“Can big data actually be made useful? Can it help us develop better products?” The rhetorical questions are being asked by Michael Kollo, London-based deputy head of research at AXA Rosenberg Equities, a €18 billion quantitative investment shop.
The answer is not yet a clear ‘yes’, and proponents such as Kollo are exploring how valuable machine learning will be. “We’re experiencing a data renaissance, but it requires investment,” Kollo said.
The stakes are high: if the answer turns out to be ‘yes’, it means machine learning will do more than build a better mousetrap: it will completely change how quant firms make their buy and sell decisions.
Kollo sat down with DigFin to explain the thinking behind how AXA Rosenberg is harnessing machine learning to its quant investment business.
Kollo has worked as a quant at asset managers including BlackRock and Fidelity. They all have their own business models. AXA Rosenberg pursues factor-model investing, using big data sets and machine learning. He is also charged with promoting innovation within the larger AXA Investment Managers organization – “Data sets are becoming institutionalized,” he said – but quant businesses are more natural adopters of machine learning.
Projects so far include tracking sentiment via social media to develop real-time indicators for portfolio managers (starting with inflation, for the fixed-income team), developing credit-scoring models (in partnership with French fintech DreamQuark), and using machine learning to assist the sales team pitch new ideas to customers.
More fundamentally, Kollo believes machine learning will help asset managers rethink their production process, rather than just help them do the same thing more efficiently.
What the machine sees
Machine learning is a subset of A.I. The science of making computers intelligent relies on machine learning to train them into making autonomous decisions. Kollo’s brief primer serves to illustrate how this will change portfolio decisions.
He shows DigFin an image of the numeral 8. The goal is to get a computer to always recognize this numeral, even though people might write it in all sorts of ways. Even so, we will recognize a numeral 8 when we see it. A machine cannot intuitively grasp what makes an 8 an 8. So it has to be taught in another way. (The images Kollo showed actually come from a series of posts on Medium by an A.I. specialist, Adam Geitgey.)
A machine will read an image of the numeral as a series of spaces that are colored lightly or thickly, or not at all. It divides the image into many small blocks. The computer assigns each micro-block a number that corresponds with the level of darkness it sees there.
When this is reproduced as an image, you can still make out the numeral 8, as a series of bigger numbers where there is ‘ink’, and zeros or ones where there is empty space. A human looking at this first rendition sees a recognizable if abstract numeral 8 among a field of numbers.
This taking inputs – a series of numbers that correlate to the amount of ‘ink’ visible in an image – and turning it into a set of numbers creates what is called a vector.
But another step – another set of vectors – is required to understand what is happening. The abstract numeral 8 amid a field of numbers only looks like an 8 because the grids are arranged in the same shape as the original image. But the computer doesn’t want or need to see this in, say, size A4. Instead the numbers simply run on, in order, but without regard to shape. To the human eye, the numbers run on into gibberish. We no longer see the 8.
But the computer does, when it is programmed to. Its neural network reconceives this string of numbers – random to the human eye – as a probability: the probability that this vector means “8”. The process of turning a number into a probability is called ‘logistic transformation’.
“There’s nothing intuitive here,” Kollo said. “It’s pattern recognition.”
So over time, as the computer is fed more examples of numeral 8s, it converts these into vectors, long strings of numbers that to us mean nothing, but to the computer correlate to “8”. Eventually when it scans an image of an 8, no matter in whose handwriting, it will correctly identify it as such.
This kind of logic applies to all sorts of things. Computers use it to identify faces using 128 such vectors. That’s what a single person is to a machine: 128 probabilistic computations that to us look like very random batches of numbers. Faces are hard. Easier, then to turn credit scores, financial statements, earnings figures and financial forecasts into vectors and recreate them as probable matches – which can then be fed into an algorithm of buys and sells.
Restructuring buys and sells
In isolation, this may be happy news for quant investors, but it doesn’t change the business model. It might just be a better way of creating a trading call.
But Kollo says machine learning will upend how quant firms make market decisions.
“We will go from looking at financial statements or a company’s product, to a company’s tradeable network. What if a machine can recognize connections between entities that matter?”
In other words, quant investing – and by extension, many forms of active fund management – will no longer be about analyzing the company. It will be about how a company’s amassed vectors correlate to other companies’.
“A portfolio manager can track a hundred stocks,” Kollo said. “But not a hundred corporate relationships.”
But by turning company information – all of its numeral 8s – into machine-readable vectors, computers can then begin to recognize connections among companies that no human could spot. Not even managers of the companies themselves.
And as computers get faster, portfolio managers will create endless permutations of questions about a stock. That means far more ways to take original views on investment decisions, giving fund managers new ways to generate alpha.
Dream versus reality
That’s still science fiction. “Today we don’t know the rules well enough,” Kollo said of the technology. For now, more effort is going towards using machine learning to validate consensus economic theory and to understand how patterns are created, rather than trying to outsmart markets. The way that computers arrive at decisions, using logical transformation, is still not understood by programmers, and they want to get a grip on it before unleashing alpha-generating (?) monsters into the world.
“It’s a distrust of purely data-led investment,” Kollo said, noting again that Rosenberg isn’t a black-box firm.
That said, AXA Rosenberg is using its machine learning experience to introduce new products. Factor timing is first. Now that smart beta has been embraced by institutional clients, can the machine understand how long-term returns behave in specific time periods, and time these to avoid market losses? It’s a project that combines A.I. with the harvesting of massive amounts of data from all kinds of sources.
Many quants are working on this, and factor timing usually involves a variation of tactical asset allocation techniques, but Kollo believes that a trading strategy based on machine learning will outperform.
Another application is ESG funds, as environmental, social and governance mandates are big business now in Europe. From a quant’s perspective, machine learning can find correlations between ESG language in company statements and the degree to which these are implemented.
But does this information produce better alpha?
“I don’t know yet,” Kollo said.