The El Niño-Southern Oscillation (ENSO) is the dominant mode of interannual climate variability in the tropical Pacific and exerts widespread impacts on the global climate system through atmospheric teleconnections.Although substantial progress has been achieved in ENSO prediction over recent decades,forecast skill assessments are still predominantly based on time-series metrics such as the anomaly correlation coefficient (ACC) and root-mean-square error (RMSE).While these metrics provide a general measure of forecast skill,they are insufficient for diagnosing forecast performance during critical stages of ENSO evolution,particularly the event peak,which largely determines the magnitude and spatial pattern of climate impacts.
Using real-time multi-model forecasts from the International Research Institute for Climate and Society (IRI) covering the period from 2002 to 2025,this study develops an event-oriented diagnostic framework to quantitatively evaluate ENSO forecast performance with a focus on peak characteristics.Two key attributes are examined:peak amplitude and peak timing,both defined based on the Niño 3.4 sea surface temperature anomaly.Rather than focusing on continuous forecast skill,the analysis emphasizes discrete ENSO events and evaluates how forecast errors evolve with increasing forecast lead time.
Results indicate that although the IRI forecast system maintains relatively high overall ENSO prediction skill up to approximately 8—9 months,substantial systematic biases emerge when individual event peaks are examined.For peak amplitude,forecasts exhibit a general tendency toward amplitude weakening with increasing lead time.This bias is strongly dependent on event intensity.Moderate and strong ENSO events are more likely to be underestimated,whereas weak events near the ENSO threshold tend to be overestimated.Such behavior suggests that prediction systems have difficulty sustaining realistic amplitudes for strong events while being overly sensitive to marginal signals associated with weak events.Clear differences are also found between model categories.Dynamical models show superior performance in reproducing the peak amplitude of moderate-to-strong events,reflecting the advantage of explicitly simulating coupled ocean-atmosphere processes.In contrast,statistical models tend to produce peak amplitude estimates closer to observations for weak events.
In terms of peak timing,forecasts systematically exhibit delayed peaks relative to observations,with timing errors increasing as forecast lead time increases.This lagged behavior is evident across all model categories and highlights persistent difficulties in accurately simulating the temporal evolution of ENSO events.Peak timing errors also display pronounced phase asymmetry.La Niña events tend to experience substantially larger delays than El Niño events,indicating lower predictability for cold-phase evolution.Comparisons between model types further reveal that statistical models exhibit significantly larger peak-timing errors than dynamical models for La Niña events,whereas the difference between the two model types is relatively small for El Niño events.
Overall,this study demonstrates that ENSO peak prediction skill in the IRI real-time forecast system is characterized by systematic amplitude weakening and delayed peak timing,with strong dependence on event intensity and phase.The results highlight the complementary strengths of dynamical and statistical models and underscore the importance of event-based diagnostics for understanding forecast uncertainty.These findings provide useful guidance for improving multi-model ensemble strategies and enhancing the reliability of ENSO peak predictions.