A common problem in climate studies is to decide whether a climate model is realistic. Such decisions are not straightforward because the time series are serially correlated and multivariate. Part II derived a test for deciding wether a simulation is statistically distinguishable from observations. However, the test itself provides no information about the nature of those differences. This paper develops a systematic and optimal approach to diagnosing differences between stochastic processes.