In November 2017, Malaysia-based asset manager Farringdon Group announced it had developed an investment strategy managed by artificial intelligence (AI), using neural networks technology.
Farringdon has since cancelled plans to launch the product to its clients. “It didn’t work,” Martin Young, the firm’s CEO, told FSA. “On the way up the algorithm had done very well,” he said. But as the markets went into a downturn in February, it did very poorly, he added. The disappointing performance has prompted the company to shelve the launch.
Neural networks attempt to replicate with software the learning and information storage patterns of a human brain. They learn from data feeds, and are expected to become “smarter” over time.
The data used by Algebra’s AI consists of traditional financial data, as well as unstructured “big data” gathered from the internet, for example the flow of social media content.
As any neural network system is effectively a “black box”, it is very difficult for humans to understand the rationale behind each decision it makes. Young said that his firm’s algorithm appeared to essentially follow the trend, but more aggressively.
“It is very much the same story that we’ve seen in the past with algorithms,” Young said. “Algorithmic trading tends to be trend-following, it tends to do very well in a rising market, because it tends to be the [most aggressive] of all the investors, but on the way down it didn’t curtail what it was doing rapidly enough.”
Farringdon’s machine-learning strategy was developed in-house, using the neural network technology provided by Maxsys AI, a US company specialising in AI for investment management.
The firm hasn’t abandoned the AI strategy but it will not released to clients at this stage. The AI works with the universe of 125 US companies from the S&P 500 index that have been certified as sharia-compliant.
Farringdon had said in a statement in November that the AI-powered portfolio is not a high-frequency trading strategy but a long-term investment approach. A person would oversee any investment decisions made by the algorithm.
Farringdon’s AI system processes US social media data and news feeds, scouring them for relevance to the selected companies. It was trained on historical data to learn the patterns corresponding to positive and negative movements in the stock prices.
Young attributes the algorithm’s failure to the lack of relevant data that can be used to train the neural network. While market data goes back decades, “big data” has a much shorter history.
“Neural networks are not trying to analyse market data per se,” Young noted. “We’ve had algorithms that can analyse market data for a long time. AI tries to bring other factors, such as social media data. If you start to put in that kind of data, you rapidly drop off a data cliff once you go back more than three-four years. After 20 years, there’s next to nothing.”
Considering that the market has not experienced a significant downturn for a decade, the algorithms don’t have enough social media data to learn how to behave when that happens.
Farringdon aims to improve the algorithm by feeding it current data to learn from. While plans for a launch of the product to clients have been shelved, Young said he will revisit the idea after the next substantial market correction.
“The last thing that we want to do is to give AI neural networks a bad rap by launching them before they are ready.”