Portfolio Construction and the COVID Cluster
· While we do not have a crystal ball and cannot know which stocks will become most impacted by some future event, we offer a solution of how to mitigate this problem.
· Using advanced machine learning techniques, we can create portfolios with increased resilience to external shocks like COVID-19.
· We demonstrate how to reduce the damage to a portfolio caused by security selections and risks we might be unknowingly taking.
· By adjusting the weights through portfolio optimizations and constraints in our machine learning platform we improved portfolio returns by over 10% during the crisis period.
· While the machine has no concept of COVID-19, it was successfully able to materially reduce its pandemic risk through some practical applications of machine learning.
My colleague Josh Pantony previously discussed how he would use machine learning to help find stocks with the largest exposures to COVID-19. Today, we will examine how using advanced portfolio construction and machine learning techniques can create portfolios with increased resilience to external shocks like COVID-19.
We will do a deep dive on a model to illustrate the techniques that will be helpful – but these techniques apply to all diversified portfolios, they are not specific to the portfolio in the examples. The portfolio we are using is one generated within our platform, Boosted Insights. The portfolio is a Russell 1000 Index model with 100 long positions and 100 short positions, in a ratio of 150% long to 50% short. It will be measured against IWB (the iShares Russell 1000 ETF) as a benchmark for performance.
All examples shown in this article use the exact same stock selection and rebalance periods – that is, the stocks in each portfolio are the same across the board. The only differences are the techniques we use to construct the portfolio and constraints applied to determine the portfolio weightings.
We intentionally selected a model with a very high exposure to the COVID-19 cluster that Josh highlighted in his article, knowing that its performance in March would be poor and its performance in April would be strong. The goal of this post is to help show some techniques that can reduce the damage to a portfolio caused by security selections and risks we might be unknowingly taking.
So, for our baseline we will use those security selections with no portfolio optimization and just weight our securities using Boosted Insights’ “alpha weight” methodology (the highest conviction picks of the machine get higher weights):
The return for this portfolio is not good (underperforming the benchmark by -16%). Compounding really hurts on the downside.
The next set of results is using more advanced portfolio construction techniques designed to reduce the overall risk and overall variance of your portfolio. Specifically, the mean variance construction method of Markowitz Portfolio Theory. Using these techniques, we can slightly reduce the risk of the stock selections:
Again, the overall performance is not fantastic (under performing the benchmark by -14%), but certainly better than the performance of the baseline portfolio. However, the negative compounding in March comes back to haunt us once again. What really hurts are the risks that we could not properly quantify in February, namely that a whole bunch of stocks, previously uncorrelated, became very correlated in March (Josh goes into more detail here). Knowing that we do not have a crystal ball and will not know which stocks will become correlated in the future, how can we solve this problem?
One solution is to apply machine learning techniques to our risk modeling and portfolio optimization. We can use Principal Component Analysis (PCA) to analyze the holdings within our portfolio and attempt to minimize risks that we are not intentionally taking. PCA is a little esoteric in that it can help identify risks within your portfolios, but it cannot name those risks. The output you get from PCA is effectively a list of risk factors and how exposed each stock is to those risk factors. Through our platform Boosted Insights, you can find the largest unnamed risk factor, the next largest unnamed risk factor not associated with the previous risk factors, and so forth. Below are the exposures to some of the latent risk factors that PCA has highlighted for the mean variance portfolio:
You can see it has a large negative exposure to some PCA factors (i.e. Machine 1) and a more neutral exposure to other ones (i.e. Machine 5). We do not know exactly what each of these risk factors are, but you can see throughout the history of the portfolio that we take large negative and positive bets on those factors. These bets are not always intentional and are likely a symptom of a particular sector or sub-sector getting historically cheap (i.e. cruise lines, airlines, hotels, etc).
Given that we are not necessarily making these bets intentionally the question is – what would happen if we limited our exposure to these PCA risk factors when optimizing our portfolios?
For this portfolio construction, we applied similar mean variance techniques but also restricted the PCA factors to a level that somewhat approximated the benchmark’s exposure to these risk factors – as you can see from the Boosted Insights screenshot below:
The result of limiting the risk of the portfolio is that the PCA constrained portfolio performs materially better than its cohorts:
The key point here is that the stock picks in each example are exactly the same, and overall, those stock picks were poor. However, just by adjusting the weights through portfolio optimizations and constraints we improved portfolio return by over 10% during the crisis period. The other important thing to understand is the machine has no concept of COVID-19 or what it is, yet it was successfully able to materially reduce its pandemic risk through this practical application of machine learning.
Feel free to reach out to me or anyone on the Boosted.ai team if you would like more information about how these portfolio construction techniques can help your investment process.