The medium‐range ensemble prediction system (EPS) at the Japan Meteorological Agency (JMA), like other operational EPSs, suffers from overconfident probabilistic forecasts in the lower troposphere and over the oceans in particular. To alleviate this issue, a new scheme to perturb the sea surface temperature (SST) is proposed and tested. The proposed method statistically generates SST perturbations that account for time‐varying errors of the SST that is prescribed to the control member. Despite its simplicity and purely statistical nature, the new method is able to produce growing perturbations whose spatial patterns appear to reflect ocean dynamics. Their growth with forecast lead‐time is completely flow‐independent but is statistically consistent with the past records of the error growth of the prescribed SST. Perturbing the SST with the proposed scheme is shown to improve the probabilistic forecast skills of JMA's medium‐range EPS. The improvement is statistically significant over the Tropics throughout the entire troposphere but the impact on the extratropical forecast is found to be neutral. The improvement is largely attributed to enhanced reliability which in turn is achieved through an increase in the ensemble spread. Perturbing SST incurs no harmful impact on the quality of ensemble mean forecast.