Transformer-Based Timeseries Forecasting for Company Returns with Multi-Provider ESG Data

Abstract

The goal of this thesis is to analyze the effectiveness and impact of ESG ratings through the lens of machine learning interpretability. We first start by motivating the thesis through studying the correlation between ESG ratings and controlled log returns. We demonstrate that there is a correlation between the growth of a company and commitment to sustainable initiatives. We then establish two new machine learning models: the Non-stationary inverted Transformer (NSiTransformer) and the Centralized Multi-Agent framework with Attention (CMAA). We compare these new models against state-of-the-art methods and benchmarks and develop ad-hoc interpretability tools to harness their insights. We then construct a dataset using traditional financial data and diverse ESG data from various providers. The NSiTransformer and CMAA are then applied to this new dataset, respectively for timeseries predictions at several prediction length and for fine-tuning of a timeseries predictor. We find that the inclusion of ESG ratings in the dataset, especially from various providers, improves the performance of the models. Through interpretability, we pinpoint which features of the dataset are contributing the most for a given prediction. We conclude that using interpretability of machine learning models is a valid approach to discover patterns that might escape traditional statistical analysis. We also conclude that ESG ratings are worth integrating in financial predictions and have the potential to increase performance. We compare this property to other slow-moving indicators that have been determined to be beneficial for financial predictions.