what they published.
What does a channel's own scoreboard prove?
That its arithmetic is sound, which is a much smaller claim than it appears to be. One channel in the archive publishes session reports several times a day, each with the session's accuracy, its win count and its loss count, and it is the cleanest example of the whole problem.
Across 98 distinct reports covering 41 days between 26 June and 5 August 2026, the channel reported 1,399 wins against 59 losses. Every single report's stated accuracy equals its own wins divided by wins plus losses, to within rounding. Pooled over the period the figure is 95.95%, and 50 of the 98 reports state exactly 100%.
The division is right in all 98. What the report never states is where the trades came from, whether every call that day appears in it, or what made a trade a loss instead of an open position. A scoreboard kept by the player is a record of what the player counted, and the arithmetic being correct tells you nothing about the counting. None of that is evidence the counting was wrong. It is the reason nobody outside can tell. Building the count yourself from the feed takes about ten minutes and is described in How to check a signal channel's track record in 10 minutes.
Which number can you compare between channels?
Only one computed the same way for all of them, which in practice means one you or a third party computed rather than one each channel computed for itself.
Three questions turn an advertised percentage back into a number, and a channel that answers all three has told you more than the percentage ever could.
- Who computed it? A figure a channel produced about itself can be checked against its own posts and against nothing else. A figure an outsider produced can at least be pointed at a rival.
- Was the identical rule applied to the other channel? Two percentages sit on the same axis only when one party produced both. Ours are comparable to each other and to nothing a channel publishes about itself.
- Is the rule written down where you can disagree with it? A stated rule can be argued with. An unstated one can only be believed, and belief is what the percentage was selling in the first place.
A figure that survives those three questions can be argued about. A figure that does not answer them is a slogan with a percent sign, and it is worth exactly as much as the 1,299 messages saying 95%.
What this does not prove
The claim counts are marker searches over free and public channel archives, so they are a floor rather than a census: a claim phrased as "9 out of 10 calls land" does not match the pattern and is not counted. The archive is 42 channels, all of them accounts we index; the replay index is larger, at 47 accounts, so the claim counts and the replay figures are drawn from overlapping but differently sized sets. Nothing here is a random sample of the signal market.
The concentration caveat travels with every figure above. Two channels supply 1,298 of the 1,299 messages claiming 95%, so "1,423 claims" describes repetition inside a few feeds as much as breadth across many.
Our own count moved during this work, and it moved for a reason worth stating. An earlier run of the same search, over 39 dump files and 23,707 messages, reported 1,328 claims with 1,322 at 90% or better. Two things moved the number since: the archive grew to 42 files and 25,932 messages, and the pattern itself was wrong. It mis-read decimal figures such as "Accuracy: 100.00%" as claims of 0%, and read one channel's return figures as success rates. Run unchanged over today's archive, the old pattern gives 1,443 claims with 1,323 at 90% or better, so the growth explains the total and the fix explains the 90% bucket. Both faults are fixed, both counts are printed side by side in our working notes, and the corrected figures are the ones used here.
The replay figures inherit every assumption on the Methodology page, including the 10% fallback stop and the 30-day window, and the replay's own treatment of a candle touching the entry as a fill. They describe our index as of 17 August 2026 and they move as signals resolve.
Nothing here alleges misconduct by any channel, and a channel whose posts cannot be parsed is not thereby doing anything wrong.
The practical read
An advertised accuracy figure is not evidence about a channel until you know what it counted, and in this archive the channels that repeat the highest figures most often are the ones publishing the least that anyone could check.
Current ratings, with the scored count printed next to every percentage, are at Signal Providers, and what our verification does and does not claim is set out in the Verification Guidelines. Nothing here is financial advice.
Sources
- Advertising claims: our own archive.
data/channel_dumps/, 42 files, 25,932 messages dated 3 January 2021 to 12 August 2026, searched on 18 August 2026. Every claim figure above comes from that run: 1,445 claims, 1,423 at 90% or higher, 1,299 saying 95%, 65 saying 100%, and the twelve-row table of claimants. - Session reports. The same archive, same run: 98 distinct daily reports from one channel, 26 June to 5 August 2026, 1,399 reported wins against 59 reported losses.
- Replay figures. Our production database as of 17 August 2026: 3,988 parsed signals, 47 accounts, 1,935 scored verdicts at 67.44%, 17 rateable channels from 55.07% to 91.3% with a median of 68.18%, 1,892 as-published outcomes averaging -0.54%, and 2,022 signals carrying no stop.
- Rule set. Methodology: the One-Take Test, for the one-take rules, the 10% fallback stop, the ten-verdict minimum and the reason for a fixed target.
- Outcome distribution behind the replay figures. 3,988 signals, replayed: what the numbers say, snapshot of 17 August 2026.
- Excluded calls. Half the trades never happen, for the unfilled-entry share and its concentration in one publisher.
- Sample size behind a hit rate. How many trades a hit rate needs before it means anything, for the interval around a figure computed on 23 scored verdicts.