Hilary Clinton looks set for victory in tomorrow’s US Election.

National Polls Give Clinton a Small Lead

Three new national polls released yesterday give Hillary Clinton leads of 1 point, 5 points, and 5 points, respectively, for an average of 3.7 points. Here are the numbers:

Sponsor Pollster Clinton Trump Johnson Stein
Investor’s Business Daily TechnoMetrica 45% 44% 5% 2%
WSJ/NBC (Not stated) 48% 43% 6% 2%
Washington Post/ABC News Abt-SRBI 48% 43% 4% 2%

With a lead of almost 4 points, it is almost certain that if these numbers hold, she will also win the Electoral College.

Prediction Models Agree that Clinton Will Beat Trump

We don’t deal in probabilities here because that only makes sense when you can repeat the experiment. If a weather forecaster predicts an 80% chance of rain tomorrow, the forecaster can be rated by looking at the most recent 100 times that the prediction was 80% and see if it indeed rained on about 80 of them. With elections that doesn’t work, so we just stick to looking at the state polls. Based strictly on state polls, it looks like Clinton will win tomorrow.

Nevertheless, other election forecasters do deal in probabilities, which always involves some model, and the models differ. The simplest model looks at the state polls and runs a Monte Carlo experiment. For example, if a poll says Hillary Clinton is going to get 47% ± 4% in Florida, on each run a number is chosen from the range 43% to 51% either uniformly or according to a Gaussian (normal) distribution. This is done for each state based on the polls of that state and the electoral votes are tallied for that run. Then a million runs are made and the fraction of runs in which she gets 270 electoral votes is called the probability that she wins. But even in this model, there are many choices to be made. Does only the last poll count? How about the last three polls or the last week of polls? Are the pollsters weighted for past performance? Is a uniform or Gaussian distribution used? Are other factors used? The wide varieties of choices the modelers make accounts for the differences in predictions. That said, here are some current predictions that Clinton wins.

Predictor Chance Clinton wins
FiveThirtyEight 65%
New York Times 84%
Daily Kos 87%
PredictWise 89%
Huffington Post 98%
Princeton Election Consortium 99%

All the models say Clinton will win, but the range is 65% to 99%. This rather broad spread, with FiveThirtyEight the outlier, has been the source of much commentary in recent weeks. And the commentary reached something of a fever pitch this weekend, when the Huffington Post’s Ryan Grim slammed FiveThirtyEight’s Nate Silver, accusing the polling guru of “unskewing” polls in Trump’s favor, through what Silver calls a “trend line adjustment.” Grim writes that, “He may end up being right, but he’s just guessing. A ‘trend line adjustment’ is merely political punditry dressed up as sophisticated mathematical modeling.”

Grim is certainly right about one thing (see below), but otherwise he’s probably overstating the case. Silver isn’t pulling his numbers from thin air (as the original unskewer, Dean Chambers, was). The results that the FiveThirtyEight model produces are, as the New York Times’ Josh Katz points out, primarily the product of two major assumptions that Silver has made. The first is that his model is ultra-sensitive to daily changes in the polls (aka “shiny object chasing”). The second is that his model assumes that shifts in one state tend to result in equivalent shifts in similar/neighboring states (aka “correlation”). This means that the FiveThirtyEight model would be expected, for example, to change rapidly in response to James Comey’s announcement 11 days ago (as it did). And it will likely change rapidly again, now that Comey has said there was nothing on the laptop.

Katz, for his part, is mildly critical of the choices Silver has made, explaining that the Times prefers to be a bit more conservative. Princeton’s Sam Wang—who has done battle with Silver before—is a bit more critical, arguing that FiveThirtyEight’s model likely double-counts swings in polls. Vox’s Andrew Proskop, by contrast, defends Silver’s approach.

As we note above, these aren’t the only assumptions and decisions that a pollster makes. And this brings us to the part that Ryan Grim is 100% correct about: Polling is a science, yes, but it also incorporates a fair bit of art. Anyone who suggests otherwise (as Silver is sometimes wont to do) is not being 100% forthright with their audience.

Betting Markets Say Clinton Will Win

While the polls lag behind the news, the betting markets are real time. The instant important news appears, bettors can place or sell bets so this might give a more up-to-date picture, although in the past, betting markets have gotten it wrong on occasion. Here is what some of the betting markets say:

Market Chance Clinton wins
Iowa Election Markets 69%
PredictIt 81%
Election Betting Odds 83%
Paddy Power 83%
William Hill 83%

Here the range is narrower than what the mathematical models say, but again, Clinton is the strong favorite.

Last polls

To hear SurveyMonkey tell it, Donald Trump’s rust belt strategy is working, since they believe he’s keeping it close in Michigan, Wisconsin, and Pennsylvania and is ahead in Ohio. We’ll soon know if they are right. We don’t believe the Nevada poll as too much of the early voting campaign occurred during the poll and it went strongly for the Democrats.

State Clinton Trump Johnson Start End Pollster
Alaska 31% 47% 13% Oct 31 Nov 06 SurveyMonkey
Alabama 36% 54% 6% Oct 31 Nov 06 SurveyMonkey
Arkansas 34% 55% 6% Oct 31 Nov 06 SurveyMonkey
California 56% 30% 7% Oct 31 Nov 06 SurveyMonkey
Colorado 43% 39% 11% Oct 31 Nov 06 SurveyMonkey
Connecticut 51% 37% 7% Oct 31 Nov 06 SurveyMonkey
D.C. 88% 7% 2% Oct 31 Nov 06 SurveyMonkey
Delaware 49% 39% 7% Oct 31 Nov 06 SurveyMonkey
Florida 45% 45% 4% Nov 02 Nov 04 YouGov
Florida 47% 45% 5% Oct 31 Nov 06 SurveyMonkey
Georgia 45% 45% 6% Oct 31 Nov 06 SurveyMonkey
Hawaii 52% 29% 9% Oct 31 Nov 06 SurveyMonkey
Iowa 37% 47% 9% Oct 31 Nov 06 SurveyMonkey
Idaho 29% 47% 7% Oct 31 Nov 06 SurveyMonkey
Illinois 52% 35% 7% Oct 31 Nov 06 SurveyMonkey
Indiana 35% 52% 10% Oct 31 Nov 06 SurveyMonkey
Kansas 36% 48% 10% Oct 31 Nov 06 SurveyMonkey
Kentucky 35% 53% 6% Oct 31 Nov 06 SurveyMonkey
Louisiana 38% 52% 4% Oct 31 Nov 06 SurveyMonkey
Massachusetts 56% 29% 7% Oct 31 Nov 06 SurveyMonkey
Maryland 60% 29% 6% Oct 31 Nov 06 SurveyMonkey
Maine 47% 39% 7% Oct 31 Nov 06 SurveyMonkey
Michigan 44% 42% 8% Oct 31 Nov 06 SurveyMonkey
Minnesota 46% 37% 10% Oct 31 Nov 06 SurveyMonkey
Missouri 40% 48% 8% Oct 31 Nov 06 SurveyMonkey
Mississippi 41% 50% 4% Oct 31 Nov 06 SurveyMonkey
Montana 31% 53% 10% Oct 31 Nov 06 SurveyMonkey
North Carolina 48% 41% 7% Oct 31 Nov 06 SurveyMonkey
North Dakota 29% 57% 10% Oct 31 Nov 06 SurveyMonkey
Nebraska 34% 52% 8% Oct 31 Nov 06 SurveyMonkey
New Hampshire 49% 38% 6% Nov 03 Nov 06 U. of New Hampshire
New Hampshire 49% 38% 8% Oct 31 Nov 06 SurveyMonkey
New Jersey 53% 37% 4% Oct 31 Nov 06 SurveyMonkey
New Mexico 40% 36% 16% Oct 31 Nov 06 SurveyMonkey
Nevada 44% 43% 10% Oct 31 Nov 06 SurveyMonkey
New York 51% 34% 5% Nov 03 Nov 04 Siena Coll.
New York 58% 31% 5% Oct 31 Nov 06 SurveyMonkey
Ohio 42% 45% 8% Oct 31 Nov 06 SurveyMonkey
Ohio 45% 46% 3% Nov 02 Nov 04 YouGov
Ohio 48% 47% Oct 27 Nov 05 Columbus Dispatch
Oklahoma 32% 56% 10% Oct 31 Nov 06 SurveyMonkey
Oregon 51% 35% 7% Oct 31 Nov 06 SurveyMonkey
Pennsylvania 47% 42% 6% Oct 31 Nov 06 SurveyMonkey
Rhode Island 49% 36% 7% Oct 31 Nov 06 SurveyMonkey
South Carolina 44% 46% 6% Oct 31 Nov 06 SurveyMonkey
South Dakota 31% 53% 12% Oct 31 Nov 06 SurveyMonkey
Tennessee 40% 49% 6% Oct 31 Nov 06 SurveyMonkey
Texas 42% 47% 6% Oct 31 Nov 06 SurveyMonkey
Utah 31% 34% 6% Oct 31 Nov 06 SurveyMonkey
Virginia 49% 39% 7% Oct 31 Nov 06 SurveyMonkey
Vermont 60% 27% 6% Oct 31 Nov 06 SurveyMonkey
Washington 51% 34% 7% Oct 31 Nov 06 SurveyMonkey
Wisconsin 44% 43% 7% Oct 31 Nov 06 SurveyMonkey
West Virginia 27% 57% 8% Oct 31 Nov 06 SurveyMonkey