Global mean surface air temperature (T) variability on subdecadal timescales can be of substantial magnitude relative to the long-term global warming signal and such variability has been associated with considerable environmental and societal impacts. Therefore, probabilistic foreknowledge of short-term T evolution may be of value for anticipating and mitigating some course-resolution climate-related risks. Here we present an empirically-based methodology that utilizes global spatial patterns of annual surface air temperature to predict subsequent annual T anomalies via Partial Least Squares Regression. The method’s skill is achieved via information on the state of long-term global warming as well as the state and recent evolution of the El Niño-Southern Oscillation and the Interdecadal Pacific Oscillation. We test the out-of-sample skill of the methodology using a “forecast mode” where statistical predictions are made precisely as they would have been if the procedure had been operationalized starting in the year 2000. The forecast errors for lead times of 1 to 4 years are smaller than naïve benchmarks using persistence and perform favorably relative to most dynamical Global Climate Models retrospectively initialized to the observed state of the climate system. Thus, this method can used as a computationally-efficient benchmark for dynamical model forecast systems.