[Paper Review] ESG2Risk: A Deep Learning Framework from ESG News to Stock Volatility Prediction

Tian Guo, Nicolas Jamet+3 more|arXiv (Cornell University)|May 5, 2020

Market Dynamics and Volatility4 citations

TL;DR

This paper proposes ESG2Risk, a deep learning framework that leverages transformer-based text embeddings from ESG news to predict stock volatility, outperforming sentiment-only models. It demonstrates superior forecasting accuracy and identifies high-volatility, low-return stocks, validating ESG news as a timely predictor of market risk across U.S. and European equity markets.

ABSTRACT

Incorporating environmental, social, and governance (ESG) considerations into systematic investments has drawn numerous attention recently. In this paper, we focus on the ESG events in financial news flow and exploring the predictive power of ESG related financial news on stock volatility. In particular, we develop a pipeline of ESG news extraction, news representations, and Bayesian inference of deep learning models. Experimental evaluation on real data and different markets demonstrates the superior predicting performance as well as the relation of high volatility prediction to stocks with potential high risk and low return. It also shows the prospect of the proposed pipeline as a flexible predicting framework for various textual data and target variables.

Motivation & Objective

• To investigate whether ESG-related news flow can predict future stock volatility more effectively than traditional structured ESG data.
• To develop a scalable NLP pipeline that transforms unstructured ESG news into numerical representations for quantitative modeling.
• To evaluate the predictive power of ESG news on realized volatility and its implications for portfolio risk management.
• To demonstrate that ESG2Risk can identify stocks with high risk and low return potential, improving investment strategy selection.

Proposed method

• The framework extracts ESG news using a domain-expert-defined ESG vocabulary and entity linking to associate news with specific companies.
• Textual ESG news is embedded into dense vector representations using a pre-trained transformer-based language model, preserving semantic and sentiment information.
• A Bayesian neural network with stochastic gradient Markov Chain Monte Carlo (SG-MCMC) inference is used to estimate model uncertainty and improve robustness.
• Predictive volatility is computed via Monte Carlo sampling from the posterior distribution of model parameters, yielding probabilistic forecasts.
• The pipeline is evaluated independently on MSCI US and All Cap EU equity markets using out-of-sample testing data from 2015–2019.
• Model performance is measured using RMSE and MAE on 1- and 2-week forward volatility predictions, with portfolio backtesting across volatility prediction quintiles.

Experimental results

Research questions

• RQ1 Can ESG news flow serve as a timely and predictive signal for future stock volatility?
• RQ2 How does a transformer-based text embedding model compare to sentiment-only models in predicting equity volatility?
• RQ3 Does incorporating ESG news improve the identification of high-risk, low-return stocks in equity portfolios?
• RQ4 To what extent can the ESG2Risk framework generalize across different equity markets and time horizons?

Key findings

• On both MSCI US and All Cap EU markets, ESG2Risk achieved significantly lower RMSE (0.289 and 0.281) and MAE (0.249 and 0.241) than the sentiment-only baseline (RMSE: 0.663 and 0.630; MAE: 0.542 and 0.515) for 1-week forward volatility prediction.
• The model’s predicted volatility was strongly correlated with actual realized volatility, particularly for stocks in the highest predicted risk quintile (Qt 4), which exhibited significantly higher realized volatility.
• Portfolios constructed based on high predicted volatility showed significantly lower average returns than lower-risk quintile portfolios, confirming the predictive power for risk and return characteristics.
• The Bayesian inference approach enabled stable and robust predictions, with Monte Carlo sampling providing reliable uncertainty estimates for volatility forecasts.
• The framework successfully identified stocks with high risk and low return potential, validating its utility in systematic ESG investing and risk-aware portfolio construction.
• The ESG2Risk pipeline generalizes well across geographies and time horizons, demonstrating consistent performance on 1- and 2-week prediction tasks.