Imagine a drunk walking down the white line in the middle of the road, with traffic passing on both sides. If we attempt to forecast his route, his “average” track is to follow the white line, and if our model was built only to work with this average, we would predict his survival.

In truth, of course, he will wander away from the white line and back in a random walk. The range of possible paths gets wider the further the drunk travels. There is a small chance that he will not drift far enough from the line to be hit by a passing vehicle; but that chance diminishes the further he goes.

The state of the drunk based on his average position is ALIVE. But the average state of the drunk is DEAD.

Prediction based on simple averages is seriously flawed. If you are a decision maker, do not accept any forecast with a single-point estimate. Adding a couple of extra scenarios (“optimistic” and “pessimistic”) is better than nothing, but you are probably still ignoring the full range of possibilities. And you are certainly ignoring outcomes that depend on the path you take to get to the future destination.

Modern computers and modelling technology make it very easy to generate a myriad of scenarios and to vary all sorts of assumptions. Whether with decision trees or Monte Carlo simulation, accept nothing less than a full risk/reward profile for any decision you have to make.

See also:

• The Flaw of Averages, an article by Sam Savage, who coined the phrase.
• Probability Management, Part 1 and Part 2, articles from ORMS in which Sam Savage, Stefan Scholtes and Dan Zweidler propose that every company should have a Chief Probability Officer.
• My review of Sam Savage’s book, The Flaw of Averages.