This work examines how climate projections are affected by using different subsets of available climate models.

It has been shown previously that simply choosing models that perform well can actually degrade climate model ensemble estimates, since the models that perform well may share the same kinds of biases, and choosing a set of models that all have different kids of biases can give the best projections, when examined together as a whole.

Here we explore the relationship between choosing a climate model ensemble to maximise ensemble performance in different variables and its effect on estimates of prediction uncertainty.

We find that optimising a climate model ensemble to perform well in one variable not only degrades estimates in other variables, but results in uncertainty estimates that are too narrow. It is only when several variables are optimised independently, and the results from each optimisation are merged, that we can achieve both improved ensemble performance and improved uncertainty estimates.

Future Australian precipitation is used as a case study to illustrate the technique.