This blog post contains the main points from the paper titled, ESG Investments: Filtering versus Machine Learning approaches.

ESG is one of the hot topics in the investment world. Every portfolio manager, wants to overlay ESG on its investing philosophy. Is there any value in incorporating ESG related ratings for various companies ? In the world of factor investing, there has been active research in validating the importance of ESG as a additional factor in the factor zoo. Also, with a glut of ESG data vendors in the market, there seems to be more noise than signal.

My purpose of reading this paper was to do some data crunching exercise on Asian stocks and figure out answers to a few questions such as

  • Can ESG can act as a factor for portfolio selection ?
  • Can ESG overlay can reduce volatility of a portfolio ?
  • Can ESG portfolio can give a better Sharpe as compared to other traditional factor based investing ?

In this brief post, I will try to summarize the main points from each of sections of the paper


  • Relationship between corporate social performance (CSP) and corporate financial performance(CFP) is fairly old theme in economic research
  • Recent empirical research highlights the link between ESG and alpha.
  • Research on the relationship between CSP and CFP seems to be mixed
  • Skeptical about basic ESG rating to act as an alpha generator
  • ESG ratings are, by construction, a composite measure that dramatically reduces the rich set of information
  • Aggregated ESG ratings are not suited to distinguish between out-performers and under-performers over the long run
  • There is alpha in the granular ESG data, but the relation between ESG and financial performance is not linear


  • Portfolios based on MSCI World Index
  • Proxy of MSCI World Index by using end of month compositions
  • Proxies for US, Europe and Asia developed benchmarks
  • Sector portfolios derived from the MSCI World Index and the regional benchmarks by filtering on stocks that belong to the same sector
  • ESG ratings from Sustainanalytics
  • 3 ratings E + S + G ; 50 narrower indices

Best-in-class approach

  • Remove stocks whose ESG ratings fall below a certain quantile
  • ESG ratings have structural, sector-driven bias and hence the filtering is done over peer groups
  • ESG best-in-class portfolio is derived from a cap weighted portfolio and removes with in each peer group, the stocks whose ratings belong to the lowest quantile.
  • ESG best-in-class filters do not bring outperformance
  • ESG objectives in a portfolio does not significantly modify its risk/return profile
  • Most fund managers report that ESG is mostly looked as a risk mitigation tool
  • Short to mid term financial performance is at best lowly correlated to ESg ratings
    • aggregated ESG ratings have too little signal-to-noise ratio
    • too reductive in nature
    • Granularity is the key
    • ESG investing takes a long time to materialize
    • market regime can affect the overall strength of ESG filtered portfolios
  • ESG filtering does not bring extra performance

Machine Learning

  • Deterministic, easily understandable ML prediction algorithm, aimed at finding consistent and statistically significant patterns
  • Predict excess return of each company over the benchmark, given the specific values taken by some of its ESG indicators
  • To avoid over-fitting, the algo selects only a finite number of rules
  • The predictions of each rule are aggregated into one prediction at each time instant
  • Link between ESG profiles and financial performance exists, but can only be accessed with non-linear techniques