Assessment of three temperature reconstruction methods in the virtual reality of a climate simulation

The performance of statistical climate reconstruction methods in the pre-instrumental period is uncertain, as they are calibrated in a short instrumental period but applied to much longer reconstructions time spans. Here, the virtual reality created by a climate simulation of the past millennium with the model ECHO-G is used as a test bed of three methods to reconstruct the annual Northern Hemisphere temperature. The methods are Composite plus Scaling, the inverse regression method of Mann et al. (Nature 392:779–787, 1998) and a direct principal-components regression method. The testing methodology is based on the construction of pseudo-proxies derived from the climate model output, the application of each of these methods to pseudo-proxy timeseries, and the comparison of their result with the simulated mean temperature. Different structures of the noise have been used to construct pseudo-proxies, ranging from the simulated grid-point precipitation. Also, one sparse and one denser pseudo-proxy network, co-located with two real networks, have been considered. All three methods underestimate the simulated variations of the Northern Hemisphere temperature, but the Composite plus Scaling method clearly displays a better performance and is robust against the different noise models and network size. The most relevant factor determining the skill of the reconstruction appears to be the network size, whereas the different noise models tend to yield similar results.