Probe lab
Does Timestral-5M's uncertainty mean what it says? Three checks on any series. Everything runs in this browser; your file never leaves it.
1 · Coverage audit
Forecast from many past points of the series and check the answer against what happened. An honest 80% band contains the truth 80% of the time.
The series and the last backtest forecast
Reliability: stated vs actual coverage
On the grey diagonal the bands are honest. Above it they are too wide, below it too narrow.
Where the truth lands inside the forecast (PIT)
Flat at 10% per bar is honest. A hump means too wide, a U too narrow, a slope means bias.
80% coverage at each cut-off
Does the band stay honest across the series? A run of low points marks a change the model has not caught up with.
The same, as a table
Coverage below 50% at a cut-off is a band exit worth flagging.
80% coverage by step ahead
Does the band stay honest further out?
80% band width by step ahead
In units of the series' typical one-step change.
Numbers
2 · Blind spot
Hide a stretch of the series. Forecast it forward from the left and backward from the right (reverse the series, forecast, reverse back). Where the two disagree, neither knows, whatever its band says.
Does disagreement predict the error?
Press "Scan 30 gaps" to test it across the series.
3 · Horizon fade
Forecast past the end of the series. The band should widen with distance. The lab stops at 1,024 steps: beyond one decode block Timestral's band collapses (80% coverage fell from 82% to 5% in our test).
The forecast
80% band width by step ahead
In units of the series' typical one-step change.