Enhanced multi-year predictability after El Niño and La Niña events

Abstract

Several aspects of regional climate including near-surface temperature and precipitation are predictable on interannual to decadal time scales. Despite indications that some climate states may provide higher predictability than others, previous studies analysing decadal predictions typically sample a variety of initial conditions. Here we assess multi-year predictability conditional on the phase of the El Niño-Southern Oscillation (ENSO) at the time of prediction initialisation. We find that predictions starting with El Niño or La Niña conditions exhibit higher skill in predicting near-surface air temperature and precipitation multiple years in advance, compared to predictions initialised from neutral ENSO conditions. This holds true in idealised prediction experiments with the Community Climate System Model Version 4 and to a lesser extent also real-world predictions using the Community Earth System Model and a multi-model ensemble of hindcasts contributed to the Coupled Model Intercomparison Project Phase 6 Decadal Climate Prediction Project. This enhanced predictability following ENSO events is related to phase transitions as part of the ENSO cycle, and related global teleconnections. Our results indicate that certain initial states provide increased predictability, revealing windows of opportunity for more skillful multi-year predictions.

Fig. 1. Potential skill dependence on initial El Niño-Southern Oscillation (ENSO) state.

The mean squared skill score (MSSS) for the near-surface temperature in the perfect-model predictions for decadal simulations initialised in El Niño (EN; a, b, c), La Niña (LN; d, e, f), and the skill difference between the different groups: El Niño – neutral (g, h, i) and La Niña – neutral (j, k, l). Rows correspond to forecast year 1 (a, d, g, j), the average of forecast years 2–3 (b, e, h, k), and the average of forecast years 4–6 (c, f, i, l). The plus sign stippling indicates grid cells where the skill score or skill differences are significant at the false discovery rate (FDR) being 0.2 (when adjusting for the false discovery rate, see Methods for details). The square symbols indicate grid cells where the skill score or skill differences are significant at the 90% level (outside the 5%–95% confidence interval based on local individual testing with 1000 bootstrap realisations).

Fig. 2. Areas of skill.

The percentage of global areas (10° × 10° grid cells weighted by cos(latitude)) where the perfect-model predictions exhibit significant skill (the false discovery rate (FDR) is 0.2 when adjusting for the false discovery rate as indicated by the ‘+’ sign stippling in Fig. 1) for predictions initialised in El Niño (red), La Niña (blue), neutral (black), and persistent neutral (yellow) conditions. The shading around each line represents the 5%–95% confidence interval based on 1000 bootstrap realisations (with replacement, see Methods).

Fig. 3. Agreement of local temperature anomalies.

Composite maps of normalised temperature anomalies (after detrending; refer to Methods for details of normalisation and detrending) from perfect-model decadal simulations started from El Niño (EN; a, b, c), La Niña (LN; d, e, f) and neutral (NEU; g, h, i) conditions for forecast years 1 (a, d, g), 2–3 (b, e, h) and 4–6 (c, f, i). The pattern correlations (weighted by grid cell areas) of the initialised (r_ini) and uninitialized (r_noIni) perfect-model predictions compared to the reference simulation are shown above each panel. Subplot titles are in black when r_ini is significantly greater than r_noIni (exceding the one-sided 95% confidence interval, based on 1000 bootstrap realisations), whereas grey subplot titles show non-significant difference between r_ini and r_noIni.

Fig. 4. Skill dependence on initial El Niño-Southern Oscillation (ENSO) status in decadal hindcasts of the real-world climate.

The mean squared skill score (MSSS) of the near-surface temperature for the real-world hindcasts provided by the Community Earth System Model decadal prediction large ensemble (CESM-DPLE) simulations initialised in El Niño (EN; a–c), La Niña (LN; d–f), and the skill difference between the different groups: El Niño – neutral (g–i) and La Niña – neutral (j–l). Rows correspond to forecast year 1 (a, d, g, j), the average of forecast years 2–3 (b, e, h, k), and the average of forecast years 4–6 (c, f, i, l). The plus sign stippling indicates grid cells where the skill score or skill differences are significant at the false discovery rate (FDR) being 0.2 (when adjusting for the false discovery rate, see Methods for details). The square symbols indicate grid cells where the skill score or skill differences are significant at the 90% level (outside the 5%–95% confidence interval based on local individual testing with 1000 bootstrap realisations). HardCRUT4-median is used as the observation ref. .