Saturday, December 28, 2013

A Sucker Born Every Minute ...

In these times, it seems superfluous to present more evidence regarding the old adage (attributed to P.T. Barnum) that there's a sucker born every minute.  Social media are rife with ignorant nonsense.  Television is dominated by shows of monumental vapidity sponsored by products preying on people's narcissistic concern for their self-image.  Gambling casinos hum with activity 24/7.  Politicians convince people to vote for things that clearly are not in the self-interest of those voters.  And so on ...

Nevertheless, I'm moved to provide yet another example: the marketing of long-range forecasts.  Virtually any respectable meteorologist knows that our ability to forecast the weather accurately in a deterministic way decreases with increasing lead time.  For those of my readers who don't know what 'deterministic' means, consider this product:

Note that the high and low temperature forecasts in this National Weather Service (NWS) product are given to within one degree Fahrenheit for each forecast time.  This conveys no information about increasing uncertainty in the accuracy of the temperature forecasts, so this is a 'deterministic' temperature forecast.  There might be various ways to show that uncertainty, but this sort of product simply makes no attempt to do so.  

It should be evident to most people that uncertainty increases with time over the period of the forecast, but nevertheless it seems that many forecast users are uninformed about this.  The product above is not a 'long-range' forecast, of course, being less than a week ahead.  Beyond a week or so into a forecast, the accuracy of weather forecasts is no better than what you would find if you simply forecast the local climatological averages for that date in the future - in technical terms, after about 8-10 days, the forecasts have no skill over a 'climatology' forecast!  A skillful forecast is one that is more accurate than some standard forecast method, such as random guessing, persistence (every day will be just like today), climatology, or whatever standard you wish to choose.  [Accuracy refers to the difference between what is forecast and what is actually observed.  Accuracy and skill are not synonymous!]

The same is true for the sky conditions and sensible weather forecasts in the product above.  However, observe that the weather forecast for "tonight" mentions a "chance" of freezing drizzle.  What does the word "chance" mean to you?  Do you think everyone interprets that word the same way?  This language is at best an attempt to describe uncertainty, but it uses words for which the meaning is unspecified.  The language of uncertainty is probability and a proper forecasts should always contain information about the uncertainty. 

In 2012, AccuWeather began issuing long-range forecasts out to 45 days, well beyond the 8-10 day limit of skillful predictability.  In those forecasts, no uncertainty information is provided, so to the user, the level of precision in the forecasts beyond the predictability limit looks just the same as the forecast for tomorrow, which is at best a deceptive practice, arguably bordering on unethical.

Recently, a study of the accuracy of the long-range forecasts from AccuWeather for selected cities was done.  That study shows what any meteorologist already knew:  AccuWeather forecasts exhibit no positive skill over climatology beyond about 8-10 days (or less) and in most cases show negative skill beyond that of climatology after that time. The important information that the uncertainty increases with time is not an explicit part of their forecast.  For NWS/NOAA forecasts out beyond a week or so, there's a different sort of product suite - see here - that provides a non-deterministic sort of forecast product.

Most users typically don't keep track of what the forecast was even a week ago, to say nothing of the forecast 45 days ago!  They also don't typically subject the forecasts to rigorous verification analysis.  Hence, they naively 'look at' long-range forecasts and perhaps even use them to make personal decisions.  It would be interesting to interview a cross-section of users of those long-range forecasts to ascertain their opinions regarding their value and how they go about dealing with the decline of accuracy with time in the forecasts.  It's hard to imagine how an unskillful product would be of much value to users ...

It seems that many people are at least attempting to use long-range forecasts somehow, and private sector companies provide their clients with what they want.  Unfortunately, such products are not what users need, which is a forecast with uncertainty information included.  When users aren't informed about forecast uncertainty, they have to guess for themselves how much faith to put in those forecasts.  Capitalizing on user ignorance by issuing deterministic long-range forecasts beyond 8-10 days is a shameful practice.  Caveat emptor - let the buyer beware!!

No comments: