On Precision

FiveThirtyEight has an article that claims Ms. Clinton has a 53.3% chance of winning compared to Mr. Trump's 46.7%. Now, this is a projection three months ahead of the election, three months in which who knows what will happen in the world, and who knows what atrocious things will be attributed to the candidates.

Yet the projection is given with an absurd number of significant digits. Really? 53.3%? Not 53.4%? I bet that the eventual results don't match the units digit, and maybe not even the tens digit.

Whoops! Those numbers were from three days ago. Today (August 3) the same article cites 68.4% and 31.6%. So, in the course of less than a week the tens digit for both candidates changed. No margin of error is given but it appears to be in the +/-100% range.

It's not just politics. Many companies sell, at great expense, predictions about how industries will perform. In our own business IHS, VDC, Gartner and others supply a steady stream of pontifications. International Business Strategies predicted a semiconductor slowdown of 1% for 2016. Just three months later they changed that to 2.13%.

Give me a break – their own forecast was off, they think, by a factor of two, but they can confidently produce a new number with three significant digits.

The same article gives January forecasts from other professional prognosticators. IC Insights figures the market will grow by four percent this year. Gartner predicts a 1.9% increase. They all diverge wildly. Only IC Insights has the humility to give a nice round number.

Gartner reported that semi sales declined by 1.9% in 2015 after predicting a 2.2% increase in July of that year.

The 1.9% figure is based on real data. It is a fact, and the accuracy implied in that number is reasonable. To predict a 2.2% increase (not 2.3%?) when history shows these forecasts are crazily uncertain, is, in my opinion, the very model of hubris.

One wonders if these folks use a dartboard to create their analytics.

Companies are paying big bucks for this data. Well, not data, which Oxford Dictionaries defines as "Facts and statistics collected together for reference or analysis." The wild inaccuracies mean these numbers are neither facts nor statistics. Speculation would be more appropriate. Or guess. Maybe divination.

But not data.

Around 1995 I was Powerpointed by an analyst who claimed, with lots of charts, figures and earnestness, that by 2000 there would be essentially no 8- or 16-bit processor market. In fiscal 2016 Microchip sold $1.3 billion of MCUs, most of which were not 32 bitters. TI's MSP430 business is going strong. Plenty of other companies are doing well in that space.

Economics is called the dismal science for good reason. Analysts don't get such a moniker because they fly far under most people's radar. I've been reading industry projections for a very long time and the most accurate observation I can make is that they are seldom right.

We had an unusual Spring here in Maryland this year. Forecasts were so far off the mark, even for a day ahead, that they became a local joke. I asked a weather forecaster if he or his colleagues were graded or rewarded on the accuracy of their predictions. He smiled and said "nope." Industry analysts seem to be of a kind with weathermen.

Must be good work if you can get it.

For the next three months the e-sphere will be consumed with presidential poll numbers that will change constantly and mean little. Pundits will confidently spout data to three significant digits while citing error bands that swamp the implied precision. Too many gullible consumers of this nonsense will be seduced by the implied accuracy.

I do have a very precise prediction for the race: Ms. Clinton or Mr. Trump will win. Unless something unanticipated happens.

But I'm hedging this by claiming a margin of error of +/-100%.

Published August, 2016