Winning the Losers’ Game: Using Machine Learning & Big Data to Optimise Investment Returns
With Patrick Armstrong, CIO Plurimi Wealth
Active management is, in aggregate, a loser’s game and we can prove this mathematically. The market is the sum of all investors, and is split between both active and passive investors. Passive investors own a pro-rata percentage of the stocks that constitute the market, the remainderis made up of active investors. Active investors are different sides of the same coin. For every active investor overweight an asset, their must be another underweight that same asset. In aggregate their returns will equal the market return, however managers on both sides of this trade will charge a fee. In aggregate, active investors returns are the market return, less all active investment manager fees charged.
To justify an active approach, investment managers must have the following:
1. An information advantage.
2. The ability to process information more effectively than their competitors.
Information advantages were more readily available historically. With the advent of Regulation Fair Disclosure (Reg FD) in 2000, companies could no longer selectively disclose important information to market professionals and certain shareholders. Reg FD has levelled the playing field for all investors. Under this regulation, companies that conduct earnings and forecast calls to update stock analysts must simultaneously issue a press release to make that information available to the general public. This improves both efficiency and confidence in markets, but it also removes a reason to invest actively. In the past market participants who had close connections with company executives could generate outsized returns based on these information advantages.
Ability to process information better than competitors
1. Human judgement
2. Big Data and Machine learning
Skilled human judgement may allow individuals to process public information and generate better returns than the market. There have been many instances of star managers, who outperform for many years, but then they have often been viewed as having a style bias rather than a judgment advantage. We think there is scope for outperformance based on this approach, but it is hard to objectively assess.
Big Data and Machine learning has a set of attributes that are particularly useful when applied in the investment industry. Vast amounts of information that are continually updated make it impossible for a human to systematically analyse and interpret all of it. The exponential growth in data means we now can know more about companies than we ever have in the past, but the enormity of the task of processing all this data is beyond the ability of any human.
Sourcing and consolidating information on companies used to be a competitive advantage, but that has now become a commodity. We are in an era where there is no difficulty getting information, the problem lies in the overwhelming amount of information.
‘Big Data’ refers to extremely large data sets that may be analysed computationally to reveal patterns, trends, and associations. Big data analytics are able to make sense of large amounts of data and uncovers these trends and patterns with in it. Machine learning can accelerate this process with the help of decision-making algorithms.
The investment industry is data rich, with daily movements on prices, new estimates continually being updated by experts, and a wide range of financial information being published, alongside live pricing of thousands of assets. This provides fertile grounds for any process which can systematically review and assess all this data.
In recent years financial models were limited to linear and logistic regression techniques. Identifying good predictors through different regression models. Today, enormous amount of data can be interpreted by machines which are able ‘learn’ how to make predictions based on the data. We are strong believers that utilising AI and Big Data is the most robust, scalable, and sustainable source of competitive advantage in investment management today.
Plurimi’s Wealth’s approach:
Active management has always been about uncovering opportunities before they are priced in by the broader market, and we believe that processing information more swiftly, robustly and objectively than our peers allows us to do this in a consistent and repeatable manner. This is the area we have focussed on in our investment process. We implement a data-driven investment approach that can objectively evaluate public companies globally, through fundamentally based and economically motivated investment themes. Our process utilises a large set of company-specific data from financial statements, as well as market data including prices, returns, volatility, and broker estimates of various financial statement items, investment recommendations and target prices. Our system endeavours to identify strong businesses in terms of balance sheet and profitability, with attractive valuations, positive sentiment, and a positive momentum in terms of price and estimates. Data is the basis of our investment process, but the additional research on buy candidates and portfolio construction processes still requires human judgement, and this is done to ensure we do not fall into any spurious conclusions provided by our machine model.
Machine learning provides pattern recognition and identification of correlations with leading lagging degrees. For example, machine learning may pick up that companies tend to beat on earnings, also tend to outperform over the next quarter. Or, that a string of downgrades on airline earnings estimates, tend to precede under performance in aircraft manufacturers. Machine learning identifies patterns such as these, but also more complex and non-linear relationships as well.
Machines do not sleep, and do not take holidays. In our process their only reason to exist is to identify stocks that have the characteristics that lead to outperformance. Our approach allows us to process incredible amounts of data, and to uncover connections that aren’t as obvious to other investors, or at least not yet obvious. We have found our machines often Identify companies in emerging trends before they become mainstream knowledge.
A good example of this was in February and March of 2020. Our system started to highlight video game companies as the most attractive stocks in the coverage universe., using this information bought Activision, Tencent and Nexon. As we do not utilise non-financial data, or our machines did not know about a pandemic, they did not know about lockdowns which would increase demand, but they did have visibility of analysts upgrading the earnings estimates and increasing target prices in this sector. They also saw companies which were trading at near market multiples but offering considerably higher growth. During the same months, the machines supplied data, ensuring that we did not have any traditional oil companies, despite highlighting very attractive valuations on many of the companies.
The machines identified that downgrades in cyclical estimates across industries would lead to poor performance from these stocks despite the attractive value scores.
Machine learning adapts as the data changes and to different market regimes. In aggregate our portfolios generally are overweight value, quality and momentum, and the machines generally value these characteristics. However, there are times when it becomes clear to the machines that traditional value factors are not driving a particular stock, sector, or even the overall market. Machine learning techniques adapt to the constantly changing data at all levels. We use it to make more informed predictions on which stocks have the potential to outperform the market.
Portfolio characteristics. Higher yield, lower valuations, less debt, higher profitability growth than the market.
Artificial Intelligence guided buys and AI authority to sell.Source: Bloomberg/PW Holdings as of 31/1/2021
While machines are very powerful in processing information, they are not infallible, and they lack nuance at times. We have adopted a ‘man + machine’ process in selecting stocks to invest in, but the machines are never overruled when they signal a sell. We view buying any stock as a risk, and prefer a human overlay to reaffirm any AI buy rating. On any given day, our AI identifies 150-250 stocks which have the characteristics to outperform the market from a universe of more than 3000. We select stocks to buy exclusively from this list. When AI recommends a sell, we immediately follow the recommendation. Selling a stock only creates a potential opportunity cost, rather than risk of capital loss, so we have less need for a human overlay on this part of the process. When a single stock is sold, we have 150-250 stocks to choose from that machines have identified with winning characteristics, so this mitigates the opportunity cost risk.
We combine AI and our own common sense during portfolio construction and diversification.
Since we launched our global equity strategy, it has delivered returns of 32.1% per annum gross of fees, which compares very favourably to the MSCI World return of 16.6% per annum. It has generated far greater returns with less volatility and drawdowns than the market.
Gross of all fees Source: Bloomberg/PWTotal return in USD terms. (30 Nov 2018 – 31 Jan 2021)
The age of AI and machine learning is here. Investors who embrace it, and take advantage of its dispassionate objectivity, superior ability in the processing of information and tireless approach to identifying investments which will outperform, have a sustainable edge vs those who do not.
Investors who invest actively and do not have an approach with a clear information advantage will inevitably underperform and transfer value to those that do.
Patrick Armstrong, CIO
Eugen Fostiak, Head of Risk and Quantitative Strategies