Telling the future off

Everyone is reading Nate Silver’s book, The Signal and the Noise: Why So Many Predictions Fail, But Some Don’t, these days. I haven’t finished it, but I think I’m getting the idea.

Unfortunately, he has a very persuasive section on the perils of simply extrapolating forward based on exponential trends. (This was the problem with the Population Bomb, for example.) That’s too bad, because I was very happy with my forecast for Family Inequality blog traffic, based on October hits in each year 2010-2012. Here is the extrapolation, shown with the Census Bureau’s U.S. population projection for comparison (on a log scale).

I’ve never had an R-square (explained variance) of “1” before with real data, but that is the fit of this second-order polynomial for the three years of data I have so far — it doubles every year. So that gave me a lot of confidence in the forecast, and I expected to have 100,000 hits in October 2015, and 1 million in 2025. After that it looked like we’d get to 10,000,000 sometime shortly around 2050.

I have done some forecasting based on extrapolation before, with good results. For example, I predicted there would be 2,848 girls named Mary born in the U.S. in 2010, and missed by 0.8% when 2,826 were born. Adding Google search data to the shape of the fertility trend, I also predicted fertility would “rebound” in 2011, which came true if you count “bottoming out” as a “rebound.”

Finally, I predicted a slow summer for weddings in 2011, based on flaccid seasonal bumps in Google searches for wedding invitations and bridal showers in 2010 — and that was born out by falling marriage rates in 2010 and 2011 compared with 2009 (according to the American Community Survey), as the Pew Research Center just reported:

Of course, marriage rates have been falling for many years, so that wasn’t such a risky prediction.

Anyway, most extrapolations are just good as conversation pieces, not actual forecasts. Here are a few recent ones in that category:

Working wives earning more than their husbands

Liza Mundy extended the trend line fore the percentage of working wives whose incomes are greater than their husbands, writing: “If you plot this rise in a linear graph, you can see that if the trend continues, the percentage of working wives who are primary earners would cross 50 in about 2036.” Here is her figure:

This is rhetorical. It is a statement about the pace of change, not intended as a realistic prediction. One obvious problem here is the boundedness of the data series. It starts in 1988 just because that’s when the series reported by Bureau of Labor Statistics happens to begin. If you started in 1800 or 1900 or 1950 you would get a different prediction.

More importantly, this is a social process; you need a reason to expect it will continue as it has from some arbitrary starting point. As I’ve written, changes in the sectoral composition of the economy aren’t helping much anymore, and women’s share of the labor force has been stalled for almost two decades. And women’s share of BA degrees hasn’t budged in more than a decade (source):

Still it works fine as an “at this rate” rhetorical flourish.

Character counts

Similarly, in an interview with the Ms. Magazine blog, Geena Davis has this exchange:

Q: In the past two decades, has there been any improvement in how women are portrayed in media?

Davis: No, there hasn’t been. Our research actually covers the past two decades, and what we found is that the ratio of male to female characters did not improve over the 20 years we studied. Well, actually, it improved by 0.7 percent. So that’s something. It means we should achieve parity in about only 700 years.

That report she released is worth looking at, by the way. It is: Gender Roles & Occupations: A Look at Character Attributes and Job-Related Aspirations in Film and Television.