Statistical approximation of high-dimensional climate models

We propose a general emulation method for constructing low-dimensional approximations of complex dynamic climate models. Our method uses artificially designed uncorrelated CO2 emissions scenarios, which are much better suited for the construction of an emulator than are conventional emissions scenarios. We apply our method to the climate model MAGICC to approximate the impact of emissions on global temperature. Comparing the temperature forecasts of MAGICC and our emulator, we show that the average relative out-of-sample forecast errors in the low-dimensional emulation models are below 2%. Our emulator offers an avenue to merge modern macroeconomic models with complex dynamic climate models.