Empirical models of climate prediction on subseasonal to seasonal time scales

We discuss the use of a seamless empirical dynamical modeling approach to construct a state-of-the-art benchmark probabilistic forecast system for forecast lead times ranging from weeks to decades. The model used, a linear inverse model (LIM) derived from observed simultaneous and time-lag correlation statistics of oceanic and atmospheric variables, can also be used to make forecasts whose skill is competitive with current coupled global forecast GCMs. The geographical and temporal variations of forecast skill are also generally similar for the LIM and CGCMs. This makes the much simpler LIM an attractive tool for assessing and diagnosing climate predictability, including determining factors that contribute to climate prediction skill as well as those that limit it, and also for diagnosing the predictability of climate modes such as the MJO, ENSO, PDO, and AMO. The LIM formalism also allows determination of which climate regimes have particularly high or low predictability, and how this affects the predictability of different system variables. This suggests that an important “best practices” aspect of climate forecasts on all time scales should be the issuance of both a forecast and a “forecast of forecast skill”, of which the LIM would be an important component. Additionally, the LIM shows forecast skill that is greater than CGCMs at subseasonal forecast leads in some areas and for some seasons, which suggests it may be used for improving forecasting now.