The field of forecasting has produced excellent work over the years. It has created models of uncertainty that can be assessed with remarkable precision and incorporated into all kinds of analyses to produce the best possible decisions. However, these models are not set up to deal with unexpected events. And the problem is, there are more unexpected events than people anticipate or acknowledge.For instance, look at oil prices. In 2008, the soaring price of oil reminded the world that inflation exists -- something that many Western economists seemed to forget. Consider the graph displaying daily changes in oil prices over a two-decade period. On first impression, the graph has a nice, symmetrical shape, with roughly the same pattern as many other daily series of data in economics and business. But it’s not the same shape as the normal distribution shown by the smooth pattern in the background. In particular, there are more extreme negative daily changes in the oil price -- both up and down -- than you would expect if the values were normally distributed. For example, between June 20 and October 11, 1990, oil prices went up 160%, from $26 to $67.30. By February 25, 1991, they were back down to $28. Throughout that period, there were vertiginous daily rises and falls, including several of the 13 values greater than 10.41% and the 21 lower than -10.35%, respectively. On the other hand, during the rest of the time, there were also more small rises than you would “normally” expect.
In short, oil prices don’t display nice, predictable uncertainty. Thus, people whose job it is to predict tomorrow’s oil price must cope with more uncertainty than they might expect. The empirical data have many more outliers and many more values closer to the mean.
This article is adapted from “Why Forecasts Fail. What to Do Instead,” by Spyros Makridakis, Robin M. Hogarth and Anil Gaba, which appeared in the Winter 2010 issue of MIT Sloan Management Review.