Published and updated as calibrations are recomputed — including when they get worse. A willingness to show a degradation is what makes the good numbers believable.
Drift between a signal bar's close and a fill five minutes later, measured on 1-minute bars. Positive means the price moved against the trade before it could be filled.
| Symbol | Strategy family | Mean | p10 | p90 | Signals | Window |
|---|---|---|---|---|---|---|
| SOLUSDT | sma_cross_1h | +1.1 | -15.9 | +14.4 | 92 | 45d |
| BTCUSDT | sma_cross_1h | +0.1 | -14.7 | +9.8 | 93 | 45d |
| ETHUSDT | sma_cross_1h | -0.7 | -14.6 | +14.1 | 106 | 45d |
The mean of this cost is close to zero on every symbol we have measured. The band is ±15 basis points. Reporting the mean alone would say "this cost does not exist"; reporting the band says "on any individual trade it can go either way by more than the commission", which is the true and more useful statement.
The strongest accuracy statement available compares what we predicted against what a venue actually charged and what the market actually did, over real fills. Errors below are predicted minus actual: positive means we overstated the cost, which is the direction that matters, because overstating is what talks someone out of a strategy that would have worked.
No comparison against live fills has been run yet. Adverse selection is a model.
Until this table has a row with at least 30 comparable round trips for a symbol, every report for that symbol repeats that sentence on its signal-to-fill drift line. We would rather say it on every page than let a fitted number pass for a measured one.
Every next-week and next-month range a What-If user follows is compared with the outcome on the day its horizon ends, and counted here whichever way it went. A range whose calibration is not published is an opinion.
The first month's calibration is thin: one scenario checked once proves little, and a band that is right eight times in ten will look wrong on any single check. The count is what grows; the claim does not.