How I Successfully Forecast The Results Of The UK General Election 2015
Abstract: Elementary quantitative finance techniques were applied in the run up to the UK General Election 2015 to try to predict the next government. Comparisons are made between elections and derivatives valuation, and this allows forecasting pitfalls to be avoided, pitfalls that most, if not all, professional forecasters fell into. The results were thus far better than those made by said professionals.
Keywords: Election forecasting; Jensen's Inequality; Transition matrix; Random matrices; Sensitivity to parameters
First let me make it clear that only a handful of people knew that I was attempting to forecast the results of this election. At the time I saw more downside than upside to making a public prediction. And since I wouldn't be writing this paper if I hadn't been rather successful then clearly I am open to being criticized as a major case of survivorship bias. In defence I can only say that I put my money were my mouth was, and more than quintupled that money!
Election forecasting and quantitative finance methodology
There are many similarities between quantitative finance and election forecasting due to the random nature of voting, the nonlinear relationship between the number of votes and the number of seats, and the information available prior to the election.
- Share prices are treated as random in quantitative finance, and voting intentions give the impression of some randomness.
- Derivatives are nonlinear in their payoffs, and with the First Past The Post (FPTP) voting system the person who gets most votes wins, even if that is less than 50\% of the votes. It is possible for a party to get less than 50\% in every constituency and win 100\% of seats.
- In quant finance we have traded prices for vanilla instruments such as bonds, and they may (or may not) contain information about future interest rates. In elections we have poll results. We might want to calibrate our election model just as we often calibrate our quant models.
One might initially think ``Baskets'' as a possible quant equivalent to elections, where each instrument in the basket is a single seat. Although that might seem to be a good comparison such an approach is not the best for it would require treating each seat separately although perhaps correlated to other seats. Think of the number of parameters that would need to be estimated. With 650 seats and half a dozen parties (at least) the number of parameters to estimate is prohibitively large. Instead in my approach I shall rely on human nature being more or less uniform, but with me being less than perfect in quantifying that uniformity. By that I mean that voting intentions can be modelled but that my insight into the parameters in such a model will not be precise.
Background and results
In the last few weeks before the General Election ... to be cont'd!