Company emissions estimates for private markets: A machine learning approach for investors

Private equity can make up 20% of a multi-asset portfolio, and global PE assets reached $8.2 trillion in 2023. Most private companies don’t report emissions, yet disclosure rules are starting to reach them. California requires companies with more than $500 million in revenue to report, public or private, and the UK, Australia, Singapore and New Zealand are moving the same way. Investors usually hold only one or two financial metrics for each investee.

Industry factors give every company the sector average

The standard workaround is a spend-based emission factor drawn from an input-output (EEIO) model. Every company in a sector gets the same emissions per dollar of revenue, so a steel company using efficient electric arc furnaces receives the same estimate as a competitor with the same revenue. The factors also offer no split between Scope 1, 2 and 3, and no measure of uncertainty.

Trained on public data, tested under private-market limits

With no independent private company dataset to test against, we trained models on reported emissions from more than 10,000 public companies in 25 countries. Inputs were limited to the metrics PE investors typically hold: revenue, total assets, employee count or enterprise value. A fifth of the data was held back for validation, and each model was tested for stability by adding 30% noise to its inputs.

Nearly seven times more accurate than industry factors

Across the reported universe, the machine learning approach produced a median error of about 70%, against about 480% for industry factors. For companies with $100 million to $1 billion in revenue, industry factors overstated emissions by 10 to 100 times. Revenue was the strongest single input. Total assets works as a fallback, and enterprise value performed worst.

Accuracy is best where emissions are material

Company-level error is still wide, and the paper says so. For mid and large caps, median estimates sit within 9–36% of reported medians, and these companies carry 10 to 70 times the emissions of small caps. Small caps are underestimated by at least 50%, and companies under $100 million in revenue are assigned a flat 1,000 tonnes because training data is thin. With typical footprints of 1,200 to 2,000 tonnes, they have little effect on portfolio-level financed emissions.

Download the full white paper for model results by input and scope.