Fish population dynamics estimation is important for the sustainable use of fisheries resources. Conventional fisheries stock assessments often rely on age-structured population dynamics models; however, future projections can vary substantially depending on the assumed stock–recruitment relationship and age-specific biological parameters. In addition, nonlinear responses to fishing pressure and environmental changes are often not fully considered, which can lead to limited predictive accuracy and, in some cases, an underestimation of risk in future projections.In this study, we applied Gaussian Process Empirical Dynamic Modeling (GP-EDM), an equation-free nonlinear time-series analysis, to the chub mackerel (Scomber japonicus) in the northwestern Pacific to improve future projection accuracy and to understand greatly fluctuating dynamics. Using leave-future-out cross-validation, we found that GP-EDM predictions based solely on the time series of spawning stock biomass (SSB) outperformed conventional stock assessment approaches based on life-cycle models linking SSB and recruitment. Including sea surface temperature as a covariate further improved prediction skill and revealed the emergence of alternative stable states, where both high and low SSB regimes coexisted under warm conditions. Our results suggest that the recent rapid decline of the chub mackerel population was driven by increased fishing pressure, and that recovery is unlikely as long as elevated temperatures persist, even if fishing pressure is substantially reduced. These findings highlight that under global warming, small changes in fishing pressure may trigger sudden and irreversible stock collapses. The approach developed here contributes to forecasting and sustainable management of small pelagic fish exhibiting complex dynamics under climate change.