Estimating companies’ Greenhouse Gas emissions to aid in environmentally-conscious investing decisions
Given the growing challenges in climate change, there has been a shift towards more climate conscious investing.
Bloomberg offers Environmental, Social, and Governance (ESG) data to aid in the decision making for socially conscious investing.
A major driver in these decisions is a company’s Greenhouse Gas (GHG) emissions; however, since there is no regulation dictating that companies must report their emissions, these must often be estimated to make climate-conscious decisions.
In this paper, we use machine learning methods trained on reported GHG data to generate “distributional estimates” for a company’s Direct and Indirect GHG Emissions.
Classical approaches to estimating emissions start by grouping similar companies into buckets.
Then, one can either take the mean of the reported values in bucket as the estimated value for the unreported ones or fit a simple model on the reported values of the bucket, typically using hand-selected features thought to be the drivers of emissions within that bucket.
For carbon modeling specifically, a typical approach is to bucket companies by industry sector and then fit a linear model in each bucket.
We instead use more complex machine learning methods to learn a full distribution of GHG emissions conditioned on features and trained over all of the data.
Bloomberg provides annual self-reported ESG data for about 13,000 global companies, with data going as far back as 2006. Over 300 fields are offered of which 147 are environment, 66 are social, and 118 are governance.
However, many climate-related fields are reported purely voluntarily, creating a need for reliable estimates of carbon emissions.
Fundamentals data includes the information from the three key accounting statements used by financial analysts to understand the state of a company’s business (the balance sheet, income statement, and statement of cash flows). We used Fundamentals features that pertain to a large subset of the company universe (such as Total Assets and Country of Incorporation) as well as industry-specific fields (such as airborne hours for airlines).
Bloomberg’s industry classification data
For industry classification, we use Bloomberg’s Industry Classification System (BICS), a hierarchical classification of industries; all industry sectors are broken down to at least four levels, with some going as deep as eight levels.
Advanced machine learning techniques
Our full machine learning stack uses Gradient Boosted Decision Trees for Amortized Inference, Recalibration using Normalizing Flows, and Patterned Dropout for regularization.
This aids in:
- Using features that are related to a company’s GHG emissions, even where it is unclear prima facie what the relationship is
- Learning relationships that hold true across all industries, such as the relationship between company size and GHG emissions
- Modeling uncertainty in GHG Emissions. We use this uncertainty to aid in “the precautionary principle”: it is better to overestimate a company’s GHG Emissions rather than underestimate so that a company is incentivized to report rather than hide their emissions.
We evaluated our advanced techniques versus the baseline approach, and found that our model has strong performance across a variety of metrics including squared error, percentage error, and calibration error.
In addition to providing distributional estimates for companies that have reported some emissions data in the past, we also calibrated our model to perform well on companies that have never reported ESG data.
Download the whitepaper to learn more about how Bloomberg is using advanced machine learning techniques to estimate companies’ carbon emissions.
This article was originally published on February 25, 2021.