Can ‘Quarbs’ Help Forecast the Outcome of an Election?

Arbitrage is the practice of making a risk-free profit by taking advantage of a price differential between two or more markets. For example, if it is possible to BUY the number of corners in a soccer match at 10 with one odds-maker (so that a profit is obtained for every corner in excess of ten) and to SELL the number of corners at 11 with another odds-maker, a profit will accrue whatever the actual number of corner kicks taken. Such examples of free money are in short supply, however, and a more realistic trading approach may be to employ what I first proposed and termed in 2000 as a ‘quasi-arbitrage’ or ‘Quarb’ strategy. The assumption underpinning this Quarb strategy is straightforward. It is that the average market opinion (where there is a range of opinions) is a better indicator of the truth than the outlier (or ‘maverick’) opinion.  Take that number of corner kicks as an example. If there are four market-makers, say, three of which offer clients the chance to BUY at 10 and SELL at 11 and the fourth which allows you to BUY at 11 and SELL at 12, the average market price is the sum of the mid-points (10.5+10.5+10.5+11.5) divided by four, i.e. 10.75. If we take this to be the best estimate of the actual expected value of the outcome, a SELL at the ‘maverick’ price of 11 is a value bet. In other words, the ‘Quarb’ strategy advocates the aggregation and averaging of the range of forecasts, implied in the odds on offer from professional odds-makers, to identify the best possible forecast of the actual outcome.

PollyVote (www.pollyvote.com) is an election forecasting site which follows this principle of combining forecasts. By aggregating the vote-share snapshot contained in traditional opinion polls with a panel of American politics experts, a prediction market and a range of quantitative forecasting models, Polly provides a daily updated forecast of the share of the two-party vote that the Democratic and Republican Presidential candidates will obtain on polling day. For their polling input they use a variation of the RealClearPolitics average and for their prediction market they use the Iowa Electronic Markets, a research and teaching-oriented marketplace which allows small bets on a range of contracts including the Presidential vote-share. The panel is made up of respected political scientists and pundits and the quantitative models (based on an analysis of the likely impact of variables ranging from unemployment and inflation to incumbency and wars) draw on an array of different methodologies. Each of these forecasts is assigned an equal 25% weight and in 2004 generated an almost spot-on forecast of the actual respective vote shares.

Is the RealClearPolitics average the best baseline to use for aggregating the polls? Is the Iowa market the best choice of prediction market? Is the chosen panel and the way its views are aggregated the best way to assimilate the combined knowledge of the political intelligentsia? Is the range of forecasting models appropriate? Does the averaging of the four methodologies perform better than any one alternative forecasting methodology?

This debate will continue, and has been informed by how well Polly has performed over the elections from 2004 to 2012 in forecasting the vote shares of the candidates. The answer is very well indeed. In 2016 things went very wrong, but this was not the fault of the combining methodology but rather that all the separate methodologies failed. Having been fed the wrong food, the parrot suddenly got very sick. Time will tell whether it was a short illness or a harbinger of things to come.