AbstractThe asymptotic predictability of global land surface precipitation is estimated empirically at the seasonal time scale with lead times from 0 to 12 months. Predictability is defined as the unbiased estimate of predictive skill using a given model structure assuming that all relevant predictors are included, thus representing an upper bound to the predictive skill for seasonal forecasting applications. To estimate predictability, a simple linear regression model is formulated based on the assumption that land surface precipitation variability can be divided into a component forced by low-frequency variability in the global sea surface temperature anomaly (SSTA) field and that can theoretically be predicted one or more seasons into the future, and a “weather noise” component that originates from nonlinear dynamical instabilities in the atmosphere and is not predictable beyond ~10 days.Asymptotic predictability of global precipitation was found to be 14.7% of total precipitation variance using 1900–2007 data, with only minor increases in predictability using shorter and presumably less error-prone records. This estimate was derived based on concurrent SSTA–precipitation relationships and therefore constitutes the maximum skill achievable assuming perfect forecasts of the evolution of the SSTA field. Imparting lags on the SSTA–precipitation relationship, the 3-, 6-, 9-, and 12-month predictability of global precipitation was estimated to be 7.3%, 5.4%, 4.2%, and 3.7%, respectively, demonstrating the comparative gains that can be achieved by developing improved SSTA forecasts compared to developing improved SSTA–precipitation relationships. Finally, the actual average cross-validated predictive skill was found to be 2.1% of the total precipitation variance using the full 1900–2007 dataset and was dominated by the El Niño–Southern Oscillation (ENSO) phenomenon. This indicates that there is still significant potential for increases in predictive skill through improved parameter estimates, the use of longer and/or more reliable datasets, and the use of larger spatial fields to substitute for limited temporal records.