Reported freeze dates gave no clear mean forecast gain beyond time trends at 41 Finnish lake sites in 1991–2020

We tested whether reported permanent freeze dates improve forecasts of complete spring ice break-up at 41 Finnish lake observation sites. Each model used only earlier spring records for fitting. The test used 1,149 winters in 1991–2020. Adding the freeze date to a local time trend reduced mean site squared error from 51.94 to 47.45 days squared, a gain of 8.66%. A 95% range from common three-year blocks was −6.50% to +15.66%. The test did not establish a mean gain beyond time trends. Against a running historical mean, the gain was 23.57%, with a range of +4.28% to +30.32%. That model still had higher error than a time trend alone. The result shows that the reference model can change the conclusion. Rolling training windows, previous break-up dates and sample checks did not give a clear mean gain beyond time trends. The source flags can be assigned after the freeze event, and report times are absent. This is a retrospective test, not proof of a verified forecast lead. The package preserves raw Syke records, source hashes, a local plan, all forecasts, code, figures and offline instructions. A separate calculation checked all main forecasts. A full repeat matched 37 result and figure files exactly.

lake ice · Finland · seasonal forecast · ice break-up · freeze-up · time trend · forecast validation · open environmental data