With her polling lead slipping, Hillary Clinton still has Pennsylvania as a firewall — for now.
David Rothschild, an economist at Microsoft Research who runsPredictWise, an online forecasting model that relies on betting markets, explained the primacy of Pennsylvania for Mrs. Clinton’s election chances during an interview this week.
He also discussed the potential importance of the ground game, the perceived failure of betting markets in predicting “Brexit” and the ethics of live forecasting on Election Day, among other topics. Here’s the lightly edited text of our email exchange.
Pennsylvania
Q. Based on the PredictWise state polling probabilities, the entire election could boil down to Pennsylvania. If Hillary Clinton wins the state, she’ll probably be president. If Donald J. Trump wins there, he’ll probably be president — because such a victory would suggest he’d also win Ohio, Florida, North Carolina. Today, PredictWise gives Clinton a 78 percent chance to win the state. This is close to The Upshot forecast (85 percent). Can you give some more insight into what makes Pennsylvania so important and what signs you’ll be looking for in the state in the next few weeks?
A. Pennsylvania has been the most likely tipping-point state since midsummer.
It has been the state to put Hillary Clinton over 270 electoral votes, should she win all of the other more likely states for her. Conversely, it’s also the state that would put Trump over the hump, if he wins all of the states that are more likely for him.
Every day, I run 100,000 simulations of the election. I use the probability of each state going for Clinton or Trump, then I mix that with a correlation matrix that defines the relationships between the states. And every day since late July, Pennsylvania has been the state that most frequently is won by the candidate who wins the election. Currently, there are just 6 percent of scenarios where Clinton wins Pennsylvania but loses the election, and just 3 percent of scenarios where Clinton loses Pennsylvania and wins the election.
Since Pennsylvania is more secure for the Clinton camp than other swing states, it’s unlikely that Clinton loses Pennsylvania and wins either Florida or Ohio or other states to make up for the necessary electoral votes. And Trump could take Florida and Ohio and North Carolina, and go over the top with some other combination of swing states. But Pennsylvania is his most likely route.
What I will be looking for in Pennsylvania over the next few weeks is simple: polls in Pennsylvania; polls in Ohio, which have similar demographics (and a lot of polling); and national polls that correlate heavily among the key swing states.
Forget Blue States, Red States: America’s Tipping-Point State
As of Wednesday’s PredictWise simulations, there were just 6 percent of scenarios where Hillary Clinton would win Pennsylvania but lose the election, and just 3 percent of scenarios where Clinton would lose Pennsylvania and win the election.
Furthermore, I will be paying special attention to the crosstabs of national polls that focus on key swing demographics for Pennsylvania, when available and reliable, including white women. Beyond the polling for the presidential election, the ups and downs of the Pennsylvania Senate race could be important. The Democratic challenger, Katie McGinty, currently enjoys a slight lead, and that get-out-the-vote campaign will heavily overlap with Clinton’s.
Further, we will learn more soon about ad buys and get-out-the-vote operations in the state. Currently Clinton enjoys a comfortable margin in both categories. If they make a difference — and if they ever make a difference it will be this year with a massive disparity in both advertising and get-out-the-vote efforts — it should give Clinton a slight advantage over the polling average.
Q. So in a hypothetical Election Day in which the rest of the nation sat out the proceedings, and the result was based solely on what happened in Pennsylvania, we would A) almost certainly reach the same verdict (around a 95 percent chance it would match the national election); B) save a whole lot of money and time!
A. But keep in mind: If only Pennsylvania voted, then the outcome of Pennsylvania would be affected by the billions of dollars swarming into the state. Your question highlights a larger problem in our democracy. Our sequential primary system, Electoral College, uneven representation in the Senate and House gerrymandering (natural or not) influence not only how the election is determined, but also the actual sentiment of Americans. It would make my job a lot less interesting, but a national popular vote for president would dramatically alter not only where the candidates campaign, but also the issues about which they, and consequently our elected officials, care.
Polling Leads and Swing Voters
Q. Clinton has a roughly three-point-percentage lead in national polling in the head-to-head matchup with Trump. That’s pretty close. But at this stage, history tells us, even such a small lead can be tough to overcome.
A. The entire polling collapse in 2012 for President Obama around the time of the first debate, including the dip that began before the debate, was only about two percentage points. So there is no question that a three-percentage point disadvantage is hard to overcome. There are just not many swing voters left by mid-September.
But the third-party contenders (Gary Johnson and Jill Stein) are polling better than in recent cycles, leaving a few more people in the possible swing category. Voters in mid-September do not swing between Clinton and Trump (my colleagues and I have dubbed that The Mythical Swing Voter), but between undecided and/or third-party support and Clinton or Trump. So the larger that pool, the larger the potential swing.
While there are certainly surprises that could influence the race, the first debate looms very large as the last major set piece for Trump to make a move. The second and third debates are generally too late. Of course, the election is won on the state level, not the national level, but movements in national polls correlate heavily with state movement.
PredictWise and Personal Feeling
Q. Speaking of that correlation, PredictWise says Clinton’s probabilities are currently in the 80s in in each of these states: Michigan, Wisconsin, Virginia, Colorado, New Hampshire. We talked about the importance of Pennsylvania (78 percent). Then, in order, it’s Nevada (59), Florida (55), Ohio (50), North Carolina (46) and Iowa (45). If you go strictly by those probabilities, along with the fact that Clinton has at least a shot of winning Georgia or Arizona, an observer would have a greater sense that she will win. A national surge by Trump would affect those state numbers, as you say, and we saw some of that this week. But do you personally feel the PredictWise number, a 70 percent chance for Clinton to win, is too low?
A. Obviously, I have constrained myself to let the model run through Election Day regardless of my intuition. If I let my intuition take over, it would almost certainly be worse than my model, on average. But still, an interesting question!
I do not think too much about Georgia, Arizona and North Carolina, because if Clinton wins these states she has certainly won the entire election. The only thing I worry about there is how the tightness affects Clinton or Trump’s ability to wage advertising and get-out-the-vote efforts in other key swing states.
There is a complex relationship between the swing states that I model considering three things: the likely correlated impact of any major events or shocks in the days before Election Day (the correlated impact would be very high); the likely correlated errors between the polling and Election Day outcome (reasonably high); and the likely correlation of random chance on Election Day that affects voting (i.e., rain or shine, and this is not as high of correlation).
The fact that the probability of victory for Clinton in the crucial swing state of Pennsylvania continues to be a few percentage points above her probability of winning the election reflects a very high correlation between the states in both how events could impact all of the states and the sizes and correlation of any potential polling error.
By the way, it is amazing that pundits spend so much time talking about how different this election is, how unique this election is, but the statisticians are looking at the same map and basically taking out Virginia and putting in Pennsylvania. Four years ago it was: Romney must take Florida, Ohio and Virginia. This year it is: Trump must take Florida, Ohio and Pennsylvania.
To answer your question directly, I am not compelled to buy or sell Clinton contracts (that pay off $1 if she wins) at $0.70 . Of course, that is because I trust my model.
No More Waiting for Results on Election Day
Q. What is your impression of the wisdom and fairness of doing live forecasting on Nov. 8? The Times reported recently on a plan that could permanently change the way Americans experience Election Day.
The company spearheading the effort, VoteCastr, plans real-time projections of presidential and Senate races in Colorado, Florida, Nevada, New Hampshire, Ohio, Pennsylvania and Wisconsin. It plans to publish a map and tables of its projected results on Slate, the online newsmagazine.
A. PredictWise will have a lot of Election Day predictions that will approach 0 and 100 percent as the day unfolds. That is the nature of real-time updating market data that powers predictions. And I was happy to call the election for Obama well before 11 p.m. Eastern in 2012.
I am not concerned about how the Slate-Votecastr plan will affect voting. Voting is a major cost for many Americans with hourly wage jobs. Those people who choose to get registered and vote are pretty committed to voting. They are more likely going to fixate on any projection that rationalizes voting than a particular projection that unmotivates them.
If Slate reports the election is tight, people will use that to rationalize their decision to vote. If Slate reports that Clinton is winning, Clinton supporters will use that to rationalize winning big. And Trump’s supporters will turn to Fox News to confirm that Trump still has a strong chance and needs them more than ever. That is why Karl Rove was still predicting a Romney victory to the very end; he knew that some people just needed someone to convince them that it was worth the investment to stay in line to vote.
The Ground Game
Q. You have said that you expect prediction markets and polling averages to converge in the weeks leading up to the election, as they typically would, but with a possible exception this time around: the influence of the ground game. If the Trump campaign’s organization is poor and the prediction markets pick up on that, there could be a “rare divergence,” you said in one of your live Facebook chats. You gave an example earlier of how this could affect Pennsylvania. But how could markets discern something like poor campaign organization?
A. The average of the polls near Election Day, for any given race, are off by two-three points on average, but we do not know which direction that will be. The error occurs because polls systemically undercount or overcount some groups of likely voters. On average, there is no bias for either the Democratic or Republican candidates.
As a researcher, it has been very difficult to determine the effect of the campaigns, because all we get to witness is the net effect of both campaigns spending a fortune. But we are poised to witness the greatest imbalance ever in advertising and ground game spending. Clinton raised $143 million to Trump’s $90 million in August.
Interactive Feature: 2016 Election Forecast: Who Will Be President?
Clinton has 57 campaign offices spread around Florida, while Trump is promising to get to 25 soon.
A smart reading of polling data is a prediction on who will vote and how they will vote. The “Who Will Vote” question, how polling translates into votes, is influenced by the get-out-the-vote operations (and the related voter-suppression efforts). Normally we assume these have a very low net effect, as both campaigns are pouring millions of dollars into states to get the marginal voters out to the voting locations. There is strong empirical evidence that this will not be the case this year. And if markets see that late in the cycle, with both spending and the results of early voting, they will assume that Clinton could be one to two percentage points better than her polling average.
A Beating on ‘Brexit’
Q. The reputation of prediction markets took a beating with Trump in the G.O.P. primaries and particularly after the British voted to leave the European Union. The polls on “Brexit” actually performed fairly well, especially by comparison. You wrote afterward: “In the past, traders followed the polls too closely and sent the prediction markets up and down. But now the opposite is happening. Traders are treating market odds as correct probabilities and not updating enough based on outside information. Belief in the correctness of prediction markets causes them to be too stable.”
How can prediction markets deal with this problem? Should there be reason to worry that markets will again be slow to accept new information?
A. On a race-by-race basis, the prediction-market-based forecasts outperformed polling-based forecasts in the primaries both on accuracy and scope. Further, the PredictWise forecast gave Trump a strong likelihood of victory starting in the beginning of 2016.
As far as Brexit goes, predicting something at 25 percent is not a failure! Seriously, markets aggregated available data and correctly predicted the polls tightening and then “Remain” taking the late lead at the end. And there are certainly distinct concerns in Brexit, not applicable to the American presidential election, around both trading incentives on anoutcome that would devalue the currency and less robust data in a unique event.
To answer your more precise question, there is certainly a growing phase as traders learn about how and how quickly new information is absorbed into the markets. The more established the markets become, the less I worry about trader inefficiency.
I am unconcerned about the market being slow to incorporate new information in the presidential election. Information is coming from many sources, and traders have a healthy feeling that they are the smart money trading against the dumb money. I am also very pleased at the measured responses prediction markets have given to major events, relative to most poll-based forecasts.
Limits of Betting Markets?
Q. Are there any worries that betting markets simply have too few traders, particularly American traders, and too little action to make it a good gauge for United States elections? (It’s illegal for Americans to bet on elections.) And how susceptible are they to manipulation?
A. It does not take too many traders to make a market liquid, so I am not concerned about that. I am concerned about the Commodity Futures Trading Commission’s (C.F.T.C) capricious regulations on the U.S.-based exchanges. PredictIt, for example, is constrained to 5,000 traders in any given contract, and no trader is able to wager more than $850 on any given contract. This led PredictIt to charge larger than normal fees to make up for the low volume they are getting per trader.
Traders respond to the restrictions and fees by over-wagering on long shots, slightly distorting the prices from the underlying values. I correct for this, as best I can, on PredictWise, but it is unfortunate that the C.F.T.C.’s restrictions limit both the overall involvement and efficient involvement of traders.
In addition to providing accurate predictions that we use to understand politics, prediction markets are valuable tools for me (and others) to learn about financial markets. Distorted incentives make that more difficult. Further, can you imagine a world where average Americans invested in prediction markets the way they do in fantasy sports? It would be great to see people researching and rooting at the debates as much as they do each weekend for the N.F.L. We should encourage access to prediction markets.
Perceptions of Probabilities
Q. People tend to have a misperception that with around a 30 percent chance to win, Trump is a big long shot. But it wouldn’t be shocking if Trump won based on those probabilities. In fact, you would expect that the underdog would win in three of 10 such elections. And there’s no preordained decree that this can’t be the year for one of those upsets.
A. 30 percent means 30 percent. So yes, at this point I would expect three of 10 elections to play out with Trump winning in the end.
I should emphasize that should the election be held today, Clinton would be above 90 percent to win it. The uncertainty is that the election is held on Nov. 8, 2016, not today, and Trump still has time to close the race. There are still lingering questions and events that will be resolved.
But she is around three points up in the national polls, and six points up or more in enough states to carry the Electoral College. Combine that polling lead with a larger advertising budget and get-out-the-vote apparatus that should help her perform well in reference to the polling. It would be a catastrophic polling failure if the election were held today and she lost.
No comments:
Post a Comment