a Location of vents and seismic station at Whakaari volcano. b RSAM (Real-time Seismic Amplitude Measurement) tremor signal over the 9-year study period, with five eruptive periods indicated by darkened colored lines. The wide-shaded bar either side of each eruption demarcates the testing interval during cross-validation. c Tremor time series (RSAM, MF, HF, DSAR) in the 4 days preceding the December 2019 eruption, indicated by the vertical red bar (UTC time). The relative positions of two adjacent windows (black arrows at bottom of c) and their associated look-forward periods are indicated below. The eruption falls outside the look-forward period (gray arrow) of window i, but inside the look-forward of i + 1, and are thus labeled 0 and 1, respectively. The ensemble mean of an eruption forecast model that accepts the windowed tremor data is given in the bottom frame of (c). d–g Frequency distribution of exemplary feature values, with values prior to the five eruptions (colored markers, corresponding to colored lines in (b) explicitly plotted. d–f show statistically significant features for eruption forecasting. Larger markers denote a window nearer to the eruption. Mann–Whitney U p values are quoted and indicate feature significance.