Impact of systematic gcm errors on prediction skill as estimated by linear inverse modeling

Abstract

Statistical prediction of tropical sea surface temperatures (SSTs) is performed using linear inverse models (LIMs) that are constructed from both observations and general circulation model (GCM) output of SST. The goals are to establish a baseline for tropical SST predictions, to examine the extent to which the skill of a GCM-derived LIM is indicative of that GCM's skill in forecast mode, and to examine the linkages between mean state bias and prediction skill. The observation-derived LIM is more skillful than a simple persistence forecasts in most regions. Its skill also compares well with some GCM forecasts except in the equatorial Pacific, where the GCMs are superior. The observation-derived LIM is matched or even outperformed by the GCM-derived LIMs, which may be related to the longer data record available for GCMs. The GCM-derived LIMs provide a fairly good measure for the skill achieved by their parent GCMs in forecast mode. In some cases, the skill of the LIM is actually superior to that of its parent GCM, indicating that the GCM predictions may suffer from initialization problems. A weak-to-moderate relation exists between model mean state error and prediction skill in some regions. An example is the eastern equatorial Atlantic, where an erroneously deep thermocline reduces SST variability, which in turn affects prediction skill. Another example is the equatorial Pacific, where skill appears to be linked to cold SST biases in the western tropical Pacific, which may reduce the strength of air-sea coupling.