Artificial Intelligence: Solving the riddle with AI
Artificial intelligence is making it possible to glean better ESG information about companies
- ESG investors have few standard metrics on which to base their decisions
- Data gaps can be filled by using AI to crawl a variety of media for defined information – from Twitter accounts to satellite picures
- Whether obtained through voluntary disclosure or from third parties, any such data contains biases and needs reinterpretation
How can factor investors quantify the factors they consider most important? That is becoming a key question in an increasingly sustainability-focused investment world.
Most standard financial metrics are mandatory disclosures for companies, or can be calculated from those that are. That is not the case for most ESG metrics. Indeed, there are few agreed standard measures in this area. Thomas Kuh, head of index at ESG analytics firm Truvalue Labs, says: “Firms new to this are often surprised by how different ESG data is to conventional financial data.”
Stephen Barnett, CEO at financial technology company Util says: “It can be hard to measure non-financial impact; this is multi-faceted and often assessed subjectively. This means it’s difficult to arrive at a consensus that’s a basis for an investment decision.”
Even those disclosures that are mandatory (depending on jurisdiction) such as CO2 emissions, will be the immediate carbon footprint, not for a product’s whole lifecycle. For example, as Ruben Feldman, head of quantitative research and licensing at RobecoSAM, explains: “The production of an electric car produces more carbon than that of a diesel car. So the analysis needs to be able to include the carbon footprint of the cars throughout their functional lives, not simply that of their production.”
The rise of the robots
Artificial intelligence (AI) and machine learning (ML) are increasingly being touted as the way to crack this dilemma. Indeed, Kuh, claims that AI is “solving the riddle of the future”.
A recent AI-focused issue of the Journal of Environmental Investing concluded that, when it comes to existing ESG data, “most metrics were created in past decades to meet the needs of SRI [socially responsible investment] investors who simply wanted the ability to negatively screen out certain industries from their portfolios”. Now, however, mainstream sustainability-minded investors “want a more comprehensive and carefully curated perspective on the companies in their portfolios – which existing ESG data sets too often cannot provide”.
Kuh lists three basic criticisms of the standard ESG framework over the past 20-30 years:
• Lack of transparency, as frameworks have been proprietary.
• Timeliness of information. For example, if a company publishes a corporate social responsibility report in June, it will likely be compiled from data from the previous January. By the time it is rated and published, the information will be nine months old. What is more, that data and rating will then be static for the coming year.
• The bias that comes from companies self-reporting, which is generally not audited.
How it works
AI offers numerous different ways to approach this data problem. The first is simply one of resources – saving time or money. Data that would take a bank of human analysts an age to crunch can be done almost instantly. But AI can do more than speed up already existing processes – it is able to glean data from places a human realistically could not, and make associations between disparate events that a human brain would find daunting, to say the least.
Feldman says RobecoSAM uses “AI that produces more of a return benefit”, with a model that generates about 1,000 data points, rolled up into 120 key performance indicators (KPIs) to corporations. This finds relationships between these KPIs and investment returns, through iterative machine learning.
Scraping and crawling
Much AI, however, draws not on the information that companies report but on what they do not disclose. The problem with data that is relevant to ESG scoring is that not much of it can be found in mandatory corporate disclosures.
Two ways of doing this are scraping and crawling. Crawling involves the use of specialist computer programs – bots – which deal with large data sets by searching throughout the internet. Scraping retrieves information from any source, not just the internet.
These methods present their own challenges. AI can impute missing values for companies with incomplete reporting – or even where there is no reporting – by inferring ESG metrics from two sources: known ESG metrics of reporting companies; and data that defines the similarity between companies. Imputed values should rely on companies within the same industry or ones that use similar ESG-related language. However, such imputed values contain inherent biases, as companies choose to self-report ESG metrics that are likely above average and do not report metrics that are likely below average. This means that assigning a value drawn from this distribution would probably over-estimate ESG compliance.
Investors have to work out what is relevant for them and then go looking for it. It is often found in unusual places. For example, satellite data can be used to gauge a variety of impacts of corporate activity, such as logging and mining. Algorithms can be used to pick up data from social media accounts, rationalising information from, say, employees’ views about their employer to get a better view of its social impacts.
However, Feldman says, “using less-understood data can be problematic, if you’re not careful. For example, if you are scraping social media as to a company’s ESG impact, if some disgruntled employees have a large social media presence, this can be misleading, so you need a framework that can filter this out.”
Util resolves this issue by going for a very specific target – academia. “We use ML to mine 50m academic publications, drawing on best-in-class opinion, mapping what a company does and what its impact is onto SDGs [the UN sustainable development goals],” says Barnett: “The focus on academic work is because the output is higher quality, given that it’s expert and peer reviewed.”
Kuh says Truevalue’s approach is not scraping, but “collecting data and interpreting sentiment from stakeholders’ views of companies. We associate these references to categories in the SASB [Sustainability Accounting Standards Board] framework, whether positive or negative and to what extent.” He says this provides “a very consistent and current output”, delivering daily short-term signals, smoothed out with information gathered over a longer period.
A difficult challenge for ESG investing is that much of it rests on capturing the alpha from sustainability changes that are not yet embedded. If all that is being done is extrapolating from an existing trend, arguably AI is unnecessary. A sheet of graph paper and a ruler will do. So can AI pick out the potential turning points in corporate behaviour?
“While all data is historical, it can be used in such a way as to make predictions as to future investment returns,” says Feldman, using the example of to what degree a company’s prediction for certain metrics, like carbon footprint, matches actual output to predict its likely future footprint.
While Barnett concedes “it’s impossible to predict the future”, he says Util’s systems analyse such metrics as historical return on investment and profit, looking at the momentum of a company or a fund, to determine its likely trajectory.
Ultimately, AI opens the opportunity to gather data and search for associations between it in a way that has not been possible before. The importance of this for ESG is that it can illuminate corporate practices that have, until now, been in the shadows.
As Kuh says, “AI opens up the opportunity for a deeper perspective on ESG issues. Nothing is more important in finance than realising this promise.”