Home / Analysis / The hidden risk that nobody is pricing

You can only see it if you look in the right places, but carbon mispricing could have massive implications for how risky an investment really is. Meanwhile, large language models are reading between the lines of earnings transcripts.

Carbon emissions represent an underappreciated risk to company stability and earnings, but have largely gone unexamined by the broader investment community, according to Allspring Global Investments.

“The 2008 credit crisis was basically due to people not pricing risk properly,” says Harindra de Silva, portfolio manager for Allspring’s Systematic Edge team. “This isn’t the same level of risk, but it’s the same kind of underappreciated risk; it’s becoming more and more significant and how different companies manage it could have a significant impact.”

While it’s not something that’s present in existing risk models or included in analyst reports, it can be seen “pretty dramatically” in the pricing of out of the money options on companies that are high emitters, which are more expensive than for a company that’s a low emitter. That behaviour in options markets emerged around eight years ago, and was described by academics Emirhan Ilhan, Zacharias Sautner and Gregory Vilkov a paper in the Review of Financial Studies titled “Carbon Tail Risk”.

“We replicated it and realised that the effect was even more acute when adjusted for industry and country effects,” de Silva says. “For example, oil producers almost by definition have a big carbon footprint. But if you adjusted for that, the highest emitters within the oil industry have the most catastrophe risk associated with them – if you believe the options prices. There’s something in these stocks that people care about and if you don’t account for that you’re missing this level of risk.”

The pricing “tends to wax and wane”; it disappeared almost completely after Trump became president, and in de Silva’s thought experiment would vanish if a cost-effective carbon capture technology was developed.

“We’ve tried to take a company’s emissions, adjust for industry and country – because the availability of energy varies by country – and then you can treat it as a risk factor,” de Silva says. “You can say that there’s a premium associated with it and take a net long position in those companies, or you can have a short position. We’re agnostic; we measure it and maintain exposures similar to the overall equity market.”

Tell us what you really mean

Allspring is also working on harnessing the power of large language models (LLMs) to process company earnings calls and discern what executives really mean. Earnings transcripts have been around for 15 years, but lack of computing power means that it’s been impossible to analyse them at scale in any meaningful way.

“It captures the difference between what people say and what people mean quite well,” de Silva says. “LLMs are very good at identifying that subtlety that’s present in the way people speak. Even when they’re saying something positive it can have a negative connotation.”

The weakness in quant strategies, de Silva says, is that they’re by nature built around the analysis of balance sheets, income data and financial statements and haven’t done a very good job of analysing text. The advent of LLMs is “pretty exciting”.

“What we’ve done historically is just look at keywords – positive groups of keywords and negative groups of keywords,” de Silva said. “That’s pretty useless because everybody has been trained to use the correct keywords. But using a language model we’ve been able to tease things out we weren’t able to in the past.”

Despite the fact that other asset managers are working on similar solutions, de Silva isn’t concerned about signal decay because it’s a “question of horizons”. Company data became more widely available in the mid- to late 70s and quants had a 20 year heyday; the number of groups now applying LLMs in the same manner still quite small. He worries more that the way people speak will change.

“If you think about what we’re doing, somebody is going to have an indicator on their desk that tells them, as they’re speaking, whether they sound positive or negative, and they’ll adjust accordingly. I think when people realise machines are listening, they’ll change their behaviour.”