How to hone NLP’s detection skills: a cue from a super-sleuth

Navigating the so-called mind palace – a skill that uses context to make mental connections – is an Ancient Greek technique deployed by the smartest of fictional detectives in one of his modern-day incarnations.

But could something like Sherlock’s method help investors realise the utility of natural language processing – the field of artificial intelligence that trains machines to read text – by using similar situational cues to examine a company’s exposure to new threats?

Some asset managers are already using NLP to light up the trail of clues that discretionary investors might use to connect these dots. They are searching with NLP for specific signs of trouble in a firm’s own document trail.

And while such firms are openly broadcasting their use of NLP, critics warn the discipline is still in its infancy. NLP engines grasp facts in a dream-like cluster – a patchwork of information lacking sense or context – the informational equivalent of a word cloud.

These limitations have in the past circumscribed NLP’s utility. NLP-derived alpha signals have proven elusive. Most NLP technology failed to see the Ukraine war coming until very late. 

And although there’s no denying NLP is a ground-breaking technology, in some respects it’s a dumb one. Asset managers have struggled to find ways to get the most out of it.

As they do so, one idea might be to use the technique in bottom-up, stock-by-stock fundamental analysis.

Instead of weighing the importance of events such as the Ukraine invasion in the news – and then gauging stock betas to that sentiment signal – the alternative approach is to investigate a company’s exposure to a specific emerging risk or opportunity.

A more focused NLP analysis can show how often a topic has been discussed in a firm’s earnings calls, say, but also in company documents as well as public information such as the firm’s news.

Case study

For Voya Asset Management, such an approach was helpful during Russia sanctions, when the effects on individual companies were hard to gauge.

According to Chrissy Bargeron, client portfolio manager covering equity at the asset manager, traditional country-at-risk models capture a lot of the more obvious direct risk when it comes to performance. What is less clear-cut, she says, is the indirect risk.

The firm is trying to find out if there are “different themes at play” behind the headline, geopolitical risks, says Bargeron.

The main goal of the firm’s AI machine learning team is to apply ML to fundamental analysis.

It’s a markedly bottom-up rather than top-down approach.

A machine can pick out more easily than a human, for example, whether revenues are declining in a geographical pattern – a signal that might trigger before news headlines alert investors to a problem.

In the Russia-Ukraine crisis, the first aim of risk management was to minimise direct ownership of Russian securities. As it turned out, Voya owned very little.

But the firm also wanted to know which companies it held were carrying some indirect exposure, such as those that sell a lot to Russia or Ukraine.

The AI team applied NLP to news and earnings-call transcripts to look for mentions of either country.

Their engine picked out 59 stocks, for which Voya was then able to gauge sensitivity to the Russian equity market using more conventional quant regression techniques.

“We found this to be very useful. And it was reactive. We did it right when the crisis happened,” says Bargeron. “We were able to quickly scan all of this and see, within our developed equity investments, where we had that exposure.”

Bottom-up, fundamental analysis. Elementary?

Editing by Louise Marshall