In a practical example of using machine learning in investing, Quant Insight presented an analysis in the aftermath of Brexit. Quant Insight is a London based quant analytics firm that brings together academics with backgrounds from Cambridge, Harvard and Princeton with investment expertise from former PMs at BlueCrest, Brevan, Fidelity and Millennium.
Together, they have built a framework to help quantify the impact of macro factors on asset prices (even down to single stocks), to optimize trade selection and build thematic portfolios, and to identify market regime changes and spot opportunities.
After the Brexit referendum in 2016, some investors became concerned about the potential economic damage from a disorderly exit from the EU. For those running equity only funds, hedging this risk was a bit of a challenge. While a disorderly Brexit might be negative for the UK economy, a resulting weaker GBP currency would be a boost for large UK firms with substantial revenues from outside the UK.
A weaker GBP simply meant that foreign income would increase in GBP terms, with no particular change in sales. Indeed, the net effect on the FTSE100 index was positive post June 2016. The main chart above the title above also shows that even the broader FTSE350 rallied sharply in mid 2016 and kept going.
The GBP/EUR exchange rate has been a good proxy for Brexit sentiment
The GBP currency has tracked Brexit sentiment quite well over the last few years. Good news on Brexit has seen GBP appreciate, while negative news has led to a weaker GBP against the EUR. How to find a basket of stocks that tracks the GBPEUR Fx rate?
To more directly trade Brexit sentiment then, the problem can be put in terms of finding a basket of UK stocks that tracks the GBPEUR Fx rate out-of-sample. Given the cheapness of processing power, this is something a correctly specified machine learning approach can do well. It can also rebalance the basket every month and ensure that the combination of UK stocks tracking GBPEUR remains effective going forward.
Here is a basket of UK stocks (see chart below), calculated using a machine learning approach, that tracks GBPEUR out-of-sample for the last 4 years, and has a 90% correlation over the period with the target variable (GBPEUR fx rate).
GBPEUR was down 17% over the period (blue line). The ML derived basket was down 16% over the period. Remarkably, the basket tracked the Fx rate almost 1 for 1 in terms of % change from the beginning to the end of the period, out-of-sample. The correlation was 90%. The basket fell sharply straight after the referendum in June 2016, which is quite handy when compared to the subsequent big rally in the other broad UK indices. Source: Quant Insight
The critical point is that this should not be calculated after the fact. In other words, looking at what happened over the last 4 years and fiddling with stock selection and weights until you find a basket that tracks the historical GBPEUR chart is a flawed approach. The odds are that this type of solution will not work as desired in the future.
One must rewind the clock 4 years to May 2015, use only data prior to that, and then formulate the “best” basket. One then needs to see how this basket performs over the next month. The in June 2015, one must repeat the exercise using only data prior to June 2015. This approach is a more realistic simulation as it only allows information to be used that was known at the time.
The outcome of the short basket position was enormously superior to a short position in the FTSE100, 250 or 350 indices post June 2016. We don’t want to get into talking about individual stocks, so we won’t show the basket constituents.