An update from prior August reduced squared oxygen forecast error by 8% against past means in 79 lake records, 2001–2017

We tested whether the last August oxygen value improved the next August forecast in open lake records. We used complete fixed depth bands at 50–80% of maximum depth. Training coverage selected 81 records in 1980–2000. Of these, 79 supplied 1,028 tests in 2001–2017. A coefficient fitted to forward training rows gave a forecast with 20.2% of the last value and 79.8% of the mean of all past values. Mean squared error fell from 1.663 to 1.530 (mg/L)², an 8.0% reduction. The nominal 95% error-ratio interval was 0.834–0.997 with region sampling and shared three-year blocks. It was 0.832–1.005 with five-year blocks. The gain over a mean of the last five observations was uncertain. Deeper bands and a narrower August window also gave lower error than historical means. These are retrospective tests in a selected and uneven network. They do not identify a physical cause of annual persistence. The package preserves the full source data, all forecasts, code, figures, checks, and offline repeat instructions.

dissolved oxygen · lake monitoring · annual forecasts · reference models · open data