Siegert & Stephenson "Forecast Recalibration and Multimodel Combination"

Siegert, S., & Stephenson, D. B. (2019). Forecast Recalibration and Multimodel Combination. In Sub-Seasonal to Seasonal Prediction (pp. 321-336). Elsevier.

Abstract

Since numerical models of the climate system are only imperfect approximations of the real world, their forecasts often do not agree with real-world observations. This discrepancy often necessitates a statistical postprocessing step, called forecast recalibration, to better fit the climate model forecasts to future observations. In this chapter, we discuss forecast recalibration from the point of view of statistical modeling. We consider numerical model forecasts as explanatory variables in a statistical model of the real world. We focus on well known regression methods, particularly model output statistics and nonhomogeneous Gaussian regression, for their simplicity, interpretability, and good performance to correct common forecast errors. In situations where more than one climate model forecast is available for the same prediction target, it is necessary to produce an optimal combination of multiple forecasts into a single forecast. We show that forecast recalibration and forecast combination are closely related, and discuss a Bayesian regression method to combine multiple forecasts. The method can assign higher combination weights to models that better correlate with the real world, while avoiding overfitting by controlling the overall variability of the combination weights. All methods are illustrated on seasonal temperature forecasts over the NiƱo-3.4 region.