Mean absolute error grows with lead time; the useful question is how fast. Every cell is printed with the number of pairs behind it, because a tidy value drawn from forty samples is not a measurement.

Temperature (°F)

6-hour precipitation (in)

Mean absolute error says how far a forecast lands from the report; the signed mean says which side. A bias is not an error a reader has to accept — it is one they can correct for. Red runs warm, blue runs cool.

Showing the 18 most-biased of 88 stations with at least 200 scored pairs.

The largest temperature residuals in the file, one row per event. A bad forecast is re-issued every cycle until its valid time passes, so the raw list repeats the same miss a dozen times; the cycles column says how many runs carried it, which is the difference between one bad run and a model that stayed wrong for two days. They are printed for the same reason the rest of the file is.

ASOS reports at :53 and the forecast is stamped on the hour, so a forecast is paired with the nearest observation within 30 minutes and otherwise left unscored.

The ledger begins when the residuals began being written down. The upstream history store keeps a rolling forty cycles, so nothing before that date can be recovered at any price — the table deepens from here rather than reaching backwards.

A residual that cannot be believed is not a miss. 4 rows were excluded here as a data defect rather than a forecast error — a forecast outside the range of anything its own station has ever reported is a broken number, and publishing it as the month's worst miss would blame a forecaster for a decode bug.

Machine edition: /api/v1/ledger.json. Built 2026-10-31T21:25:59Z.