Why Do Forecasts for “Near Normal” Often Fail?

It has been observed by many that skill of categorical forecasts, when decomposed into the contributions from each category separately, tends to be low, if not absent or negative, in the “near normal” (N) category. We have witnessed many discussions as to why it is so difficult to forecast near normal weather, without a satisfactory explanation ever having reached the literature. After presenting some fresh examples, we try to explain this remarkable fact from a number of statistical considerations and from the various definitions of skill. This involves definitions of rms error and skill that are specific for a given anomaly amplitude. There is low skill in the N-class of a 3-category forecast system because a) our forecast methods tend to have an rms error that depends little on forecast amplitude, while the width of the categories for predictands with a near Gaussian distribution is very narrow near the center, and b) it is easier, for the verifying observation, to ‘escape’ from the closed N-class (2-sided escape chance) than from the open ended outer classes. At a different level of explanation, there is lack of skill near the mean because in the definition of skill we compare the method in need of verification to random forecasts as the reference. The latter happens to perform, in the rms sense, best near the mean. Lack of skill near the mean is not restricted to categorical forecasts or to any specific lead time. Rather than recommending a solution, we caution against the over-interpretation of the notion of skill-by-class. It appears that low skill near the mean is largely a matter of definition and may therefore not require a physical-dynamical explanation. We note that the whole problem is gone when one replaces the random reference forecast by persistence. We finally note that low skill near the mean has had an element of applying the notion forecasting forecast skill in practice long before it was deduced that we were making a forecast of that skill. We show analytically that as long as the forecast anomaly amplitude is small relative to the forecast rms error, one has to expect the anomaly correlation to increase linearly with forecast magnitude. This has been found empirically by Tracton et al. (1989).