Growing a Green Decision Tree—Machine Learning and ESG

Growing a Green Decision Tree—Machine Learning and ESG

Assuming that the growth of environmental, social, and governance-based investing is not yet at, but is approaching, a tipping point, and assuming that knowing that tipping point with precision will yield alpha, can machine learning help find it?

That is the question posed in a recent paper by the global head of credit markets for LBBW, Joachim Erhardt. He holds that the tipping point will come about both because of regulatory pressures around the world and because of product innovation, as exemplified by green bonds and green ETFs.

Inadequate Evidence and an Expected Tipping Point

Erhardt considers the evidence that ESG improves performance and finds it inadequate. He calls this “difficult to isolate” in a statistically significant way. The results of studies on this point “vary with time frame, the investment strategy, portfolio constraints, the specific ESG criteria and whether the chosen benchmarks are at all appropriate or representative on a broader scale.”

In Erhardt’s view, the evidence is so unconvincing because idiosyncratic or cyclical downturns are especially unforgiving for the ESG-relevant portion of a portfolio. This is what will be changed when a tipping point is reached, and the question is no longer “why use ESG criteria” but “why not?” That will reduce volatility for the  assets, improving performance.

How can this point be identified? If by machine learning, what can we say about how the algorithm(s) will warn us once we are about there?

Boosting the Decision Trees, and What It Teaches

The system that Erhardt has in mind is a gradiant-boosted decision trees system. A “decision tree” is an easy enough idea. Either A or B. If A, then C or D. But if B, then C/D is out of the question and what follows is either Y or Z. Such forking of branches makes up a tree-like diagram pretty quickly. “Boosting” means the use of learning algorithms in a series to create a strong learner out of weak learners. At each stage, care must be taken to keep the boosting from becoming an over-fitting, to allow the system as a whole to generalize. This care-taking involves the “gradiant” of our “gradiant boosting.”

Such methods have been applied to issues of market prediction and have determined that there are leading markets and there are following markets, with a relationship of Granger causality between them. They have also revealed that Japan is intertwined with the rest of the world with strong macroeconomic interdependence, that China is less connected with the world than Japan, but more connected now than it was before 2010, and that Europe and the U.S. are leading markets .

But none of that is what we want to know. So, again, how do we determine when the ESG tipping point is upon us?

Quoting Erhardt at Some Length

“At each point in time the algorithm picks an optimal portfolio that aims to outperform the market during the pre-defined training window available. Central to the algorithmic selection process are predictive variables (termed ‘Features’ in the field of machine learning) that either identify or quantify characteristics of ESG standards and are of high importance to a learnt model…. It is the relative importance of ESG Features to other Features available to the model–Features that clearly influence fund performance but are not immediately ESG relevant–and how such importance ranking evolves over time that is critical to understanding how ESG influences the fund performance.”

Such a system can monitor for the expected ESG tipping point in real time. A sign will be an increase in the relevance of the ESG features in the “unconstrained model.” (That is, they will be important because the market realities are developing in the direction of their importance, and the system is picking up on that, not because coders are finangling to ensure that they are taken into account.) This suggests that feature importance scores need to be an output of the system and that this output has to be monitored.

As the ESG-constrained and the unconstrained models converge, as Erhardt tells us in the final words of his paper, “that will be the point in time when ESG standards have become a market factor, which in turn allows fund managers to generate ESG alpha.”

Interested in contributing to AllAboutAlpha? Drop us a line at content@caia.org

Be Sociable, Share!

Leave A Reply

← Measuring the Inflation Risk of Three 'Inflation-Hedges' Starting Early: The Growing Need for Financial Literacy in the United States →