Like the rest of the world I was caught off guard by the result of the US presidential election. I was talking to some IFAs the other day about the Trumpflation investment story (every major US stock index has hit a new record high on expectations of fiscal stimulus) when a question caught me cold: markets thought Trump was going to lose, just as they thought “remain” would win the Brexit referendum. Why should we take any notice of them now?
It got me thinking. Markets are not infallible. They assign probabilities, not certainties, to the future. Some events are so complex that assigning a probability is difficult enough; the variables are numerous and potential correlations unpredictable. Other events are random, unexpected, unknowable.
But the US presidential election was not so complex. There were only two likely outcomes. On the one hand, this limited choice can increase volatility as investors take fright at getting caught on the wrong side of the bet. On the other, the information markets use to price the election was straightforward: opinion surveys undertaken by pollsters, election forecasts derived from the polls and the interpretation of these by the prediction markets.
As the Financial Times noted before the US vote, the divergence between prediction markets and opinion polls was striking. While Hillary Clinton’s poll-of-polls lead was just 1.7 per cent by the Monday of election week, the prediction markets continued to ascribe a very high probability to a Clinton victory.
Predictwise, which aggregates prediction markets, went into election day with an 89 per cent chance of Clinton winning. That is just an 11 per cent chance of Donald Trump winning. The New York Times overall forecast was similar at 85 per cent for Clinton, while Pollster and PEC (aggregates of opinion polls) called a 98 per cent and 99 per cent chance respectively.
These forecasts fed into market sentiment. If the predictions reflected the best evidence available, then the problem is not markets’ predictive abilities. An unlikely event happened, but that does not discredit the probability ascribed before the event. But did such high probabilities assigned to a Clinton victory accurately reflect the information available about the outcome?
Over a month on, it is worth noting that Clinton not only won the popular vote but is now 1.7 per cent ahead in the popular vote tally – very much in line with the vote poll predictions. But the US presidential election is a first-past-the-post electoral college. This is not new. The pollsters know, as do the markets, that what matters are individual state polls.
Which brings us to the election forecasting models and the opinion polls on which these forecasts were built. Polls are the one major data point for predicting elections. Markets scour them in the hunt for information about election outcomes. But if the methodology is flawed, so too is the information.
There is some evidence of flaws in the methodologies that underpinned the US polling. Here, it is important to distinguish between the polls and the forecasting models, which in most cases simply aggregate the former.
Nate Silver of FiveThirtyEight, famous for his accurate probability-based approach to the 2008 and 2012 presidential elections, assigned a much higher likelihood to Trump winning (29 per cent) in his final forecast than his peers.
His greater caution reflects a recognition that election forecasts are, in the end, dependent on a single data source: polls. In Silver’s words, “People mistake having a large volume of polling data for eliminating uncertainty. Polls tend to replicate one another’s mistakes.”
Thus, the high probabilities of Clinton winning assigned by the forecasting models. They failed to account not for the possibility that polling errors existed, but for the possibility that polling errors are correlated. Quantity of information (lots of polls) was mistaken for quality (uncorrelated polling errors) in many quarters. Including by the markets.
The Vix “fear gauge” stayed low because Trump was not expected to win. Trump was not expected to win because of the data the markets crunched.
This data was single source and markets failed to price in the possibility that errors in individual polls might be correlated.
The risk of relying on a single data source went unexamined, even by market participants skilled at exploiting pricing mistakes. With more significant elections/referendums still to come across Europe, one wonders how markets will react. Understanding political risk has never been more important to developed market investors.
Gregg McClymont is head of retirement savings at Aberdeen Asset Management