Recent projects

The hours after an election are very special times for financial markets. It can be argued that during this time only the results of the election, and nothing else, drives prices. Using this idea combined with some standard restrictions from the asset pricing literature, we show that betting and financial markets should demonstrate a very particular type of time-series property. This is called cointegration. Simply put, this states that there is only one fundamental trend, related to the odds of the election outcome, driving prices. Any deviations in prices from this common trend should be fleeting and revert.

We test the model on three elections where the results were released over a small number of hours overnight. We find strong evidence for our model for two of these events. However, we find that for the 2014 Scottish independence referendum the outcome became apparent almost immediately. The move in prices was small and quick, meaning that the data did not generate significance for our model. For the two other events, we find deviations from the common trend, and hence market efficiency, of the order of minutes to tens of minutes. This appears to be due to the betting markets leading financial markets, a phenomena also observed for the 2014 independence referendum. This is an important finding. It shows that betting markets reflected the information contained in the vote more quickly than financial markets. We also produce a realistic trading strategy using Bollinger bands that profits from deviations from the common trend for the night of Brexit. This strategy however fails for the 2016 US Presidential election. This was apparently due to the presence of risk aversion whereas for Brexit, investors behaved in a risk neutral fashion.

Policy choices, politics and elections can have profound effects on financial asset prices. Further, election markets have been shown to be excellent forecasting tools, superior to both polls and experts. It follows then that relationships should exist between financial asset prices and election markets. This paper presents a model linking the prices of betting and financial markets in the weeks and months preceding an election.

We use an assumption that a political event has a constant effect on the difference of the expected prices of an asset, given the outcome of that event. An example of this constant effect could be if UK gilts were expected to yield 50bp more if Boris Johnson had won the recent UK conservative party leadership election, rather than Rishi Sunak. Exactly this view was expressed by a pundit and reported in the Financial Times a few days before the election. This assumption yields a model whereby asset price returns are driven by a political factor, related to changes in political markets, and a residue, related to non-political information. The model is naturally extended to equities using the common 5 Fama-French factors to describe the part of returns not related to political risk.

The model is tested on six recent elections from the US and UK. We find strongly significant political factors for four of the events events (once the Fama-French factors are controlled for). There is weak evidence for the model for one and no evidence for a single election. The conclusion for this latter event, the 2017 UK general election, is that this was not informative for asset prices. This was perhaps due to the political chaos in the UK following the 2016 vote for Brexit. An exploration of the political factor weights and firm characteristics reveals some pleasing relationships. We find that firms can diversify away from local political risk with offshore sales. Other relationships are also revealed between firm location and nationalisation risk. For example, the 16 companies identified by Jeremy Corbyn as nationalisation targets in his 2019 manifesto are shown to be significantly more sensitive to a Labour party win than other UK companies. These results, along with the finding of a new, albeit temporary factor driving equity returns, contributes to the literature on asset pricing and portfolio choice.

The UK's vote for Brexit on 23rd June 2016 was the first great political shock of that year (the second being Trump's win). This result was seemingly against all expectations - with betting markets odds for Brexit bottoming out at 10% shortly after voting closed. Much of the literature on the topic suggests the possibility of the presence of mass bias in beliefs about the UK's voting intentions, with a "bubble" in opinion for remain.

This paper studies the question of whether or not this bias persisted on the night of the vote itself. We achieve this by analysing the results from the 382 voting areas of the UK that announced over the course of a few hours overnight. We use a Bayesian machine learning electoral prediction model that uses a parametrised prior, which encodes ex-ante beliefs about voting before results are announced. Our model updates in real-time as results arrive. Using sensible parameters values, we find that the markets were slow to price Brexit by around 3 hours. However, one could argue with the particular choices of parameters in our model. The approach is flexible though and allows us to test results for different choices of prior. One of the parameters in the model is related to the correlation, or dependence, between results of different voting constituencies. This is found to be around 0.3 for general elections and stable. However, a value of under 0.01 is required to produce a model consistent with price action from the night. This is a simply implausible figure that suggests different constituency votes are close to being completely independent. It was as if the first few results that announced (which had surprise votes-shares for Brexit) contained no information about constituency results that were yet to declare.

The conclusion of the paper is that without a shadow of a doubt the "bubble for remain'' persisted well into election night.

Between 2004-2007 I worked with Professor Steve Gull in the Cavendish laboratory at Cambridge. I had originally begun working with Steve on a trading based start-up backed by a bank, and ended up doing research with him. We worked on the development and application of some of Steve's Bayesian procedures, called the maximum entropy method, to the problem of training neural networks. Applying such methods to machine learning is common today but back then this was innovative work. Our first paper, published with a computer scientist, became one of the most cited machine learning papers in Internet traffic classification. Our later work involved replacing a time consuming step in the Bayesian MCMC analysis of the Cosmic Microwave Background (CMB) - Astrophysicists and cosmologists use data collected from satellite probes measuring the CMB to develop and test theories of the universe. We used a neural network to replace the computationally intensive physical calculation of the mapping from a set of theoretical cosmological parameters, to the observed CMB spectra. A package called CosmoNet was published. It was widely used by the community to analyse CMB data-sets. Similar applications of machine learning and neural networks to other fields are widespread today. For example, a good friend of mine is working at Google DeepMind now using much larger and sophisticated neural networks to replace a functional in a common quantum mechanical modelling method.