Fig. 1: Variability and predictability of length of day.

a, Variations in the length of day (LOD) showing the prominent interannual variability of around 0.5 × 10⁻³ s in observations (black) and the first year of ensemble mean model predictions starting in November each year (red). b, Correlation of predicted seasonal length-of-day anomalies in the ensemble mean with length-of-day anomalies from single model ensemble members (black), with radio telescope observations of Earth’s rotation (blue) and with atmospheric reanalysis (red). The perfect model predictability (black) is smoother than the prediction skill against observations (red, blue) due to averaging of the correlations with each ensemble member in the model case. Note the non-monotonic variation with lead time and the peaks at leads of 3 and 15 months in winter. Statistical significance at the 95% level according to a one-sided t test for positive correlations is shown by the dotted line.