SNF SPARK

Title

Hybrid Approach for Robust Identification and Measurement of Investors Driving Corporate Sustainability and Innovation. Design of Policy Tools for Evaluating the Impact of Specific Investors and Assessing the Quality of Companies’ Investor Bases.

Abstract

Following several decades of profit-oriented research in finance and economics, we have recently been observing a profound transformation in investor perception and a visible shift towards a sustainable financial system. The list of UN PRI signatories includes already over 2,000 large asset owners and keeps growing every year. Several reports (US SIF 2018) indicate that over 20% of professionally managed assets in the U.S. is already being invested according to the principles of socially responsible investing (SRI). While an increasing number of institutional investors are indicating their commitment to social and environmental sustainability, it remains unclear which investors have the most substantial and lasting effect on the sustainability of companies. In this project, we intend to measure the extent to which specific investors influence the sustainability of companies they invest in. We focus on the measurable side of sustainability, in particular the environmental impact of the activity of companies, and the level of corporate innovation, measured using patents data. By combining a previously untested dataset with a novel, hybrid methodology, we seek to answer deep-rooted scientific and practical questions such as whether the investor base affects the sustainability and innovation potential of a company. If so, can we identify and highlight investors who are effectively driving the future development of companies across the globe?

By answering these questions, we will provide clear guidance on the design of policy tools to support investors and monitor the investor bases of companies. To maximize the societal and scientific impact of this project, we will use our findings to design two practical tools for policymakers and investors, which will empower them to make more viable and future-oriented decisions concerning sustainability and innovation. The first tool will enable an evaluation of the impact of a specific investor on sustainability, based on their historical behavior. The second tool will provide information whether company’s investor base is likely to promote its sustainable development. Our approach differs sharply from other projects in this field:

• While the bulk of existing studies either use investor groups, we focus on the impact of individual investors on the evolution of sustainability of the companies they invest in. This will involve building and using untested dataset on investors and company data.
• We extend the scope of the analysis to include crucial investors such as insurers, banks, pension funds, hedge funds as well as sovereign wealth funds.
• We focus simultaneously on sustainability and innovation, which has not been investigated in a comprehensive analysis so far.
• We include European and U.S. companies and investors, in contrast to U.S.-focused studies.
• The proposed hybrid methodology combines linear and non-linear approaches as well as machine learning. It provides a variety of novelties over conventional approaches such as: adaptivity and time variation, addresses the spurious regression problem, and combines linear and non-linear dynamics. We also include a multi-tier data aggregation technique.
• We provide user-friendly assessment tools and seek to maximize usability of results.

Through contacts with national and international organizations as well as public and private sector investors, we will disseminate the research and tools to an academic and non-academic audience

Research approach

In an attempt to combine the explainability of the conventional approach with the analysis of non-linear data patterns, we suggest a novel, hybrid approach that combines the inference-focus of statistical methods with complementary ML’s ability to capture nonlinearity.

In the first stage we propose to tackle the spurious regression problem by using a novel, linear time series approach — the multivariate direct filter approach (MDFA) — as specified for example by Wildi (2005) or McElroy and Wildi (2019a, 2019b). The MDFA addresses autocorrelation of the data explicitly, and, therefore, the results (estimates and statistical tests) are unbiased. Moreover, the MDFA addresses an overfitting of the rich multi-dimensional data by a set of regularization features which are apt to account for the structure of panel data.

In the second stage we propose to apply a ML-based approach — the multilayer perceptron (MLP) — to the residuals of the linear (MDFA-) model in order to investigate whether the neural net can predict the residuals of the linear model. This neural net will focus explicitly on the non-linear features of the data as embedded in the residuals of the linear model, thereby mitigating potential overfitting issues. In terms of the choice of the classical type of neural network as a starting point, we are governed by the properties of the task and of the dataset. MLPs are suitable for regression problems where data are provided in a tabular format so that the results will be used as a baseline point of comparison.

The benefits of this hybrid, two-stage modeling approach are manifold. First, the resulting parameter estimates and statistical tests are unbiased, so an empirical analysis of the active performance drivers is not affected by misspecification issues anymore. Second, we allow for a decomposition of the dynamics into linear as well as non-linear components, the latter being a novelty of our proposal. Third, we address overfitting of the rich multi-dimensional data by a set of novel regularization criteria, which can account for the structure of panel data. Fourth, we mitigate the traditional black-box objection, typically raised against neural nets, by decomposing the dynamics into linear and non-linear features, so the former allows for a comprehensive analysis of the dependency structure. Finally, we can quantify and verify the evidence of non-linear dynamics of the data by formal statistical tests applied (and restricted) to the second stage of our hybrid approach.

Employing ownership data of individual investors together company-level data, over time, plus a variety of controls may result in an input matrix that is too wide or contains substantial proportions of zeros. To confront this challenge, we will consider the data aggregation approach based on building a similarity network between individual investors. Motivated by the approach in Sakakibara et al. (2015) and Blocher (2016), we will construct a sustainability- or innovation-based metric that provides the relative distance between investors by applying the standardized Euclidean distance between each pair of institutions’ feature vectors. In a practical sense, this will allow us to identify communities of investors that are similar across all observable features. This will allow us to aggregate the information on the ownership structure in communities of investors while maintaining a panel setting.

Contribution

The project will have an important societal, economic, and scientific impact:

• policymakers will be able to track and evaluate the quality of the investor base of key companies
• institutional investors (pension funds, etc.) will be able to select funds based on their effective historical impact on underlying companies
• companies and their stakeholders will obtain information on the sustainability impact of their investors
• researchers and society will increase their awareness of the dynamics between sustainability, innovation, and the role of investors in corporate development.

To maximize the applicability and outreach of the research and tools, we contacted and obtained assurances of non-financial support from the following international bodies: UN PRI, WWF, CFA Institute Research Foundation, and Swiss Sustainable Finance. We will also approach data providers, technology companies, pension and insurance funds to obtain user feedback.

References

• Blocher, Jesse. 2016. „Network externalities in mutual funds“. Journal of Financial Markets 30:1-26.
• Sakakibara, Takumasa, Matsui, Tohgoroh, Mutoh, Atsuko, und Inuzuka, Nobuhiro. 2015. „Clustering Mutual Funds Based on Investment Similarity“. Procedia Computer Science 60: 881 – 890.
• US SIF Foundation Releases. 2018. „Biennial Report on US Sustainable, Responsible and Impact Investing Trends“,
• Wildi, Marc und McElroy, Tucker, 2016. „Optimal real-time filters for linear prediction problems“. Journal of time series econometrics.8(2).