Skillful Climate Forecasts of the Tropical Indo-Pacific using Model-based Analogs

In this talk, we explore the idea that for forecast lead times sufficiently long that the forecast model has largely drifted to its own climatology, model initialization should be directly in the model phase space and not in the phase space of nature. There are many sophisticated approaches that might be taken to address this problem. We will instead take the deliberately simple-minded approach of constructing an ensemble of the nearest model states to each observed initial state, where the model states are taken from a long control run of a CGCM that corresponds to the forecast CGCM, and then using the ensemble mean of the evolution of those model states as our forecast. That is, we will construct analog models of CGCMs, where the data “library” is not observations but rather the free-running model itself.

This “model-analog” technique is applied to half (four) of the CGCMs comprising the North American Multimodel Ensemble (NMME): NCAR-CCSM4, NCAR-CESM1, GFDL CM2.1, and GFDL CM2.5 (FLOR). The length of the corresponding control model runs ranges from 700-4000 years, allowing a test of the sensitivity of the technique to both data availability and ensemble size. Model-analogs are determined only over the tropical IndoPacific domain, using monthly SST and SSH anomalies. In a perfect model framework, where the first 200 years of each run are withheld for verification purposes while the remaining model output is used for the (seasonally-varying) analog data library, we obtain tropical Pacific SST skill found by other perfect model studies, with a Nino3.4 correlation above 0.5 for forecast leads up to 24 months. Interestingly, this perfect model predictability appears to be largely linear.

We then apply the technique to make hindcasts for the years 1982-2009 for forecast leads of 1-12 months, initializing with observed SST and SSH anomalies over that period. The model-analog hindcast skill is found to be comparable to (and often slightly better than) the corresponding CGCM hindcast skill throughout the tropical Pacific. This suggests the possibility that any CGCM with a sufficiently long control run may be used to produce skillful forecasts of monthly tropical IndoPacific SST anomalies, without the need for additional development of its data assimilation system.