Thom Tillis, forecasting enigma.

How to Beat Nate Silver, Revisited

Jonathan Bernstein is a Bloomberg View columnist. He taught political science at the University of Texas at San Antonio and DePauw University and wrote A Plain Blog About Politics.
Read More.
a | A

It's a good time to discuss how the quality of individual candidates and campaigns -- as opposed to the party balance of the electorate and national forces -- might affect the midterm election results.

Dave Weigel of Bloomberg Politics, on Twitter this morning, tweaked projection models for one big miss: early forecasts that Republican Terri Lynn Land would be a solid candidate for the open U.S. Senate seat in Michigan. She's been awful. National Republicans have pulled out, and she appears to have scant chance to win. Weigel, singling out FiveThirtyEight, said that "datajournalists who judged Land to be a good candidate should have done reporting."

It's correct that good political reporting can, in the right circumstances, beat early forecasts from Nate Silver (or the Monkey Cage, or anyone else using objective indicators). But that hardly discredits what these modelers do or how they do it. Remember, too, that even good reporters can and do get it wrong sometimes; after all, there's a whole army of spin artists out there doing their best to confuse the picture.

To think about elections in bulk (the 30-some Senate races in each cycle, or the 435 House elections, or however many gubernatorial seats are up), shortcuts are necessary.

So election modelers use various formulas to assess candidates. For example, simply dividing House candidates into those who have won electoral office before and those who haven't has proved effective. Different models use different objective factors (they may even use pundits' assessments to try to get at less tangible indicators). There's no way to avoid some quick and easy measures if you're try to predict that many different elections.

Those objective assessments are useful for many reasons. Good reporting is scarce, and we simply don't have it for many candidates, especially early on. So if we want to know six months ahead how the Senate will shape up, the forecast models are extremely helpful -- even though informed observers should be able to beat those models in some cases.

In other words, it's certainly not the case that political scientists always dismiss the importance of individual candidates. What matters is the context. In presidential general elections, candidates aren't often important because anyone who survives the nomination process is going to be solid, and campaigns aren't too important because both sides are going to have ample resources and talent available. Those conditions don't hold in races below that level.

The North Carolina Senate race provides another illustration of how tricky objective candidate assessments can be. Republican challenger Thom Tillis is speaker of the North Carolina House, but whether that means he's a good candidate (because he has serious political experience) or a weak one (because his post isn't a statewide elected office) isn't clear. Given how few Senate elections there are with similar candidates, it isn't always possible to figure out how to treat any specific qualifications, and modelers may disagree, leading to differing early projections. That's fine; as long as the forecasters are transparent, we consumers of forecasts can learn from their differences just as we can learn from how they reach consensus.

All the modelers can do is to study the evidence and use their best judgment to tell us what tendencies to expect. That's a pretty good contribution, even if it's not the only way to learn about elections.

This column does not necessarily reflect the opinion of Bloomberg View's editorial board or Bloomberg LP, its owners and investors.

To contact the author on this story:
Jonathan Bernstein at

To contact the editor on this story:
Katy Roberts at