Robo-advisors: Machine learning in trend-following ETF investments
While the term Robo-advisor is quite a broad one, we will try and address the field of Robo advisories from a unique vantage point. In our opinion, you can use machine learning as an optimising tool to help identify potential trends early for ETF investments.
Generally, the objective is to provide consumers with an investment product that is hedged (hence, really safe in the short-term as well as in the long-term), yields an internal rate of returns (IRRs) like private equity, benefits from the structural trends in global macro, and yet is also as passive/simple as an ETF.
Let us now delve deeper into the concept by breaking down the topic into two components:
1. What drives the trend following in ETFs?
2. How can machine learning spot these trends in advance and help provide us an edge?
What drives the trend following in ETFs?
So, what drives the trend in ETFs? It’s simple! If there is a trend in the asset class in which the ETF represents, that trend would be reflected in the ETF too. For instance, if there is a long-term uptrend in S&P 500 index, the same will be reflected pretty much in SPDR S&P 500 (SPY). This same line of thinking can be applied to any index and corresponding ETFs. In addition, two other factors contribute to trends in ETFs.
This is a phenomenon that stems from the difficulty of compounding volatile assets. Known as hysteresis to electrical engineers, it usually affects levered and/or inverse ETFs. For instance, a levered ETF (BRZU) based on the Brazilian equity market was trading above $1,000 per share before the pandemic struck. After the pandemic struck, it lost more than 97 per cent of its value. However, while the Brazilian index (from which BRZU is derived) almost recovered, BRZU never did. The following pictures are illustrative:
Term structure of Futures: A near persistent positive cost of carry dampens ETFs based on Futures in the long run. Similarly, a near persistent negative cost of carry boosts ETFs based on Futures in the long run. This occurs since the cost of carry (whether positive or negative) eventually decays to zero by the time a Futures contract expires.
Using machine learning to spot trends early
Now, let us turn our attention to the machine learning (ML) aspect. How can machine learning sharpen these ideas? In the absence of further knowledge, we can
represent the above phenomenon as random processes (time series concept) with unknown parameters and allow the ML algorithm to arrive at the actual parameters. For instance, you can represent a trending time series process as a self-exciting Hawke’s process without knowing its parameters and allow the ML algorithm to choose the right parameters for prediction.
Finally, what are the nuances to be kept in mind while using ML to solve problems like these and other market problems in general?
1. Try to model market problems as classification problems rather than regression problems.
2. Pay attention to the confusion matrix. There are bound to be false positives and false negatives. Take a call on what is more tolerable and do your problem framing, feature selection, and algorithm selection based on the same.
3. Keep a healthy amount of data as test data. Make sure this is not randomised. Shuffling of data can work in other domains but not in the financial markets.
It is a good idea to opt for strong fundamentals, drag and negative/zero cost of carry, and bet against weak fundamentals, positive cost of carry. How should you decide what fundamentals are strong/weak? Easy: let the machine figure out that for you as well.
What might our passive portfolio look like? Here’s a tentative one but please do not take it to the bank.
· Long global tech
· Long US equities
· Short emerging market-levered equity ETFs
· Long online retail
· Short brick and mortar retail
· Long precious metals
· Short-levered industrial commodities
· Long developed market currencies
· Short emerging market currencies
· Short-levered agricultural
· Long electric vehicles
· Short crude producers/refiners
To conclude, we can say that machine learning is powerful! What is more important is to use a nuanced approach while incorporating machine learning to help drive returns in the capital markets. Machine learning is not a magic button, which you can press to instantly put money in the bank; instead, it can be used as an optimising technique to help improve an already existing, sound idea. If you do it right, you will find yourself cultivating an edge that others are not privy to.
(Penned down by Raghu Kumar (Co-Founder) and Bharath Rao (Chief Investment Officer), RAIN Technologies)