Dr. Cristian Homescu, Director, Portfolio Analytics, Chief Investment Office at Bank of America Merrill Lynch
Environmental, social, and governance (ESG) criteria are the set of factors through which a company measures their contributions towards sustainable development and societal impact, according to Blackburn (2007). Interest in ESG-based investing has been growing at an increasing pace, leading many to try to identify ESG as a substantial source of alpha, as shown by Chen and Mussalli (2020). ESG metrics for financial analysis are often created from a wide range of un-related figures, which by the nature of its creation, makes the ESG data often non-parametric. This, as explained by De Franco (2019) and Didur (2018), makes machine learning promising and well suited towards analyzing complex ESG data and identifying patterns, which may uncover materiality of ESG on financial performance. In this project, the team analyzed a modified version of the Fama-French Three-Factor model (ESG and Global Compact (GC) factors added) via factor regressions using multiple linear regression and Generalized Additive Model (GAM), two machine learning techniques. This approach was taken so that the team could determine if ESG and/or GC show materiality on excess stock returns. From there, the team constructed various baseline and ESG-based portfolios and compared their performances to determine whether or not integrating ESG into portfolios can lead to financial outperformance. Looking at the available literature, the team’s approach is novel and the results of the project effectively adds to the knowledge base of ESG-based investing.
Keywords: Environmental, Social, and Governance (ESG), Global Compact (GC), Factor Models, Fama-French Three-Factor Model, Machine Learning, Multiple Linear Regression, Generalized Additive Model, Portfolio Creation, Portfolio Analytics
Factor regressions by equity were carried out first, with excess return as the dependent variable and the five factors of the extended Fama-French Three-Factor model (market risk premium, small minus big, high minus low, ESG, and GC) as the independent variables. Table 1 shows the results of the factor regressions by equity, in particular showing the number of companies with/without a statistically significant ESG/GC factor, as well as the percentage of total companies with a statistically significant ESG/GC factor.
The results reveal that more companies have a statistically significant GC factor than a statistically significant ESG factor. These results demonstrate that the GC metric is more material than the ESG metric. Therefore the team decided to use GC as the factor in representing environmental, social, and governance criteria in portfolio construction. A portfolio of the equities was created and factor regressions using multiple linear regression and GAM were carried out on the whole portfolio.
The machine learning techniques were used to determine the explanatory power of the factor model, and therefore demonstrate the importance including a measure of ESG criteria. The results for both machine learning techniques can be compared. The coefficients and p-values of the factors for each machine learning method are shown in table 2.
The results show that excess market return, ESG, and GC are all highly statistically significant in relation to the portfolio’s excess return. This supports the notion that ESG and GC have materiality on financial performance. Since GC has smaller P-Values than ESG, it also supports the notion that GC has more materiality on financial performance than ESG. The results comparing linear regression and GAM are shown in table 3 below:
GAM outperforms multiple linear regression in almost every metric. It has a higher adjusted $R^2 $, comparable p-value, lower AIC, and lower GCV. The landslide outperformance of the GAM method indicates that the relationship between returns and the factor model are often non-linear.
Testing and comparing the results of the factor regression using the ESG Factor model, the GC scores showed more relevancy to the returns of equities across all sectors. Therefore, in accordance with the factor model results, the team selected GC scores to be used as the objective metric when developing ESG-based portfolios. The ESG-Simple, equally weighted and price volume portfolios were created using the original returns matrix consisting of the returns of all 578 equities, multiplied by their respective weights. The returns matrix had to be split at the GC reference level in order to create the ESG underperformers and overperformers.
These matrices were needed to build the ESG Leaders, Best in Class and Laggards portfolios. The reference level for creating the ESG underperformers and overperformers matrix was determined to be 60. This was found by the trial and error method of splitting the returns matrix by different GC scores until the distributions of the GC scores were oppositely skewed.
Once all six of the portfolio classes were created, the performance metrics were calculated, organized and compared, as shown in table 4 shown below:
The portfolio comparison shows that as the GC score increases both the cumulative returns and returns decrease. This holds true until hitting a portfolio value of 60 whereas the GC steadily increases along with returns cumulative and annualized. The risk of the portfolio was inversely proportional to the GC; thus as the GC increased for each portfolio, the risk decreased. The maximum draw-down also showed a decrease generally with the increase in the GC score.
Both the Sharpe ratio and return/risk metric show a great improvement with the increase in GC. This comparison table confirms the notion that the main benefit of integrating ESG is for the return/risk benefits. The table confirmed suspicions that investing against ESG metrics, i.e. the ESG Laggards portfolio, could lead to financial underperformance. The ESG simple portfolio was expected to perform better based on the research conducted in the development of this study. However, according to Giese et al. (2019), this could be due to the fact that US equities tend to perform worse than their peers in other regions given any level of ESG.
The Price Volume portfolio tremendously outperforms the ESG Best in Class and ESG Leaders in all categories but risk. Therefore, investors should not attempt to employ ESG-based investing if they seek to maximize returns.
The Price Volume portfolio also had a comparably better risk return profile overall, which suggests that demand based strategies will often outperform ESG based strategies.
Table 5, shown above, contains all of the portfolio statistics. The statistical properties of the ESG-focused portfolios demonstrated that the returns distribution is characterized by slight negative skewness and a low kurtosis. Negative skewness means that a greater proportion of the data piles up on the right side of the distribution curve. Thus, there are greater positive returns. Low kurtosis means that most of the data is piled up close to the mean; thus, there are less extremes in the movements of the returns.
Therefore, the returns of an ESG focused portfolio consists of frequent small gains with uncommon small losses. The price volume portfolio also shows these characteristic traits with the more negative skewness, and lower kurtosis thus is relatively better for long term returns. The simple, ESG laggards and ESG simple portfolios all show less negative skewness and are more heavy-tailed than Price volume, ESG Leaders and ESG Best in Class portfolios.
There are promising signs that ESG can be integrated into investing in such a way as to meet sustainability goals as well as financial goals. In this project, the team sought to add to the existing and constantly evolving knowledge base of ESG-based investing. A novel approach involving factor models modified by adding ESG and GC scores as factors was used with machine learning techniques to show that ESG and GC display materiality on financial performance.
The main takeaways for the team at the conclusion of this project are that machine learning techniques like GAM can be effectively used to analyze complex ESG data. This analysis shows that ESG and GC do exhibit materiality on financial performance. However, it is still not certain whether or not it can, at this point in time, achieve alpha, or financial out-performance, consistently and reliably in investments.
The main benefit the team found was that ESG-based portfolios tend to have better return and risk profiles, or Sharpe ratios, compared to other portfolio strategies. This aligns with the theory that ESG minded companies tend to have less risk, and demonstrates that a ESG based investment strategy is suitable for those wishing to minimize risk without sacrificing much return. Overall, the team feels that the project did produce novel and interesting results that add to the knowledge base of ESG-based investing, and hopefully there will continue to be interest in studying ESG and its role in investing.
Blackburn, William R. The Sustainability Handbook. Environmental Law Institute, 2007.
Chen, Mike, and George Mussalli. “An Integrated Approach to Quantitative ESG Investing.” The Journal of Portfolio Management, vol. 46, no. 3, 2020, pp. 65-74., doi:10.3905/jpm.2020.46.3.065.
De Franco, Carmine. “Machine Learning Meets ESG: How Computing Power Can Extract Alpha from Complex ESG Data.” Natixis.com, Natixis, 2019.
Didur, Konstantin. “Machine Learning in Finance: Why, What & How.” Medium, Towards Data Science, 11 July 2018.
Giese, Guido, et al. “Performance and Risk Analysis of Index-Based ESG Portfolios.” The Journal of Index Investing, vol. 9, no. 4, 2019, pp. 46–57., doi:10.3905/jii.2019.9.4.046.