Seeking signals from ESG data
In recent years, more and more attention has been paid to Environmental, Social, and Governance (ESG) issues and the data companies release to measure their impact. Along with that attention has come a debate about whether ESG data benefits investors in evaluating companies for their portfolios. There is some academic evidence that ESG factors can be positive indicators for an investment, with some studies demonstrating that companies with high sustainability drove better investment returns than their low sustainability counterparts.
Nevertheless, it has typically been very difficult to translate ESG data into a workable investment signal. The data is notoriously sparse and subject to potential bias due to the lack of required or standard reporting frameworks. Bloomberg researchers set out to investigate whether they could overcome these obstacles through the application of machine learning techniques to Bloomberg’s large ESG data set.
Bloomberg’s ESG data
Bloomberg provides annual ESG data for about 13,000 unique companies globally as far back as 2006 and daily governance data for a subset of nearly 4,500 companies back to 2013. This study focused on publicly traded companies in the U.S. with daily data available, which at the time of the study included over 3,000 stocks that could be used for a hypothetical ESG portfolio.
Machine learning strategies
Since ESG-focused investing is seen as generally focused on long-term performance, the approach of this study was to tie ESG metrics to long-term returns directly by building a machine learning model that uses Bloomberg’s ESG data to categorize the annual excess return of each stock in the dataset. The annual excess returns were measured relative to the Russell 3000 index and were bucketed into three categories: greater than 15%, less than -25%, and in between.
A gradient boosting tree (GBT) model was used to predict a company’s excess return category based on data about that company’s ESG approach. One of the significant advantages of a GBT is its inherent ability to handle missing values, without the need for special treatment.
Interpretability
After the machine learning approach produced positive results (a Sharpe Ratio of 1.25 compared to 0.73 for the benchmark, and an alpha of 8.7%), the researchers also investigated the trained model to see if it held up to an interpretability test. The key data features that were driving the model performance were identified using a technique from game theory called Shapley Additive Explanations. The result was a simple and fully interpretable regression model built on those features, which was used for comparison and also to re-train a machine learning model on those features alone.
Not a common factor
All of the portfolios constructed during the research were tested against a Fama/French five-factor attribution. The largest and most robust alpha relative to the factor model came from the machine learning model trained on the top features from the interpretability test. Moreover, only two of the five factors were statistically significant, and that at a low magnitude. This indicates the potential for ESG data to be a unique source of investment insight that is not correlated with commonly used factors.
Download the white paper to learn more about the study and how machine learning strategy can be applied to ESG data.