Dominant Initial Uncertainty Sources for El Niño Forecasting

El Niño prediction uncertainty is highly sensitive to initial-condition uncertainties. The present study explores the sources and dynamics of initial uncertainty using the Coupled Conditional Nonlinear Optimal Perturbation (C‑CNOP) method. By imposing physically consistent and rapidly growing coupled initial perturbations, a series of ensemble forecast experiments were conducted for El Niño events from 1982 to 2023, with initializations in different seasons. The resulting ensembles demonstrate high reliability for predictions initialized in January, April, and July, effectively characterizing prediction uncertainty. Conversely, October-initialized predictions show persistent under-dispersion, as perturbation growth is suppressed by an overly stable model background state, indicated by a low Bjerknes stability index. Building on the reliable framework, key sensitive regions were identified across the Pacific, Indian, and Atlantic Oceans, where initial uncertainties significantly contribute to prediction uncertainties in the tropical central‑eastern Pacific, with patterns that vary seasonally. Beyond reaffirming the significant impact of extratropical North Pacific initial uncertainties, the results also highlight the role of mid-latitude South Pacific regions. Cross‑basin remote effects are also identified. Specifically, interactions between the tropical Atlantic and Pacific are confirmed for January and July initializations, alongside a reaffirmed influence from the tropical Indian Ocean. Statistical evidence also suggests potential pathways originating from the mid-latitude South Atlantic in January-initialized predictions and the subtropical South Indian Ocean in April-initialized predictions. Validation experiments demonstrate that reducing initial errors in these identified regions enhances prediction performance and reduces overall prediction uncertainty. Moreover, utilizing perturbation information from these regions to select ensemble members improves both deterministic and probabilistic prediction performance. These findings clarify the initial sources of El Niño prediction uncertainty and provide a practical foundation for optimizing targeted observation strategies.