Estimating the corporate carbon footprint: A meta-model machine learning approach

About 83% of listed companies don’t report their full Scope 1, 2 and 3 emissions. Those that do report 12 to 18 months after the year they cover. Sector averages and input-output models fill some of the gap, but they give every company in a sector the same intensity and can’t split emissions by scope.

30,000 models, narrowed to the best 10

This research estimates all three scopes from the financial data every listed company already reports. We trained more than 30,000 machine learning models on reported emissions and FactSet fundamentals, then scored each on accuracy and on stability when 30% noise is added to the inputs. The top 10 for each scope form the Meta-Model, which takes the median of their estimates. One version uses prior reported emissions where they exist. The other uses financials alone, covering up to 45,000 companies and producing estimates as soon as results are published.

Accuracy is highest where emissions are concentrated

Tested against reported 2022 data, median error was 13% for Scope 1, 18.7% for Scope 2 and 19.2% for Scope 3. In the most carbon-intensive sectors it falls to 5–10%. Weighted by emissions across reporting constituents of seven major indices, including the S&P 500 and ASX 200, median error is about 15%, 17% and 19% across the three scopes.

For investors, the weighted figure matters most. Large errors cluster in small companies with small footprints, while most portfolio emissions sit with large emitters, where estimates are tightest. Every estimate carries a stated uncertainty that can be rolled up to portfolio level.

Where the estimates are weaker

The paper is clear about the limits. Company-level estimates for firms that have never disclosed carry much wider error. Scope 2 is hard to predict because a switch to renewable power doesn’t show up in the financials.

Download the full white paper for the methodology and the complete validation results.