3,988 signals, replayed: what the numbers say

We replayed every parsed signal from 47 Telegram channels against real 1-minute candles. The largest single outcome is neither a win nor a loss: it is a trade a follower could not have entered, as of August 2026.

What we did

We parse the signals published by Telegram channels — symbol, direction, entry, targets, stop — and replay each one against real 1-minute candles from the exchange, following one fixed rule set: enter at the published entry with the whole position and no leverage, take profit at a 2% move, exit at the channel's own published stop, and give the trade 30 days to resolve. Fees are charged in the replay. The full rules, including what we do when a channel publishes no stop, are on the Methodology page.

This is a snapshot of what came out, as of 17 August 2026: 3,988 parsed signals across 47 channels, published between November 2021 and August 2026 across 497 trading pairs, of which 3,927 have produced a resolved outcome.

The outcome distribution

Two things stand out, and neither is the hit rate.

The largest single category is neither a win nor a loss. 1,958 outcomes, 49.9% of everything resolved, were cancelled: the market ran more than 2% past the published entry before a follower placing the order as written could have been filled. A separate and much smaller group, 31 signals, never saw the entry price at all.

That 49.9% needs its caveat stated in the same breath, because it is not spread evenly. One channel accounts for 1,501 of the 1,958 cancellations. Across the other twenty-one the rate is 22.79%. The honest sentence is "about one in five across most channels, and four in five in the worst one", not "half of everything, everywhere".

This matters more than it looks. A channel with a large share of unfilled entries is publishing calls that are unfalsifiable by construction: if the market moves, the channel called it; if it does not, nothing was risked and nothing counts against the record. Excluding them from an accuracy percentage — as we do — is the honest treatment, but it also means the percentage describes a minority of what the channel actually posted.

Of the signals that resolved into a verdict, 67.4% reached a first take before the stop - 1,305 against 578 stopped out and 52 that expired, from 1,935 scored. Read carefully, that number is less flattering than it sounds. A 2% take is a low bar, deliberately chosen so every channel answers the same question; it says nothing about how large the 578 losses were. A channel can clear this bar and still lose money for its followers, which is why we publish it as a reachability measure and label it that way on every page it appears on. Under the second, stricter replay that follows each channel's own published ladder, the average outcome across the 1,892 that carry a profit-and-loss figure is -0.54%.

Only 17 channels of 47 have a publishable record

Of the 47 channels in the dataset, 17 have enough resolved verdicts for us to publish a hit rate, as of 17 August 2026. The rest fail on one of three grounds: too few parsed signals, no usable price levels in their posts at all, or fewer than 10 scored verdicts — our minimum, because three signals at 100% is an anecdote wearing a percentage.

Among the seventeen, hit rates run from 55.07% (on 69 verdicts) to 91.3% (on 23), with the median at 68.18%. Five further channels fall below the minimum, on 5 to 9 verdicts each, and are shown as a dash instead of a number - two of them would otherwise read 77.78%, which is the point of the cutoff: the highest-looking figures in the set are the ones carrying the least evidence.

The spread between a hit rate on 23 verdicts and one on 420 is not a ranking of skill. It is mostly a ranking of how much we know, and how many trades a percentage needs sets out how much.

The clone cluster

The finding that changed how we read the rest of the data came from comparing signal streams across channels, which is a different exercise from scoring each channel alone.

Three separately branded "AI" channels published identical signals with 100% overlap, at a median lag of two to four seconds between them. All three still carry near-identical records: 296 to 297 scored verdicts each and hit rates of 56.9%, 56.9% and 56.76%, as of 17 August 2026, because all three are the same feed with different names on it.

Anyone subscribed to all three for diversification was paying three times for one stream — and taking three times the position size on every call, which is the opposite of what diversification is supposed to do. We now link channels that repost each other directly on their pages.

What this snapshot does not prove

• The sample is not random. These are the channels we could parse, weighted toward those that publish structured signals. Channels that post screenshots or voice notes are absent, and they are not a random subset of the market.
• It is a window onto one moment. Signals resolve continuously and the numbers move as more do. Treat the figures here as of 17 August 2026. This article was first published with 1,110 signals behind it, recomputed on 2026-08-07 against 3,227, and recomputed again on 2026-08-17 against 3,988; the shape of the distribution has held through all three.
• This recount corrected an error of ours, not just a stale date. Until 17 August our resolver judged each signal once, shortly after import, so a call published hours earlier was scored against candles that necessarily stopped at that day - and "neither within 30 days" was being recorded on evidence covering hours. 119 verdicts moved when we fixed it, nearly all of them out of that row and into a reached take or a stop. The hit rate above is 1.8 points higher than the same measurement made on the faulty data, and every per-channel figure on this site was recomputed the same day.
• Parsing is imperfect. Free-text signals in unusual formats can be misread. Channels can dispute a specific calculation, and when we fix a parser we recompute every channel under the same rules, and never adjust one by hand.
• A hit rate is not a return. It counts how often a call went 2% the right way before the stop. What following the channel would have earned is a separate calculation with fees, sizing and slippage in it.

The practical read

If you are evaluating a channel, three questions come out of this data directly: how much of what it publishes ever becomes a fillable trade, how many scored verdicts sit behind the percentage it advertises, and whether its signal stream is its own. The first two are on every channel page we publish; the third we now detect and label.

Channel ratings are at Signal Providers; the warning signs that appear before the numbers do are in 10 red flags of a signal channel. Nothing here is financial advice.