How machine learning is helping investors find ESG stocks

In the leafy enclaves of Boston, Mike Chen and his team have taught machines to read Man­darin slang. They know Chi­nese investors often use homonyms, pairs of words with the same spelling but dif­fer­ent mean­ings, to trick state sen­sors on pub­lic forums. By crack­ing this code, the machines can find out what investors real­ly think about China’s lead­ing com­pa­nies.

This is just one of sev­er­al lin­guis­tics projects that Chen, direc­tor of sus­tain­able invest­ing at PanAgo­ra asset man­age­ment, has devel­oped using machine-learn­ing algo­rithms, which can iden­ti­fy pat­terns from the data they receive and learn from their own results. When a com­pa­ny releas­es an update, these algo­rithms enable PanAgo­ra to analyse the reac­tion from its stake­hold­ers and the pub­lic. 

Sim­i­lar advances are being made across the invest­ment com­mu­ni­ty. Investors are using machine learn­ing to mine all kinds of infor­ma­tion, from the minu­ti­ae of earn­ings dis­clo­sures to the con­tent of LinkedIn posts. Using this so-called sen­ti­men­tal data, they can sift through the hype around envi­ron­men­tal, social and gov­er­nance (ESG) issues and get an accu­rate pic­ture of a company’s cre­den­tials. 

“By under­stand­ing what peo­ple are say­ing about these com­pa­nies, we can get a true pic­ture of their per­ceived brand val­ue,” says Chen. “Not only that, but we may be able to detect whether man­agers are ‘green­wash­ing’ when they talk about their firm’s ESG pol­i­cy.”

Across the invest­ment com­mu­ni­ty, researchers and engi­neers are using machine learn­ing in equal­ly dis­rup­tive wys. They’re analysing lin­guis­tic infor­ma­tion from con­tent, includ­ing earn­ings dis­clo­sures to LinkedIn posts, using sen­ti­men­tal data to see whether a com­pa­ny is tru­ly com­mit­ted to ESG and what sort of impact this com­mit­ment has on its stake­hold­ers.

The val­ue of this intel­li­gence is huge. Around a third of all assets under pro­fes­sion­al man­age­ment are now sub­ject to ESG cri­te­ria. In the sec­ond quar­ter of 2020, when the world was reel­ing from the coro­n­avirus cri­sis, investors poured more than $70 bil­lion into green stocks. As issues such as cli­mate change and social jus­tice gain trac­tion, and US Pres­i­dent Joe Biden’s pro­gres­sive regime replaces the ESG-scep­tic Trump admin­is­tra­tion, this pop­u­lar­i­ty looks set to grow still fur­ther. 

To crit­ics, how­ev­er, the meth­ods used to eval­u­ate ESG cre­den­tials have not kept pace with demand. Until now the ESG rat­ings sec­tor has been dom­i­nat­ed by a hand­ful of providers, such as MSCI, FTSE Rus­sell and Sus­tain­a­lyt­ics. Each provider has its own rig­or­ous set of met­rics, but with no uni­ver­sal stan­dard for what con­sti­tutes “good ESG”, their method­ol­o­gy, and thus their rat­ings, dif­fer marked­ly. 

Some providers score com­pa­nies on an absolute basis, so every­one is judged by the same cri­te­ria, but oth­ers score rel­a­tive­ly, which can reward the least bad com­pa­nies in less pro­gres­sive indus­tries. The flaws in this approach were high­light­ed last sum­mer when fast-fash­ion retail­er Boohoo received an AA rat­ing from MSCI  just weeks before reports that some of its work­ers were alleged­ly being paid less than the min­i­mum wage.  

The data these providers rely on is also heav­i­ly influ­enced by peri­od­ic cor­po­rate dis­clo­sures. This data isn’t just prone to bias, as com­pa­nies omit the fac­tors that paint them in a bad light, it is also back­ward look­ing. If a com­pa­ny hired a new female board mem­ber 11 months ago, how can any­one know whether this affects the mar­ket now?

Making ESG data more accessible

As these weak­ness­es have become more obvi­ous, machine learn­ing has become more demo­c­ra­t­ic. The advent of cheap­er off-the-shelf algo­rithms, com­bined with advances in com­pu­ta­tion­al pow­er, mean invest­ment funds and the ana­lysts who serve inde­pen­dent finan­cial advis­ers can cre­ate their own machine-learn­ing mod­els to process huge amounts of data, both finan­cial and non-finan­cial, in real time.

Key to these mod­els is nat­ur­al lan­guage pro­cess­ing (NLP), a sub­set of machine learn­ing that enables machines to under­stand human lin­guis­tic pat­terns. Using NLP, researchers can go beyond tra­di­tion­al mar­ket reports to analyse both writ­ten and ver­bal com­mu­ni­ca­tion to under­stand how ESG com­mit­ments are both pre­sent­ed and received. 

Ana­lysts most­ly use NLP to analyse the lan­guage com­pa­nies them­selves use; whether their dec­la­ra­tions are con­crete or vague, whether they use the first per­son or take refuge in the third. 

The team at HSBC Glob­al Research in Lon­don, for exam­ple, uses lin­guis­tic analy­sis to sift through cor­po­rate earn­ings calls. Mark McDon­ald, head of data sci­ence and ana­lyt­ics, says: “One of our main focus areas is how the pre­sen­ters han­dle impromp­tu ques­tions from ana­lysts, as the answers are often much less pos­i­tive than the pre-pre­pared state­ment at the start. This can give a truer pic­ture and you can aggre­gate the sen­ti­ment across mar­kets and regions.”  

Oth­er firms are focus­ing on how exter­nal par­ties react to these state­ments. They comb thou­sands of news sto­ries to get instant reac­tion and often com­bine this with com­ment scraped from social media.

At Act Ana­lyt­ics in Toron­to, researchers have trained a machine-learn­ing algo­rithm to scour a care­ful­ly curat­ed list of sources from News API, a com­pendi­um of around 30,000 real-time out­lets.

“By its nature, real-time news is meant to pro­voke some sort of response, whether pos­i­tive or neg­a­tive, since that’s what sells,” says Act Ana­lyt­ics’ head of ESG rat­ings Elgin Chau. “What needs to be deter­mined is the degree to which an event is mate­r­i­al and the extent of its impact on asset prices. By apply­ing our algo­rithm across mul­ti­ple news sources, we can cal­cu­late an aggre­gate sen­ti­ment score for a par­tic­u­lar event for a par­tic­u­lar com­pa­ny.

“Much of ESG port­fo­lio per­for­mance relies on con­jec­ture or anec­dotes. Some­times you get some ridicu­lous reports that say your port­fo­lio has tak­en 200 cars off the road or plant­ed 1,000 trees. These are essen­tial­ly sound­bites, which can’t real­ly be mea­sured accu­rate­ly and rely on a rat­ings provider’s sub­jec­tive inter­pre­ta­tion of a company’s cor­po­rate dis­clo­sures. In oth­er words, these types of reports are imper­fect prox­ies for a portfolio’s actu­al ESG per­for­mance. 

“By using machine learn­ing applied to the news, an invest­ment man­ag­er can effec­tive­ly high­light the exact ESG actions a com­pa­ny is tak­ing to pro­mote pos­i­tive impact.”

How machine learning is complementing data

Data providers need not be wor­ried by this trend. Machine-learn­ing advo­cates agree that it will not replace tra­di­tion­al data sources, rather the algo­rithms will work to analyse these sources more effec­tive­ly and put the data into con­text. As Chen points out, investors can now gauge not just the num­bers that a com­pa­ny puts out, but the “believ­abil­i­ty” of its sus­tain­abil­i­ty plan­ning.

As con­sumers and stock-pick­ers pour ever­more mon­ey into eth­i­cal com­pa­nies, so the poten­tial for green­wash­ing will increase. Just this month, research by the UK’s Com­pe­ti­tion and Mar­kets Author­i­ty found that four in ten cor­po­rate web­sites were offer­ing mis­lead­ing envi­ron­men­tal infor­ma­tion on their web­sites.

With machine learn­ing, investors can now find the truth behind these claims. Quite lit­er­al­ly, they can read between the lines.