No proprietary model, no black box. The method comes down to six steps, and every one of them is built so that it can be proven wrong.
Closed positions, rebuilt from public records. No third-party feed, no user-reported figures.
A minimum number of closed positions is required before any ranking. An operator with three trades is not comparable to one with three thousand.
An operator's benchmark is computed with that operator left out. Without this precaution, everyone beats their own average.
The rank is established on one period, then checked on the following ones. An edge that does not survive a change of period is not an edge.
If the method produces the same result on shuffled data, it measures nothing. This control runs on every rebuild.
Never updated by appending. The database starts from zero and sweeps the complete history, including the part that predates the tool itself.
If any one of them fails, that night's database is not published and the previous one stays in service. A site that displays a wrong figure is worse than a site that displays nothing.
| Check | What it prevents |
|---|---|
| Operator count | A database truncated by a read incident |
| Number of bands | A broken split that would make ranks incomparable |
| No operator without a position | Empty rows counted as operators |
| Total losses present | The exact flaw that had skewed our first measurement |
| Win rate increasing across bands | A reversed or random ranking |
| Total losses decreasing across bands | The same check, taken from the other end |
| Size of the benchmark grid | Benchmarks computed on too few cases |
On the amount-weighted indicator, the second and third bands swap places. The inversion comes from a handful of very high-volume operators that weigh heavily in a weighted average. We could have dropped the indicator or smoothed the extremes. We chose to leave it visible and explain it: the two unweighted indicators rise steadily, without exception.