Evidence
Two competitors sell price forecasts with no published validation anywhere on their sites. This page is the alternative: the measurements, the denominators under them, and the register of numbers we have not measured at all.
The zero is the number to weigh. Every backtest figure we publish was chosen and scored on the same data, which is closer to memorising a window than to proving an edge. We are recording continuously; when there is enough history to score settings against data they never saw, that figure changes and this page changes with it.
These are invented priors. They are named with an ASSUMED_
prefix so a scan finds them, every signal they price carries
profit_basis="assumed", and a test refuses to let any of
them into the set that publishes by default.
| Strategy | Constant | Value |
|---|---|---|
| event | ASSUMED_EVENT_UPLIFT | 0.2 |
| seasonal | ASSUMED_FEAST_UPLIFT | 0.15 |
| mayor | ASSUMED_PERK_UPLIFT | 0.3 |
This list has gone down once. Two strategies that rested on priors were dismantled rather than labelled, and their measurements are served as instruments instead. A prior removed beats a prior disclosed.
Anyone can show a green number. The same day, run four ways: doing nothing returns 0.000%, random entries over the same spread return −80.364%, a random NPC round trip that cannot lose returns +7.510%, and what we publish returns +864.433% in-sample. The top three say the harness is not inventing profit and that trading badly costs what it should. The fourth is the one you should discount, and it is drawn off the axis on the front page so it cannot pretend otherwise.