Train on history. Prove it on unseen weather.
Learn a signed temperature correction from earlier rows, then judge it only on a chronological holdout. No live forecast, market, trade, or PnL data is implied.
Training rows from start
Prediction before reveal
Bundled sample ready: 5 training rows and 3 unseen rows.
Correction = mean(actual − forecast) on training rows. Holdout rows are never used to fit it.
Learned correction+3.00°F
Raw holdout MAE2.00°F
Calibrated MAE1.00°F
MAE improvement50.0%
KEEP: correction improves unseen MAE.Prediction confirmed.
DateForecastActualCorrectedRaw errorNew error
Why: the training set repeatedly undershot by 3°F, so the model adds 3°F before evaluation.
Holdout errors varied from training. The correction helped, but did not make every forecast exact.