Errors in retrospective ensemble forecasts of Nino 3.4 index from the NCEP Climate Forecast System (CFS) are analyzed. The forecast data used comprise 23 years of 15-members ensembles issued once a month. These ensembles are formed by collecting forecasts initialized within the previous month (referred to as “lagged” ensemble generation scheme). By design, the initial error of the ensemble mean in a given time is usually larger than if all members were initialized simultaneously. The study quantifies the skill loss in the lagged approach by partitioning forecast errors into initial-related errors and model-related errors. Results indicate a skill loss of about 10 days; about 5 days if ensembles are generated using the latest 8 members instead. Implications of these results when designing a sub-seasonal prediction system will be discussed.