Over the last few days, there has been a lot of debate on how much chance Donald Trump has of pulling off an improbable, come from behind victory in the US presidential elections. While it is almost universally acknowledged that Trump has indeed been able to close the huge gap in the polls that existed after the third debate, there is no consensus on what the exact implication of that would be on his chances of a victory. The wide disparity in the results thrown up by the various statistical forecasting tools highlights this issue. The probability of a Trump victory in such models ranges from 35% in the Fivethirtyeight one to a measly 2% given by the Huffington Post and 1% by the Princeton Election Consortium. The models of Upshot and Predictwise are somewhere in the middle, assigning probabilities of 16% and 14% respectively to Trump becoming the next President.
In fact, the probability estimate provided by the Fivethirtyeight model has been much more Trump leaning than the rest of the estimates through most of the election cycle. The reasons behind this are:
- The Fivethirtyeight model assumes a higher degree of uncertainty in the outcome compared to the other models.
- The Fivethirtyeight model is more sensitive to how it adjusts trends when new information comes out in the form of fresh polling data, especially this late in the election cycle.
There are a number of reasons to err on the side of caution and agree with the probability estimates of Fivethirtyeight than the other models, especially the ones showing Clinton has a probability of more than 98% of winning:
- There are a high number of undecided and third party voters even this late in the campaign. As per the RCP average, the percentage of such voters is 12%, as against only 5% who did not vote for either Obama or Romney in 2012. The relatively higher percentage of such voters is mainly because of the high unfavorability ratings of both the major party candidates.
- There is a lot of volatility in the polling data. For example, in the recent polls that came out in New Hampshire, one of the so-called firewall states for Clinton, Clinton has led by as much as 7 points in one and trailed by as much as 5 points in the other. To add to the uncertainty, some of the competitive states like Michigan and Minnesota have not been polled well recently.
- The polls have also swung wildly in either directions – oscillating between a decisive lead for Clinton and a narrow lead for Trump.
- There are a still an unusual number of swing states that remain in play in the election. The polls indicate that Florida, North Carolina, Nevada and New Hampshire are close to toss ups. And while Trump is leading in Arizona, Ohio and Iowa and Clinton is leading in Pennsylvania, Michigan, Colorado, Virginia, Maine and Wisconsin, these leads are not safe there.
- While Clinton is leading in national polls, the lead is currently at or around 3%. This means even if polls are off by a small percentage, the election may become too close to call. It may be noted that the average of national polls taken one week before the election have deviated from the actual election outcome by around 2% in the last 12 elections. In case you think polls have become better over time, they missed the final result by around 2.7% in the 2012 election. And if the recent performance of polls in the UK parliamentary election, the Israeli Knesset election and the referendum on Brexit were any indication, we should not be too confident in the polls. Polls of recent events have also recently shown increasing tendency of having herding bias (i.e. reverting to the mean instead of providing contrarian results) which has led to increased instances of a number of polls being wrong at the same time.
- The errors in the state polls are very often correlated with each other. For example, if the polls have understated Clinton support in Florida by a few percentage points, they are likely to have understated her support in North Carolina as well. So, the polling errors are not likely to cancel each other out in various states.
- This year has also seen the possibility of Clinton winning the presidential vote share and losing the popular vote. This is because Clinton is expected to perform better than Obama among Hispanics, Mormons and college educated white voters. These voters are not particularly concentrated in swing states. For example, Hispanics may swing the vote in Nevada, Colorado and Florida, but their impact outside these battleground states is minimal. It does not help Clinton if Latinos come out in larger number to vote for her in California (which is already expected to vote for Clinton) or in Texas (where the increased support from Hispanics may not be large enough to swing the state in her favour). On the other hand, the turnout of African-Americans may be less than in 2008 or 2012, because Obama is no longer on the ballot. Trump is also expected to perform much better than non-college educated white Americans than Romney did in 2012. Black voters are critical in North Carolina, Michigan, Ohio, Florida, Virginia and Nevada while blue collar white voters are also significant in number in Iowa, Ohio, Florida, Nevada and Pennsylvania. As a result, Clinton may perform better than Obama in blue and red states, but may perform worse in critical swing states. As a result, even though Clinton is ahead in the national polls, her situation in the Electoral College is iffier. Even if she manages the same lead over her opponent that Obama did four years ago, she will probably lose Ohio and Iowa and may also lose Florida, Nevada and Maine First Congressional District among the states that Obama won.
- One may argue that the recent swing in the polls in favour of Clinton has been on account of the spate of negative news (rising premium of Obamacare, Wikileaks revelations, more news on the email story) discouraging her supporters from participating in the polls. But one may similarly argue that the allegations of sexual harassment against Trump had in the same way dissuaded Republican voters from coming out in support of him, thus depressing his support level in the polls taken in that period. These voters are returning back to him, thus resetting the race to the default state. In fact, one does not know whether the default state of the election is an almost tied level of support, a wide victory for Clinton or anything in between. The economic fundamentals do, however, support the notion that the election should be a tightly contested one. This is also reflected in the extremely tight polling being seen in most of the swing state senate races.
- News which have moved polls have emerged late in October and even two days back, when AP broke the news that Melania Trump, wife of Donald Trump and an immigrant from Slovenia, worked in US before getting her work visa, thus violating the law. The presence of a volatile and problematic candidate like Trump as well as a distrusted and scandal-prone one like Clinton is that damaging revelations might keep coming till the Election Day, rendering the race more uncertain.
- The Fivethirtyeight model is essentially similar to the one which predicted a preternaturally stable race that happened in 2012, indicating that the uncertainty is on account of the race and not the model. Further, the betting market odds and the movement of peso (considered a proxy for market sentiment on Trump victory) are in better agreement with the Fivethirtyeight model than other models.
This is reflected in the following comparison of how the odds of a Clinton victory has fluctuated in the various models vis-a-vis the odds assigned to her in the betting markets and the movement in peso over the last two months:
Betting Market Odds vs Forecasts:
Exchange Rate of Peso vs Forecasts:
As is apparent from the first chart, the betting market odds have been closest to the odds given by the Fivethirtyeight model for most of the last two months, whereas the odds given by the Upshot and Predictwise have always been more favourable to Clinton. The movement of Mexican Peso has also been closely been correlated with the movement in the Fivethirtyeight forecast of a Clinton victory, especially in the last one month where they have moved very much in tandem.
The charts, however, show that both the betting market as well as the foreign exchange market has become more bullish on the probability of a Clinton victory in the last few days. This is largely because of the good turnout news that Clinton has obtained in a number of early states, particularly in Nevada and to a lesser extent in Florida and North Carolina.
Around two-third of the expected voters have already voted in early voting in Nevada. Since Nevada discloses the number of early voters according to party registrations, analysts have estimated that Clinton is already leading by a margin which will be difficult, if not impossible, for Trump to overturn on the Election Day. Similarly, the early voting data also shows that voters of the Latino heavy Clinton leaning districts of Florida are coming out to vote in larger numbers than expected. Even in North Carolina, voters who usually have low propensity to vote (and who are generally Democratic voters) have voted in more numbers this year so far.
None of these is a guarantee that Trump will lose in all the three states on Election Day. However, Trump will now have to ensure that the Republican ‘get out the vote’ operation works overtime and is able to turnout supporters at a better rate than what was previously expected. And remember this – Clinton has invested significantly in ground game, against next to nothing of Trump who is mainly relying on the Republican National Committee to turn out voters.
Polls, however, have shown that these states have been extremely tight and even leaning slightly in favour of Trump. Given Trump’s already narrow path to victory, he needs to win all of these states to remain competitive; at the very best, he can afford to probably lose Nevada. However, if he loses either Florida or North Carolina, it is very difficult to see him compensating elsewhere, unless he pulls off major upsets in Pennsylvania and Michigan.
The various forecasting tools, including Fivethirtyeight, do not take into consideration early voting data into their models (unless the same is already reflected in polls which are taken as inputs in the models). As a result, in a number of states where early vote has happened, significant differences have opened up between the percentage of Clinton victory estimated by the betting market and that projected by the statistical model.
The probabilities assigned by the betting market and the Fivethirtyeight model are in close agreement in almost all the swing states. The major outlier is Nevada where based on the data from early voters, the betting market is way more confident of a Clinton victory than the model which does not consider the early voting data. Similarly, the betting market is also more bullish on a Clinton victory in North Carolina and Florida where early voting data in recent days have given a slight edge to the Democrats there.
Now coming back to our original question – what is the probability of a Trump victory? As discussed above, among the various models, Fivethirtyeight probably does a better job of calculating the odds, given the polling information available. However, it has not taken into account the early voting data which has more favourable news for Clinton in some critical swing states.
It will thus be fair to look at the betting market which has closely tracked the Fivethirtyeight forecasts so far in the cycle, but has diverged in favour of Clinton as the early voting data has become available. The probability assigned by the betting markets currently stands at around 78%, higher than the Fivethirtyeight output of 65%. Thus, it would be fair to say that Clinton has something in the range of 75%-80% chance of winning the election. Correspondingly, the chance of Trump stands at anything between 20% and 25%.