The Lab

1. Basis points (bps). A hundredth of a percent: 100 bps is 1%. 10 bps on $10,000 is $10. Buying and selling once (a round trip) at typical fees costs about 10 bps, so an effect smaller than that loses money.

2. The five groups (fifths). Take every hour of every coin and sort them by the signal, lowest to highest. Cut the pile into five equal groups. For each group, average the price move that followed. Each bar shows its group's move above or below the average of all hours, so a market that simply rose all month doesn't count. If the idea is right, the bars climb (or fall) steadily from one end to the other.

3. Rank correlation (Spearman's). One number for "when the signal is higher, is the move after higher too?" Put the hours in order by the signal (1st, 2nd, 3rd…), and again by the move that followed, then compare the two orders. Five hours in exactly the same order in both give +1; exactly reversed, −1; no pattern, about 0. Markets are noisy, so real effects are tiny: 0.03 held for years is notable. We work it out for each coin, then take the average over the coins.

4. t: more than luck? Even with no real link, an average comes out a little above or below zero by chance. How much it usually wobbles (the standard error) can be worked out from how much the coins disagree with each other. t is the average divided by that wobble. For example, an average of 0.02 with a usual wobble of 0.005 gives t = 4: four times what luck usually makes. Luck alone passes ±2 about 1 time in 20, and ±3 only a few times in 1,000. We ask for ±3, because coins move together and neighbouring hours overlap, which makes luck look surer than it is.

5. "The price bouncing back" taken out. Some signals mostly echo the price's last move: a fall, then a bounce. We remove what that last move already explains and check that what is left still lines up with the move after, with t still beyond ±3.

6. Many settings tried. Try twenty settings and one will look good by luck, so every setting tried is shown under "All the numbers", not only the best.

7. Old data and new data. The settings tested (distances, horizons, thresholds) were set on 12 Oct 2026, after a first look at the numbers (a study added later gives its own dates). The record from 13 Oct 2026 on was unseen when they were chosen, so it is the fair test: each study judges it apart, as "holds on new data", once there is enough of it.

The verdicts. Holds up: every test passed. Real, but small: every test passed but one: the effect is too small to pay for a trade's costs. Partly: more than chance, but some tests failed. A hint: t beyond ±2 but not ±3: perhaps something, perhaps luck. No sign: nothing the record can tell from chance. Too early: under 2 weeks of record. No reliable edge: a write-up's word: what looked good did not hold up against chance or on unseen data. No edge after fees: a write-up's word: there is something, but trading costs eat it. Not shown: a write-up's word: the record could not show that the idea helps. Passed, on a weak bar: a write-up's word: it cleared the bar set in advance, but the bar proved easy; the stricter reading, added after, is shown apart. No reliable gain: a write-up's word: the versions that looked better did not clear the bar set in advance: all three coins, at both a day and a week. Failed on fresh data: a write-up's word: a lead from one study, tested again on prices kept apart, did not hold. Not proven, a lead: a write-up's word: it failed the test set in advance, but pointed the same way throughout; worth a fresh test. Holds, with the coins paid for: a write-up's word: it held up through every test, but only as set out: the coins bought outright, nothing borrowed. Not confirmed; the lead holds: a write-up's word: on fresh data it fell just short of the bar set in advance, but still pointed the same way.