From German News to European Market Trends: ESG Sentiment Modeling with Machine Learning

This thesis investigates the impact of environmental, social, and governance (ESG) sentiment expressed in German newspapers on European equity index returns. It begins by outlining the historical development of ESG and impact investing and the evolution of textual sentiment analysis in finance, highlighting the transition from dictionary-based approaches to transformer-based language models. The empirical analysis relies on a self-constructed dataset of ESG-related articles retrieved from the GDELT database over the period 13/06/2022–13/06/2025. Sentiment is extracted using German FinBERT, a transformer-based model tailored to financial text, enabling context-aware measurement of article polarity. Daily sentiment is aligned with equity market data and examined along four dimensions: timing, geographical spillovers, persistence, and predictability. The results show that ESG sentiment has a statistically and economically meaningful impact on short-horizon returns, particularly when measured before market opening. Effects extend beyond German indices to broader European benchmarks, indicating cross-market spillovers. Multi-horizon regressions reveal that sentiment-driven return effects are concentrated within a few days and decay rapidly. Finally, both linear and non-linear classification models confirm that ESG sentiment provides out-of-sample predictive power for return direction. Overall, the findings suggest that transformer-based ESG sentiment captures relevant information that influences short-term market dynamics.