55/45 is a really close race

Life lessons from gambling about understanding election probabilities.

After a drought, we finally have enough interesting polling to do a Model Talk on where the race stands: perhaps we’ll take another tour through the swing states, for instance. But we’ve been leaning heavily on the paywalled side lately, so let’s save that for later in the week and instead run a free column that clears up a big pet peeve of mine in how people interpret our model’s possibilities.

I’ve never seen an election in which the forecast spent more time in the vicinity of 50/50, and I probably never will:

There were a couple of times when the election seemed to be getting away from Donald Trump and Kamala Harris respectively, but they proved to be false starts. After she took over for Joe Biden and her polls immediately began rising, it seemed possible for a moment that Harris would ride her momentum to a relatively straightforward win. But then her polls began declining after her convention when typically this is one of the best periods of polling for a candidate.

And then after the convention, there was a perfect storm of factors that had the forecast drifting toward Trump: shifts against Harris in a series of high-quality national polls, plus some economic headwinds (there’s since been better data) and the lingering effects of the model’s convention bounce adjustment. The debate came at this moment, however — and Harris won and reestablished what’s been a consistent lead of about 3 points in national polls — close enough where the Electoral College is roughly 50/50 or maybe a slight Harris edge. (The model’s hypothesis that the momentum had been shifting against Harris before the debate will never quite be appropriately tested, I suppose.)

However, on average, since our forecast relaunch on July 30, Harris has won 49.4 percent of simulations, and Trump has won 50.2 percent. (These don’t quite add up to 100 because of the slim possibility of a 269-269 Electoral College tie.) People understand intuitively that a 50/50 or 49/51 forecast is a toss-up.

If the forecast is 55/45 in some direction instead, however, confusion can abound — even though this isn’t any different from 50/50 for most practical purposes. Some of the problem is that people can confuse this forecast for a prediction of vote share: if Harris were to win 55 percent of the vote and Trump 45 percent, that would be the biggest landslide in an American election since Ronald Reagan in 1984. But that’s not what this forecast is saying. Rather, it’s that Harris will win the Electoral College about 11 times out of 20 and Trump will win it 9 times out of 20: still basically a toss-up, just with the coin weighted ever so slightly in Harris’s favor. This point from the evolutionary biologist Claus Wilke is absolutely correct:

Gamblers understand this, perhaps like nobody else. Yes, if you think Harris is really a 55/45 favorite and prediction markets have it the other way, you have a profitable wager — but you’re still going to lose the bet almost half the time. (There have been much bigger discrepancies between models and prediction markets in the past.) In sports betting, the breakeven point on a point spread bet is generally 52.5 percent — that’s just enough to cover the vig — but most experienced bettors would avoid the 53/47 games¹ to reduce variance and because there’s ultimately no way to know that your model is right. In poker, you can’t avoid making tough decisions — you always have to fold, raise/bet, or check/call — but many decisions in the game are so close that you’re literally supposed to randomize your play, employing a mixed strategy because being unpredictable outweighs any slight edge from one line being superior to another.

There’s another, more advanced property of probabilistic forecasts that sports gamblers (or otherwise highly quantitatively-minded sports fans) intuitively understand. It’s late in this election and the Electoral College is close. When a game is late and close, then tiny things — a single point or goal or run scored — can make a huge difference. Here, for instance, is an estimated win probability chart for an NBA game for a team that has possession of the ball with two minutes left to go in the fourth quarter; say it’s the Knicks down 106-105 to the Celtics. If Jalen Brunson is fouled and makes both free throws, the Knicks’ win probability shoots up from 45 percent to 69 percent; that’s absolutely huge.

This property is even more profound in baseball, where single-run leads are relatively durable² because runs tend to come in bunches and the default outcome is not to score at all. If the visiting team trails by a single run with nobody out and nobody on in the top of the 8th inning, their win probability is just 22 percent — but if they lead by a run, it’s 76 percent (!) instead. Both games would be considered close, potentially coming down to the outcome of a single play like this gem from the San Diego Padres’ Jurickson Profar, but the odds shift to almost 3:1 in one direction or another. (Framing probabilities in odds ratios like this can sometimes be helpful. A 3:1 advantage is meaningful in the way that 11:9 just isn’t.) Conversely, if the visiting team is already up by 5 runs, a leadoff home run hardly makes a difference: it’s already almost sure to win, well past the point of diminishing returns.

If you look closely at our election forecasts, they follow this same principle. Our forecast of the popular vote has barely moved, other than when RFK Jr. quit the race and both candidates gained ground:

But the probabilities have, because the translation from the popular vote to the Electoral College follows the same baseball game principle. If she goes from a 2-point popular vote lead (rounded down³) to a 3-point lead, Harris’s win probability zooms up from 28 percent to 59 percent.

Some news event in the final four weeks may move the numbers — the hurricane response or if Trump agrees to another debate. But if I had to bet, I’d guess that the probabilities stay somewhere in the 50/50 or 55/45 or 60/40 range. With extremely few undecided voters (Harris and Trump combine for 95.5 percent of the vote in our national average and third parties typically get another 1 or 2 percent) there just aren’t enough votes in play to really move the needle. And polls simply aren’t accurate enough to provide for much more confidence than that. This is where the sports analogies fail too: in elections, the “final score” (the last polling average before Election Day) isn’t always correct.

1. And maybe the 54/46s, but you can’t be too picky: edges are generally small.

2. This is why closers are overrated, but I digress.

3. Or more technically, lead of 2 to 3 points versus a lead of 3 to 4 points.