More statistical power arrived. The funerals sped up.

01What ‘needs more data’ turns into when the data shows up

Three hypotheses about trader positioning — long/short ratios, the ‘follow the smart money’ family — were retired in low-power limbo earlier this summer, with t-statistics around one and sample windows of three to six weeks. Exactly the kind of result a newsletter calls an early signal. The first one was re-tested on 1,129 days instead of 37, under a criterion frozen before looking: the t-statistic went from +0.93 to −0.27, seven of fifteen symbols positive, sign opposite to the pre-registered one. It did not consolidate; it regressed to nothing. The promising number was sample noise wearing a suit.

The cross-sectional cousin is the better story. It was shelved in June on a 21-day lockbox at t=+1.09 — also ‘promising’. With the panel grown 124-fold, to 8,326 rebalances, the t-statistic finally cleared the significance bar at −1.90. Note the minus: significant in the opposite direction from the pre-committed sign, and stable across both halves of the window. Following the smart money in this dataset is not merely unprofitable; it is measurably backwards. And no, the sign does not get flipped and re-sold as a fresh discovery — the researcher has already seen the result, so any ‘contrarian’ version of the same panel would be pre-registration theater. Falsified, closed.

The middle case is the instructive one. The four-hour variant actually survived the cheap re-screen — ten of fifteen symbols, t=+1.4 — and for about six hours was the only living hypothesis on the books. Then the real walk-forward gate ran it with costs: zero passes out of thirty runs, negative out-of-sample Sharpe on every single symbol, long-short and executable long-only alike. A signal that survives a correlation screen and dies at the cost model is the most common corpse in this graveyard, and the reason screenshots never mention fees.

That closes the entire positioning-ratio family at full power: five hypotheses, zero survivors, each dead for a documented reason — noise, costs, or an inverted sign. The lesson travels well beyond crypto: ‘promising, needs more data’ resolves to ‘no’ almost every time, and occasionally to ‘no, and it was backwards’.

02Then a Journal of Finance anomaly took its shot

The rules forbid testing a ninth variation of momentum without a genuinely new economic reason, so the price-and-volume family got exactly one more candidate, and it arrived with credentials: George and Hwang’s 52-week-high effect [1] — investors anchor on an asset’s own one-year high, anchoring being the bias Tversky and Kahneman named half a century ago [2], absorb good news too slowly near that reference point, and leave a drift behind. In US equities it famously subsumes plain momentum. It is also the only signal in this whole series that uses the full path of price rather than two endpoints. If anything built from price alone deserved one last test, it was this.

Pre-registered before running: sign fixed — near the high goes long — permutation p below 0.01 on the executable long-only variant, and it must beat simply holding the fifteen coins equal-weighted. The result is the most nuanced funeral of the family: the sign came out right, and the long-short book posted an out-of-sample Sharpe of 0.85 with a permutation p of 0.025 — the closest any price-and-volume mechanism has ever come here. On most of the internet, that line is a victory-lap thread. Under the frozen criteria it is a rejection: 0.025 is not 0.01, and the bar was set that tight before running precisely because the best of eight related attempts has a very real chance of hitting 0.025 on luck alone. The blunter fact settles it anyway — the information ratio against equal-weight was negative in both variants, and the executable long-only version’s permutation p was 0.515. A coin flip with excellent academic references.

So the scoreboard moves to fifty-one pre-registered, fifty-one falsified, and the price-only conclusion graduates from provisional to formal: raw, idiosyncratic, or path-dependent, at every horizon tried, price alone carries no exploitable net-of-cost signal in this universe of fifteen liquid coins. Four categories of mechanism — return dynamics, attention, a liquidity premium, and now anchoring — tested, documented, closed.

03Making the receipts tamper-evident

The quiet runs in between went into the referee-grade version of all this: a statistical appendix rebuilt from the database rather than from memory. Its multiplicity section is the one worth quoting. Take every comparable p-value across the 58 registered hypotheses and apply no correction, Bonferroni, or a Benjamini–Hochberg false-discovery-rate correction [3]: the verdict is identical under all of them. Zero survive. There is no candidate close enough to the margin for the choice of correction to matter — which is its own kind of result.

The power analysis in the same appendix explains why this page keeps siding against the screenshots: at this project’s own thresholds, certifying a Sharpe-1.2 edge would take on the order of twenty-one years of out-of-sample data. For calibration, Bitcoin buy-and-hold — the backdrop every green screenshot rode — scores a deflated Sharpe of 0.21 under the same lens. A weekly performance claim is not early evidence of anything. It is astrology with an API key.

And the pre-registration itself stopped being take-my-word-for-it: since late August the research repository’s head gets anchored weekly into Bitcoin via OpenTimestamps [4], so ‘every threshold was fixed in advance’ is now checkable by a third party who has never seen the machine. When a research program’s headline result is that nothing works, the least it can do is make that claim impossible to quietly rewrite.