Fig. 3: Habitat forecasts capture low-frequency variability.

The forecast skill of multi-annual averages of habitat area (panels a-c), as characterised by the Pearson correlation coefficient (r), is shown for the grand-ensemble and persistence forecasts. In addition to the single-year values also plotted in Fig. 2 (solid lines), the skill of multi-year averages (3, 5, and 9 year centred means) are also shown (broken lines with symbols). Lead-time is defined as the length of time from the issuing of the forecast (1 January) to the middle of the running mean window. Multiyear forecasts are significantly better than multiyear persistence for all lead times (p < 0.01, one-tailed test, as estimated by bootstrapping). Time-series of habitat metrics (panels d-f) show habitat estimates based on observations (triangles connected by dotted line) with their three-year running means (solid black lines). Habitat metrics forecast by the grand-ensemble (solid red line) with a 5-year lead time are shown with the corresponding 90% range of realizations (shaded area). Time series are shown for the full range of years used to estimate the forecast performance (i.e., 1961–2018 for mackerel and bluefin tuna, 1985–2018 for blue whiting). Panels (a) and (d) show results for the area of mackerel habitat around south Greenland, panels (b) and (e) bluefin tuna habitat south of Iceland, and (c) and (f) blue whiting spawning habitat west of Great Britain and Ireland. Source data are provided as a Source Data file.