Built on the SatNOGS open network
Before a satellite clears the horizon, its geometry is already known. This is how good the observation will be.
Loading the forecast model…
The science underneath
Not correlations from a black box. On 7,499 waterfalls across 402 stations, a case-control causal study estimated the odds ratio of each pre-registered factor, with partial-pooling for station and satellite identity and a battery of refuters. These are the effects the forecast is built on.
An honest instrument
An accurate-looking model is easy. An honest one is the point. This forecast is deliberately weaker than the numbers a naïve model would show — because most of that strength was an illusion.
Measured across unseen stations (grouped cross-validation), the forecast scores AUC 0.806 ± 0.02. The tempting 0.88–0.95 came from letting the model memorise station coordinates and collection dates — skill that vanishes for a new operator. We report the number that survives.
In a naïve model the collection month alone carried 37% of the predictive power, and dates together 51% — pure artifact of case-control sampling. A future pass has no "collection month", so we removed the whole family: the leaked AUC dropped from 0.88 to 0.66, and the forecast was rebuilt from geometry and RF alone.
The study found peak elevation and duration improve quality (OR 1.84 and 1.58). We hard-wired that direction into the model as a monotonic constraint — so it can never tell you a higher pass is worse. Science as a guardrail, at no cost to accuracy.
Time of day looked significant (OR 0.72), but a three-way placebo refuter still found an "effect" after we shuffled the treatment. It was a structural artifact, not a cause. So the forecast ignores the hour entirely. No evidence, no feature.