The 24-year retrospective forecast data set from the NCEP Climate Forecast System (CFS) is analyzed to study its idealized predictability of precipitation and temperature under its current configuration. The analysis approach assumes the forecasting model and system to be predicted share exactly the same physics so that the idealized predictability is calculated and serves as the upper limit of the predictive skill in practical forecasts. The analysis shows that CFS is not capable of predicting itself over much of the mid-latitudes land areas for precipitation and temperature anomalies having small temporal (monthly) and spatial (2.5¿ ¿ 2.5¿ grid) scales at lead-times longer than a month. Anomalies become more predictable with the increase in temporal and spatial scales and with the decrease in lead-times, as illustrated with results from the central US region. The results imply that additional care should be taken when using climate model seasonal forecast products.