In the current implementation, however, the focus has been fully placed on model improvements, whereas the forecast error representation remained the same. Here, we will diagnose several important issues related to the error sampling in the probabilistic forecasts systems that will need to be addressed in the future. First, the combined effect of the increase in horizontal resolution, introduction of a new modernized physics package and initial condition recentering leads to a major reduction of forecast error in the new REPS, especially in the lower troposphere. As a consequence, the system now tends to be over-dispersive. There are also more fundamental error-spread discrepancies that are characteristic of all versions of the REPS. For example, REPS tends to be initially under-dispersive at sub-synoptic scales below 1200 km, which is compensated by being over-dispersive at larger scales. After 24-48 hours of integration, the error-to-spread ratio becomes more uniform across the scales. Furthermore, the predictive skill of the ensemble spread in representing the time-varying model error is practically non-existent in the early stages of integration. The spread-error temporal correlation at the local scale is vanishingly small at the initial stages of the forecast. It increases somewhat during the first 24-48 hours but remains relatively low at all lead times, which is also the case in the Global Ensemble Prediction System (GEPS) from which the initial REPS perturbations are obtained.
These findings imply the need for (1) readjustment of the amplitudes of initial perturbations and (2) rethinking the stochastic physics perturbations in order to improve the model error representation in the probabilistic forecasts.