Increasing Foresight and Forecast Quality with Skillful Low-Cost Empirical Models

Thursday, 18 December 2014
Hailiang Du1, Leonard A Smith1,2, Emma Suckling3,4 and Erica L Thompson4, (1)University of Chicago, Chicago, IL, United States, (2)London School of Economics, London, WC2A, United Kingdom, (3)University of Reading, Reading, United Kingdom, (4)London School of Economics, London, United Kingdom
Simulation models are widely employed to make probability forecasts on seasonal to annual time-scales and increasingly on decadal scales. While simulation models based on physical principles are often expected, in principle, to outperform purely empirical models, that claim must be established empirically for any given generation of models; direct comparison of the forecast skill of simulation models and empirical models provides information on progress toward that goal which is not available in model-model intercomparisons. More importantly, the blending of forecasts from both sources can lead to better operational forecasts. Direct comparison can also reveal the space and time scales on which simulation models exploit their physical basis effectively, perhaps indicating the origins of their weaknesses. The skill of seasonal and decadal probabilistic hindcasts for global and regional mean temperatures from the ENSEMBLES project and CMIP5 are interpreted in this context. Physically inspired empirical models are shown to display probabilistic skill comparable to that of today’s state-of-the-art simulation models as well as to that of the multi-model ensemble. The inclusion of empirical models (blending) with simulation models is shown to significantly improve forecasts. Inasmuch as the cost of building or running empirical models is negligible comparing to large simulation models, it is suggested that the direct comparison of simulation models with empirical models become a regular component of large model forecast evaluations, that rank order evaluations include empirical models whenever the timescales allow, and that blending simulation models with empirical models becomes a regular component of seasonal and decadal forecasting.