Research note · Cross-asset testing
We published a failed test. Then we found out why it failed.
The strategy was tested on Ethereum, lost money, and we said so in the published description. Re-measured under control, the failure turned out to be ours: we had run Bitcoin’s stop widths on an instrument that moves about a third more per bar.
What we published
Selepro Veltus is a trend system built and validated on Bitcoin perpetual futures, on the 5-minute chart. Before publishing it we tested whether it survived being pointed at a different asset. It did not. So the script description said so, in the section listing what the system does not do:
“It does not transfer to other assets. Tested unchanged on ETHUSDT.P across the same five years, it lost or broke even in four of them.”
That sentence went out with the product, before anyone had paid for anything. It is the kind of line most vendors leave out, and leaving it in is the whole reason this business exists: a record you can only see when it flatters the author is not a record.
The part we got wrong
Read that sentence again and notice the word unchanged. It is doing more work than we realised when we wrote it.
The system does not only decide when to enter. It also decides how far to let a position run against it before cutting, how far to let it run for before taking profit, and how tightly to trail the remainder. Those distances are expressed as percentages of price, and they were calibrated on Bitcoin.
Ethereum is not Bitcoin. By average true range on the same timeframe over the same period, it moves roughly a third more per bar. Running Bitcoin’s stop distances on it means positions get cut by ordinary noise before the move they were waiting for has a chance to happen.
So the test measured something. It just did not measure what we said it measured. “This asset has no edge” and “these stops are too tight for this asset” produce the same losing equity curve, and we reported the first one.
The controlled re-measurement
To separate the two explanations you need to change exactly one thing. So: the same code, the same instrument, the same year, the same commission and slippage, the same entry logic, the same efficiency-ratio thresholds. Only the risk distances changed.
The year re-run was 2023. One year is not a cycle, and this note does not pretend otherwise — the point of a single controlled year is to establish whether the variable matters at all, not to size the effect.
| ETHUSDT.P, 5m, 2023 | Net | Profit factor |
|---|---|---|
| Bitcoin’s risk widths (the original test) | −10.17% | 0.828 |
| Risk widths calibrated on Ethereum | +7.12% | 1.118 |
Seventeen percentage points, from one variable. The entry logic never changed. Whatever the classifier and the trend filter were doing on Ethereum, they were doing it the whole time; the exits were throwing the trades away before it mattered.
What the widths actually are
The interesting part is not that the numbers differ. It is that they differ in a specific, structured way.
Both instruments use the same internal template: a first stop at some width S, a wider secondary stop at 2S, a first target at 1.8S, a trailing distance of S, and an asymmetric short target at 2S. That shape was never imposed — it fell out of calibrating each asset independently. Only the scale differs: S = 1.5% on Bitcoin, S = 2.0% on Ethereum.
Expressed against each instrument’s average true range, the two calibrations land within 0.9% of each other — coefficients of 0.3277 and 0.3249. Two independent optimisations, on two different assets, over five years, arriving at essentially the same number relative to volatility.
Then we tried to use that, and it failed
A coefficient that stable is tempting. If risk width is really just a function of an instrument’s volatility, it should be possible to predict the right width for an asset that has never been calibrated, and skip straight to the answer.
We wrote that prediction down before running it — which matters, because a hypothesis recorded after seeing the data is not a hypothesis — and tested it on Solana.
| SOLUSDT.P, 5m, five years | Net |
|---|---|
| Risk width predicted by the volatility relationship | −31.1% |
And it was not that the predicted width was a bad choice among good ones. Widths either side of it were tested as well. On Solana there was no width that worked.
So the relationship survives as a way to scale risk between instruments that already work, and dies as a way to answer the question anyone would actually want it to answer: does this system work on this asset? That question still has to be paid for in full, one asset at a time.
Five hypotheses were put to the test in this body of work. Four died. Three of them were about risk width: the volatility relationship as an asset passport, above; stops set as a multiple of a rolling ATR, which turned out to work backwards, because a fixed percentage already narrows in ATR terms exactly when volatility rises, and that is the part that helps; and a stop that switched with the measured market regime, which improved Ethereum in five years out of five and would have cost Bitcoin nineteen points. That last one was dropped on the Bitcoin number alone.
Two things this cost us, worth writing down
A parameter that looks good in one year falls apart over five
This happened three separate times in the same body of work. In each case a value looked clearly best on the year in front of us and clearly worse once the full cycle was run. Nothing gets adopted now without the complete period, and the discipline is not optional: the whole appeal of a shortcut is that it saves the check that would have caught it.
With five observations you cannot measure a correlation
At one point we measured a correlation across five annual figures, said out loud in the working notes that five observations are not enough to measure a correlation, and then drew a conclusion from it anyway. Measured properly on monthly data, the number was different enough to reverse the conclusion. If it cannot be measured, it does not get concluded.
What we are not publishing yet, and why
The obvious question is what the corrected Ethereum calibration does across the full five years. We have that number. It is not on this page.
Two checks are still open: a five-year control run with Bitcoin’s widths, so the comparison covers the whole cycle rather than the single year shown above, and a cross-venue validation on a second exchange. Until both are closed, publishing a headline five-year figure would mean asking you to trust a number that has had less scrutiny than the ones already on this site.
That is the same rule that produced the mistake this note is about. The original cross-asset test was published in good faith on a measurement that had not been controlled properly. The correction is not to measure less carefully next time and hope — it is to say what is known, say what is not, and wait. The figures will be published when the checks are done, whatever they say.
What changes for the product
The description of Selepro Veltus currently carries the sentence quoted at the top of this page. Its framing is now wrong and will be corrected in the next update of the published script, alongside the calibrated second market.
The sentence that follows it — that tested unchanged, it lost or broke even in four years out of five — remains true exactly as written, and stays. It was never the false part.
Nothing in this note changes the Bitcoin record: the same five years, the same two losing years, the same 28.1% worst drawdown across the period and about 35 months spent below the April 2022 peak. Those figures were measured against the published script and are unaffected by anything here.