The Brexit vote: A case study in causal inference using machine learningThe Brexit vote: A case study in causal inference using machine learningDeveloper Advocacy Manager

In this blog post, we’ll answer the question, “How did the Brexit vote impact exchange rates between the British Pound and US Dollar?” To do so, we’ll use causal inference techniques to estimate the impact of what statisticians call a “treatment,” in this case a policy decision.

Please note that this is a technical blog post aimed at educating about concepts and tools with public data, not any political or economic implications. The techniques we’ll discuss here can apply to all kinds of scenarios, such as the impact of a marketing campaign or product introduction on sales.

Causal inference is needed because we don’t have a controlled experiment for this scenario. An ideal experiment contains carefully matched groups, except for the explanatory variable being investigated. Many real-world situations in which we are trying to find meaning don’t meet those conditions.

We’ll need to find another time series that closely follows the US Dollar : British Pound exchange rate, but was not impacted by the Brexit vote. From this other time series, we’ll derive the counterfactual: what was expected to happen, had the Brexit vote not occurred. We’ll estimate the effect as the difference between the counterfactual and actual time series.

Our scenario

After the Brexit vote on June 23, 2016, the British Pound (GBP) dropped from 1.48 versus the US Dollar (USD) to 1.36 the following day, and continued to decline.

In contrast, the Euro:USD exchange rate did not change much, despite being highly correlated to the GBP:USD exchange rate. The daily values of the two exchange rates had a Pearson correlation coefficient around 0.75 during the 5 year period prior to the event. So, we’ll use the Euro:USD exchange rate as a control.

To estimate the effect, we’ll consider the following 4 weeks as the post-treatment period. We could extend this period out further to estimate the full effect. However, the longer of a window we use, other factors come into play, and it becomes more difficult to isolate the effect of the treatment alone.

Below you can see a chart of both exchange rates, along with the shaded area indicating the post-treatment period:

Leave a Comment