Abstract

Chile Ni & ntilde;o is a seasonally modulated coastal warm mode that affects northern-central Chile and challenges early-warning systems because events are short-lived and geographically confined. Here we assess the predictability of the Chile Ni & ntilde;o Index (CNI) using a hierarchy of data-driven inverse models, ranging from a baseline Linear Inverse Model to Seasonally Varying and nonlinear extensions, and a lightweight hybrid scheme in which a Long Short-Term Memory network is trained to correct systematic forecast residuals. We also evaluate Gross Coastal Warming (GCW), a complementary metric of total coastal SST anomalies over the Chile Ni & ntilde;o region, to distinguish the predictability of total coastal warming from that of the residual coastal mode represented by the CNI. We first show that the Linear Inverse-Model framework reproduces key characteristics of coastal sea-surface temperature variability associated with Chile Ni & ntilde;o, supporting its suitability for predictability assessment. Deterministic and probabilistic verification identifies a clear window of forecast skill for austral-autumn initializations (April-May) at short lead times (1-3 months), together with a secondary but shorter-lived enhancement for early-winter initializations (June-July) that is largely confined to 1-2-month leads. Within these windows, the Extended CS-LIM yields the strongest correlations, the lowest normalized errors, and the most reliable probabilities of warm and cold coastal conditions. The hybrid correction does not substantially increase the short-lead forecast skill maximum, but it provides consistent relative improvements at intermediate lead times (4-8 months), particularly when warm eastern Pacific events precede the target period. The GCW experiment shows higher forecast skill than the CNI and remains skillful through lead 4, indicating that part of coastal-warming predictability is associated with the basin-scale component retained in total coastal SST anomalies. A case study of the 2017 event, one of the strongest on record, illustrates the added value of the hybrid correction: the inverse model captures the onset of coastal warming but underestimates its magnitude and delays the phase transition, whereas the hybrid scheme better maintains warm persistence and improves the timing of the reversal. Overall, our results indicate that a seasonally explicit stochastic Inverse-Model framework, augmented with a parsimonious data-driven correction, provides a physically interpretable and computationally efficient basis for diagnosing and forecasting coastal warming in the southeast Pacific.