Do crypto signals actually work? I graded 305 of them against exchange data

Correction, 1 August 2026. This post states that source errors are never repaired - that a typo stays a typo and the signal goes down as ungradeable. An audit run the day this went live found one signal where that did not hold: a language model silently corrected a misplaced decimal in a target, and the signal was graded rather than voided. That row has since been voided.

The audit also established something the post does not say: outcomes are graded from a model's rewrite of each signal, not from the original message. 307 of 311 rewrites were byte-identical to their source on every price. Four were not.

No figure below has changed - the voided row moved the aggregate by less than a hundredth of an R. What changed is that the description of the method is now accurate. The full account is in the limitations section of the scoreboard.

Posted by the operator of RebateMax

Search this question and you will find a dozen articles that all end the same way: demand a transparent track record, don't trust advertised win rates, look at the data.

Not one of them shows you any data.

That is the whole industry in miniature. Everyone knows the right test and nobody submits to it, because the test is expensive — you have to publish the trades that went wrong, and no amount of good writing makes a losing month look good.

So here is the data. Every signal RebateMax has forwarded, graded automatically against exchange candles, published at rebatemax.trading/scoreboard and updated as the resolver runs.

This post is the answer to the question, and the answer is more interesting than yes or no.

The setup

Seven Telegram channels. Every signal they post gets pulled in, parsed, and scored against one-minute candle data from the exchange. No human decides which trades count.

The scoring rules were frozen before there was any data to score, which is the only version of this that means anything. A signal fills when price touches its entry zone. Stops and targets resolve on touch. If a candle contains both the stop and a target, the stop wins — the pessimistic reading, chosen deliberately. Seven-day timeout. Source errors are never repaired; a typo stays a typo and the signal goes down as ungradeable.

Everything is measured in R — profit or loss in units of the risk taken on that trade. A signal that hits its stop is exactly −1.00R by construction. R is the metric those other articles keep telling you to ask for, and they're right to.

305 graded trades. Five weeks. 18 more that couldn't be scored at all.

The answer depends on one decision

Here is the finding, and it reframes the question.

Same signals. Same candles. Same rules. One is a profitable set and one is not, and the only thing separating them is whether you close part of the position when a target prints.

That is not a subtle effect. It's the difference between a strategy worth following and one that slowly bleeds.

What causes it: 114 trades

Of those 305 trades, 114 touched a profit target and then reversed into the stop.

Book at the target and they're wins. Hold and they're full losses. One decision, 114 trades, and the entire record flips sign.

This is the mechanism behind every argument about whether signals "work," and almost nobody separates it out. A signal that goes 60 pips your way before turning around is a good call and a losing trade at the same time. Which of those it is on your statement depends entirely on you.

Why the win rate everyone quotes is the wrong number — including mine

Under the scale-out convention, 69 of 305 trades were outright losses. Stop hit, no target ever reached.

That's a 23% loss rate. Which means, if I wanted to, I could write "77% win rate" on the homepage in large type.

It would be technically true and completely misleading, and it's worth walking through why, because this is exactly the trick the guides warn about without showing you an example.

A "win" in that number includes trades that reached the first target and reversed. It includes trades that returned +0.01R — a cent on a dollar of risk. A win rate counts events, not money. Push your first target close enough to the entry and you can manufacture a 90% win rate out of a losing strategy, which is precisely how the 95%-accurate channels do it.

Under the other convention, the same trades produce a 60% loss rate. Identical data, wildly different headline. That's how much room a win rate leaves for creativity.

The R figure can't be gamed the same way, because moving your target closer shrinks the R you collect. That's why it's the number worth publishing and the number worth demanding.

The channels differ enormously

Seven sources, and the spread is wide enough that "do signals work" isn't answerable as a category.

The strongest performer by volume averages +0.494R scaling out and −0.002R holding — a channel whose entire edge lives in the decision to take profit. The weakest averages −0.427R and gets worse if you hold. One goes from +0.221R to −0.621R on the convention alone.

Only two channels are positive under both conventions, and both have fewer than fifteen graded trades — which is to say the data cannot yet tell you whether that means anything.

Several have so few graded trades that their confidence intervals span zero, meaning the data genuinely cannot say whether they're good or bad yet. Those are published with that stated plainly, because a mean over nine trades is decoration, not evidence.

The useful conclusion is not "signals work" or "signals don't." It's that source selection and exit discipline both matter more than the existence of a signal, and only one of those is something a provider can sell you.

What this data does not tell you

Four limits, named because leaving them out is how a track record becomes marketing.

Every figure is gross. No fees, no funding, no slippage. A strategy averaging a small positive R before costs can be negative after them. On leveraged positions held for days, funding alone can exceed the trade fee several times over.

Five weeks is not a track record. It's enough to see the shape of the thing and nowhere near enough to project forward. Anyone quoting a month of data as proof of anything is overreaching, including me.

Mid-trade management isn't modelled. Two of the seven channels post follow-up instructions — take partial here, move the stop to entry. The grader doesn't read them. It scores the original signal as posted, which both misses the protection those instructions give and credits runs that active management would have exited earlier.

These are the signal's levels, not your fills. Slippage, latency, position sizing, and whether you took every signal or picked among them all change the outcome. Nobody's real record matches a backtest, and this is closer to a backtest than to a record.

The conflict, declared

Signals are graded against exchange candles in this order: Binance futures, then OKX, then BingX.

BingX is our referral partner. We earn commission on trading fees generated there, so grading partly against BingX data is a conflict and you should know about it before you weigh any of the numbers above.

It sits third in the order so it's only reached when the neutral sources can't cover a pair — mostly smaller alts the other two don't list. Before it was added those signals couldn't be graded at all, and their absence made every channel look better than it was, because the pairs that go ungraded skew toward the thin and volatile. Adding it moved most channels' numbers down.

So, do they work?

Sometimes, from some sources, if you take profit at targets, before costs, on five weeks of data.

Every clause in that sentence is doing work. Strip any of them out and it becomes the kind of claim this whole post exists to argue against.

The more useful question isn't whether signals work in general. It's:

Does the provider publish losses? Not "we're transparent" — actual losing trades, visible, in the same place as the wins.

Do they publish R, or only a win rate? A win rate without average win and average loss beside it is a number chosen because it flatters.

Can you verify a single one? Take any signal, find its entry, stop and targets, pull the one-minute candles for that pair, and see which level price touched first. That's the entire method. It takes about five minutes.

If you run that check on our scoreboard and find one graded wrong, say so publicly and it gets corrected there with a note.

Why publish this at all

Because the honest version is a worse advertisement and a better filter.

A trader who reads this and decides five weeks isn't enough evidence has read it correctly. A trader who notices that the −0.167R figure is the one that applies if they don't take profit, and adjusts, has got more out of it than any win rate would have given them.

RebateMax makes money when someone registers on BingX through the referral and trades — 50 of the 60 commission points BingX pays us go back to the user as a lifetime fee discount. That model works best with traders who stay, and traders who stay are the ones who checked the math first.

The scoreboard exists because it's the one claim a competitor can't copy without publishing their own losses.

---

Next in the series: BingX versus Bybit — fee schedules side by side, and which venue suits which kind of trader.

Past results do not predict future results. This is general information, not financial advice. RebateMax aggregates third-party signals and does not generate them. Figures are as published on the scoreboard and change as the resolver runs — check there for current numbers.