AbstractThis letter explored the influence of atmospheric convection scheme on El Niño–Southern Oscillation (ENSO) predictions using a set of hindcast experiments. Specifically, a low‐resolution version of the Climate Forecast System version 2 is used for 12 month hindcasts starting from each April during 1982–2011. The hindcast experiments are repeated with three atmospheric convection schemes. All three hindcasts apply the identical initialization with ocean initial conditions taken from the European Centre for Medium‐Range Weather Forecasts and atmosphere/land initial states from the National Centers for Environmental Prediction. Assessments indicate a substantial sensitivity of the sea surface temperature prediction skill to the different convection schemes, particularly over the eastern tropical Pacific. For the Niño 3.4 index, the anomaly correlation skill can differ by 0.1–0.2 at lead times longer than 2 months. Long‐term simulations are further conducted with the three convection schemes to understand the differences in prediction skill. By conducting heat budget analyses for the mixed‐layer temperature anomalies, it is suggested that the convection scheme having the highest skill simulates stronger and more realistic coupled feedbacks related to ENSO. Particularly, the strength of the Ekman pumping feedback is better represented, which is traced to more realistic simulation of surface wind stress. Our results imply that improving the mean state simulations in coupled (ocean‐atmosphere) general circulation model (e.g., ameliorating the Intertropical Convergence Zone simulation) might further improve our ENSO prediction capability.