← zhqdsk.app
Backtest notes: what did not hold
Every performance figure we have ever put on this site, along with the
result of re-running it on data we had not tuned on. Including the failures.
Why this page exists. A backtest that only ever gets run once is not
evidence, it is a screenshot. If we publish a number, we should also publish
what happens when someone else runs it.
Note 1 — The default template
This is the figure that appeared on our landing page for several weeks:
Claimed
t = 4.33
Universe195 perpetuals
Span712 days, 178 periods
Total return+189.8%
Max drawdown10.52%
Re-run on data we had not tuned on:
Binance — 74 perps, 250 days, default weights
t = −1.04
Hyperliquid — 176 perps, 104 days, default weights
t = +0.13
Hyperliquid — best of 8 factor sign combinations
t = +1.55
It did not reproduce. We think three things went wrong, roughly in order
of how much they matter:
- Overlapping windows. A 96h hold on 4h bars means consecutive
periods share most of their holding window. The effective sample size is
far below 178, so the reported t-statistic was never real.
- One funding regime. The original window had persistently positive
perpetual funding. The funding term was carrying most of the result.
- 712 days of one universe. The longer the window and the narrower
the universe, the more likely you are fitting a regime rather than an edge.
What we changed: the claim came off the site. The tool stayed. It is sold as
a backtester, not as a strategy.
What we would want from you
If you have done cross-sectional work on perpetuals, how do you handle the
overlap problem properly? We used a crude 4-block split and we do not fully
trust it either. Write to
[email protected].
Reproducing this
The engine is shipped inside the client and runs on your device, so you can
re-run any of the above and check us. Data hash for the original run:
c2a80ffb85a74c95.
Last updated: 2026-10-22 ·
back to site