JOSIAH RUDOLPH: Investing and artificial intelligence
Can asset managers use artificial intelligence models to make better stock market calls?
Can asset managers use artificial intelligence (AI) models to make better stock market calls? At Anchor Stockbrokers we recently published a detailed report showcasing how investors can use AI tools to identify stocks that will out-or underperform in the SA market.
These tools are different to the traditional modelling techniques that have been used in the past, which are ill-equipped to handle the highly complex relationships found in financial markets.
For example, higher dividend yields are generally considered favourable to lower ones, but abnormally high yields can also indicate potential bankruptcy risks.
The problem with traditional linear models is that they are ineffective at capturing these nonlinear relationships.
A recent academic paper titled "Ten Applications of Financial Machine Learning" by Marco López de Prado summarises the key areas where AI has begun to prove its usefulness in complex and interconnected markets.
These areas include more effectively modelling asset price behaviour, risk management and portfolio construction.
Trading stocks is another area of application, as AI can model liquidity in markets that trade infrequently.
Natural language processing tools are also commonly used to measure sentiment from corporate reports or social blogs.
Fundamental asset managers are successfully using AI to calibrate bet sizes on stocks by creating an independent measure of confidence in their predictions.
Outside finance, examples where AI is coming into its own include development of autonomous cars and computer vision to interpret X-rays.
As far as stock markets go, work in this field has been applied mainly to the US equity markets, with a focus on predicting index returns.
Selecting stocks with high AI forecast returns would have helped outperform the SA market by a significant margin since 2007
At Anchor, we have built on the work done by others in predicting the performance of listed shares, which we have extended to SA markets.
We use a supervised learning model called a support vector machine, which learns the highly complex relationship between all the typical characteristics that influence a share price (such as company valuations, earnings performance and macroeconomic indicators) that fundamental investors consider before investing in a company. It looks at these relationships and subsequent performance in an unbiased and statistically robust manner.
We found that selecting stocks with high AI forecast returns would have helped outperform the SA market by a significant margin since 2007.
The flexibility of AI models to changing market regimes offers an edge over traditional linear approaches.
Consider that 2020 was nothing like a normal year, but our model outperformed substantially in the first quarter when markets fell with the onset of global lockdowns.
In fact, in January it was already positioned for what was to come by being overweight gold and platinum counters, as well as some rand hedges, while being underweight domestic SA, like banks and retailers.
A common criticism of these kinds of approaches is their high turnover: that is, quantitative or rule-based investment processes often require frequent trading. So much so that in some cases these frictional costs can eat into any improvement in performance gained. But we can show that the signal generated is actually highly persistent and does not require frequent rebalancing or trading.
In fact, you can train these models to hold stocks for periods of more than a year, making them attractive to longer-horizon fundamental investors.
Another criticism of AI is that it is a "black box" where interpretation of the stock recommendations is difficult or sometimes impossible.
However, there is academic work to suggest that traditional linear models also have significant drawbacks.
Take the growth in Covid-19 infections, for example. These have followed an exponential growth rate, and had we modelled the pandemic’s progression using a linear model we would have materially underestimated the number of infections.
On the other hand, there are more and more tools available to interpret machine learning results to make the recommendations more tangible and understandable to "quant" and "nonquant" investors alike.
There is still much to be done in this space. However, with judicious and sound application it is an exciting and growing area of financial market research, as evidenced by the many academic papers published on this topic.
While it may not be the holy grail in investment, it certainly offers incremental improvement over traditional approaches to investing.
Together these advances add up to something much bigger, which can help fundamental active managers fight the passive onslaught.
- Rudolph is head of quantitative research at Anchor Stockbrokers
Would you like to comment on this article or view other readers' comments?
Register (it’s quick and free) or sign in now.
Please read our Comment Policy before commenting.