Non-zero model error indicates that our model is missing explanatory features. In practice, we don’t expect to get rid of all model error—there will be some error in the forecast from unavoidable natural variation. Natural variation should reflect all the stuff we will probably never capture with our model, like measurement error, unpredictable external market forces, and so on. The distribution of error should be close to normal and, ideally, have a small mean. We get evidence that an important explanatory variable is missing from the model when we find that the model error doesn’t look like simple natural variation—if the distribution of errors skews one way or another, there are more outliers than expected, or if the mean is unpleasantly large. When this happens we should try to identify and correct any missing or incorrect model features.
It’s an interesting article, especially the bit about cross-validation, which is a perfectly acceptable technique in non-time series models.