ichnoraDon't trust us. Check the tests.
Every model looks good on the data it was built on. So we froze the models on data up to Sep 2025 and tested them on the next twelve months (2025-10-02 to 2026-09-23), a falling market where Bitcoin lost about 30%. All returns are after 0.3% trading costs per coin. Here is what held up, and what didn't.
per call above comparable random coins (90% range +1.0% to +7.7%), positive from every start day. Passed all four tests.
per call above comparable coins, but the range includes zero (-0.1% to +2.2%) and at 1% trading costs the average shrinks to +0.1%. A watchlist, not a trade.
average per call on this site (281 graded calls) vs +1.1% for the top-200 average and +0.1% for Bitcoin. Earlier calls came from older model versions.
1 · Is it just picking wild coins?
Momentum models like volatile coins, and volatile coins land in the top-10 gainers more often by chance. So each pick is compared with a random coin from the same day with similar volatility and size (200 random draws per day).
| Avg return per call | Coins that went up | Top-10 hit rate | |
|---|---|---|---|
| 7-day picks | +3.7% | 43.0% | 17.6% |
| Matched random coins | -0.4% | 41.1% | 8.6% |
| Top-200 average | -0.5% | 43.1% | 5.7% |
| Bitcoin | -0.5% | – | – |
Matched random coins hit the top 10 more often than pure chance (8.6% vs 5.7%), so part of the headline hit rate is volatility. But they lost money on average (-0.4%) while the picks made +3.7%. The edge in returns is not explained by volatility or size.
The picks do not go up more often: 43.0% of picked coins rose vs 43.1% of all coins. The gain comes from the winners rising much further than the losers fall. Hold all 10 in equal size; single picks are a coin flip.
2 · Is it luck from a few lucky weeks?
Daily 7-day calls overlap, so 356 calls are really about 51 independent weeks. We re-ran the test on non-overlapping windows only, for every possible start day, and resampled the weeks 4,000 times to get a 90% range.
| Edge per call | Average | 90% range | Independent windows | Start days positive |
|---|---|---|---|---|
| 7-day vs matched random coins | +4.1% | +1.0% to +7.7% | 51 | 7/7 |
| 7-day vs top-200 average | +4.3% | +2.1% to +6.6% | 50 | 7/7 |
| 7-day vs Bitcoin | +4.1% | +1.5% to +6.7% | 50 | 7/7 |
| 3-day vs matched random coins | +1.0% | -0.1% to +2.2% | 120 | 3/3 |
| 3-day vs top-200 average | +1.1% | +0.1% to +2.2% | 120 | 3/3 |
| 3-day vs Bitcoin | +1.0% | -0.1% to +2.2% | 120 | 3/3 |
For the 7-day calls the whole 90% range is above zero against every yardstick, and every one of the 7 possible start days is positive. For the 3-day calls the range touches zero against comparable coins and Bitcoin: not proven.
3 · Does the score mean what it says?
The 7-day score is the model's chance that a coin finishes in the top 10 of ~200. A good score should come true about as often as it says.
■ what the score said ■ what happened · grouped by score, all ~200 coins, held-out days
The 10 picks scored 17.7% on average and actually hit 17.6% of the time. Lower scores run a little optimistic (they say ~4%, reality ~3%), so read small scores as "unlikely", not as exact odds.
4 · Can you actually trade it?
| 7-day calls | Result |
|---|---|
| Median 24h trading volume of picks | $28M (all coins $25M) |
| Picks under $1M / $5M daily volume | 1.9% / 9.9% |
| Thin coins (<$1M) swapped for the next pick | +3.6% per call |
| At 0.3% / 1% / 2% round-trip costs | +3.7% / +3.0% / +2.0% per call |
| Bought a full day after the signal | +3.5% per call (edge vs market +1.7% to +6.4%) |
The picks are as liquid as the average coin, survive 2% trading costs, and still work if you buy a day late.
The market light
TRADE when Bitcoin is flat or up over 7 days, CAUTION when it is down. Held-out year, 7-day calls: +5.6% per call on TRADE days vs +2.1% on CAUTION days, with smaller drops on TRADE days. The difference is likely but not proven (roughly 88–94% in resampling), the site's live record shows the opposite so far, and it had no effect on 3-day calls. One input, not a rule.
What could still be wrong
- One test year. Twelve months, mostly a falling market. Other conditions may behave differently.
- Survivor data. The history holds coins that are in today's top 200, so coins that collapsed and dropped out are missing. This flatters every basket, picks and comparisons alike; the comparisons above are fairer than the raw returns.
- Big-winner dependence. A few large moves drive the average. A stretch without them will look poor.
- Live record mixes model versions. Calls before Oct 20, 2026 used earlier models. The live numbers will settle the question over time.
Everything is graded in public: every call, 7-day dashboard, track record as JSON. Not financial advice; crypto can lose all of its value.