Rally Radar
Does a Crypto Scanner Work in a Bear Market? We Ran Ours Over Six Years
Every number we had published about our scanner came from one stretch of market: April to August 2026. That is the obvious hole in the one positive result we have, and we said so at the time. A ranking that concentrates movers in one kind of market may do nothing in another. So we went back to the start of 2020 and asked the same question of every year since, bull and bear. Then, with six years of data to train on, we tried to beat our own hand-written score with a machine-learning model.
The question, held fixed
The easiest way to get a flattering answer from six years of data is to ask it many questions and report the best one. So we asked one. It is the cell we already publish on the track record: among coins whose net score reached 85 or more, how many touched +10% within 24 hours, compared with coins the scanner did not flag, priced in the same hour, with similar volatility, momentum and liquidity? That last comparison is the one that matters. The score is built from volatility and momentum, so beating random entry is cheap. Beating an equally jumpy coin at the same moment is not.
Nothing else was computed. The threshold, the target and the window were fixed before the history was downloaded. Each year was scored by the same code the live tool runs.
The data
Binance publishes its full candle archive, so we pulled hourly candles for 667 USDT pairs from January 2020 to September 2026, about 18.5 million bars. The live tool scans USDC pairs, which are a smaller set of the same coins; the price action is the same asset, the liquidity ranks are not identical. Scoring every bar of every coin took a little over six hours.
One honesty note before the table. The score was tuned on spring 2026, so 2020 to 2025 is data it never saw. But the people who built it had lived through those markets, and a rule like "volume climbing while the range tightens" is folklore that came out of them. This is a check across regimes, not a blind test.
The result, year by year
The last column is the one to read. It is the lead of 85+ coins over their matched look-alikes, in percentage points, with a 95% interval built by resampling whole days, because three hundred altcoins rallying in one afternoon are close to one event, not three hundred.
| Year | Bitcoin | 85+ episodes | 85+ reached +10% in 24h | Random entry | Lead over matched look-alikes |
|---|---|---|---|---|---|
| 2020 | +303% | 2,599 | 21.7% | 14.9% | +0.9 pp [−1.7, +3.7] |
| 2021 | +59% | 4,299 | 28.4% | 22.9% | +4.1 pp [+2.2, +6.0] |
| 2022 | −65% | 5,372 | 18.7% | 12.9% | +4.4 pp [+3.0, +5.8] |
| 2023 | +156% | 7,984 | 15.6% | 8.0% | +4.1 pp [+2.9, +5.2] |
| 2024 | +120% | 6,183 | 15.8% | 10.2% | +4.1 pp [+3.1, +5.2] |
| 2025 | −7% | 5,938 | 13.9% | 9.1% | +3.0 pp [+1.8, +4.2] |
| 2026 (to Sep) | −5% | 6,145 | 16.0% | 8.2% | +4.8 pp [+3.6, +6.0] |
| All years | — | 38,520 | 17.7% | 11.4% | +3.9 pp [+3.4, +4.4] |
Three things stand out.
The bear market is not the exception. 2022 is the year Bitcoin lost two-thirds of its value, and it is one of the better years in the table. The raw touch rate fell, as it should when the whole market is sinking, but so did the look-alikes' rate, and the gap between them stayed. The scanner was never meant to tell you whether the market is going up. It sorts the coins that are about to move from the ones that aren't, and that sorting was about as good in a crash as in a boom. 2025, a flat year, is the weakest of the six that cleared zero, and it still cleared it.
2020 did not clear zero. The matched lead that year is under one percentage point, with an interval that spans zero. The raw gap was large, close to seven points, so the score looked fine until it was compared with coins of the same volatility. 2020 was the thinnest year, with 188 coins and the fewest episodes, and the one where almost everything rallied at once. We are reporting it because it is in the table, not explaining it away. A single year coming back near zero could be noise in the thinnest year of the set, or it could be a year the effect simply wasn't there. The data can't tell those apart.
The lead is small and steady. Four percentage points on a base rate of eight to twenty is not a dramatic number. It is the same size, give or take, as the five-month result we published in October, and that is the point of the exercise. Had the six-year number come back at zero, or at fifteen, the October result would have been the thing to doubt.
Then we tried to beat it with a model
Six years of scored history is also a training set. If a hand-written score made of eight signals can separate movers from matched look-alikes by four points, a model with 50 inputs and millions of examples should do better. We wrote the test down before training anything: a gradient-boosted classifier trained on 2020 to 2024, tuned on 2025, and judged once on January to September 2026, which it never saw. The comparison was paired by hour. Each hour, the hand score picks its top ten coins, the model picks its top ten from the same eligible set, and we ask how many of each reached +10% within a day. The model wins only if the per-hour difference has a day-clustered interval entirely above zero.
It won. Over 4,859 hours on 272 entry days, the model's top ten reached the target 12.5% of the time against the hand score's 10.8%, a paired difference of +1.6 pp with an interval of [+1.3, +2.0]. Random entry over the same hours was 8.1%. The two lists barely overlapped, sharing less than one coin per hour on average.
And then we looked at the matched row, the one the whole post has been about:
| Holdout, Jan to Sep 2026 | Hand score | Model |
|---|---|---|
| Top-10 picks reached +10% in 24h | 10.8% | 12.5% |
| Lead over matched look-alikes | +2.5 pp [+2.2, +2.8] | +1.7 pp [+1.4, +2.0] |
The model picks coins that reach +10% more often. But once each pick is compared with an unflagged coin of the same volatility, momentum and liquidity in the same hour, the model's lead is smaller than the hand score's. The two findings are not in conflict. The model learned, correctly, that the cheapest way to hit a +10% target is to pick coins that swing 10% all the time. A single input, the coin's recent average true range as a share of price, accounted for 65.7% of everything the model learned. It is mostly a volatility filter with a very good fit.
That is exactly the failure mode we built the matched comparison to catch, and the reason the win rate was never the metric. So the model is not live. The hand score still ranks the site, and will until something beats it on the matched row rather than the raw one.
What happens next is measured in public
Everything above is history. The remaining question is whether the live tool, scanning mid-bar on USDC pairs, with real gaps and real latency, keeps doing in the future what the archive says it did in the past. Since late October, the track record page has carried a forward scoreboard: every hour, the top ten picks are recorded, and 24 hours later they are graded against matched look-alikes from the same scan. Nobody can tune a number that is written down before the outcome exists. If the lead drifts to zero, that page will say so before we do.
What this does not say
Nothing here is a profitability claim. A touch means the price reached +10% at some point in the window. There is no stop, no exit and no fee. The out-of-sample test of the score as a trading strategy found no edge, and nothing in six years of touch rates changes that. A coin that reaches +10% can first fall 20%.
The yearly table is descriptive. There is no holdout in it and no pre-registration beyond having fixed the cell in advance. The model comparison was pre-registered and read once; the yearly table was not a test, it was a look.
Different pairs, different entries. The archive is USDT pairs entered at bar close; the live tool scans USDC pairs mid-bar. Absolute rates will not match the live scoreboard's. Only comparisons within the same data, such as hand score against model on the same hours, are like-for-like.
It is not a forecast for any single coin. Four percentage points across thirty-eight thousand episodes says nothing about the next one.
So, does a crypto scanner work in a bear market? Ours sorted movers toward the top about as well in 2022 as in any other year, and that is the honest size of the claim: a sorting aid that holds up across regimes, not a signal that makes money in them. Every method and every table, including the rows that went against us, is on the track record.
See what is at the top of the list now
Rally Radar scans 200+ Binance pairs every 15 minutes and ranks them by net score, with the signals behind each one. Use it to decide what to look at first, and do the deciding yourself.
Open the scanner →For the five-month result this post set out to check, read how to find altcoins about to move. For which of the score's components actually carry it, see RSI, MACD and volume together. And for the tests any scanner's claim has to pass, do crypto scanners actually work lays them out.