Inefficiency and predictability in the Brexit Pound market: a natural experiment

Abstract

Exploiting the near-experimental conditions provided by the GBPUSD exchange rate during the Brexit vote of 2016, we quantify a significant delay of the market price in reflecting the increasing probability of a Brexit outcome over the vote counting period. We claim that the Brexit outcome could realistically have been predicted hours before the market adjusted to the outcome. This inefficiency is identified by comparing the market-implied probability of a Brexit outcome with a separate probability, estimated by a standard Monte-Carlo algorithm based on a simple linear regression model, representative of what should have been easily possible in real time. The core of the method is the real-time re-calibration of ex-ante ‘pollster’ predictions for the voting district outcomes by regressing the observed voting results onto them. For comparative purposes, a study of the MXNUSD exchange rate in the 2016 US Presidential Election was done, finding that the market-implied and model-estimated probabilities moved more consistently toward the Trump outcome. Put together, this identifies a somewhat anomalous breakdown in market efficiency in the case of the Brexit vote, which we attribute to its novelty as well as a kind of political bubble and subsequent crash, generated by confirmation bias and social herding.

Keywords:

• Prediction
• natural experiment
• market failure
• efficient market hypothesis

JEL classifications:

• C53
• D72
• G17
• C51

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 This sounds like a Bayesian setting, where the posterior reflects the updated prior belief, however we do not use Bayesian methods and hence use the terms ex-ante, real time, and ex-post to indicate quantities before, during, and after the event.

2 The mistake in our working paper (Wu, Wheatley, and Sornette 2017) pointed out in Appendix B of Auld and Linton (2019) was a misprint in the paper. The analysis and code were in fact correctly implemented.

3 The regression method is the standard Weighted Linear Regression with the weights set as electorate size. The fit and coefficients are highly significant (F-test and parameter p-values $\le {10}^{-16}$), but the data is noisy ($R^2$ measure about 0.3).

4 We also analyzed a two-factor model, which exploits turnout information by linear regression re-calibration of Brexit vote turnouts onto district turnouts from the 2015 General Election. However, we do not find significantly improvements for this two-factor model.

5 Data availability is an important issue for the US Presidential Election. The state-level data and the media calls for each state are the only available data for the public. According to The New York Times (Grynbaum 2016; Rodriguez 2016), The Associated Press and the US's five major news networks, ABC, CBS, CNN, Fox, and NBC, are all members of a consortium called the National Election Pool, which has provided election night information including the vote count, analysis, and projections since 2003. These networks use what are called ‘decision desks’, which employ dozens of statisticians and pollsters to project winners based on their analyses and the new organizations' proprietary statistical models.

Additional information

Funding

This work was supported by H2020 Marie Skłodowska-Curie Actions [643073].