Investor sentiment and statistical moments of the return distribution in the German stock market. A three stage empirical analysis.

Abstract

This dissertation contributes to an increasing body of literature through a holistic perspective in research and model development to investigate the relationships between investor sentiment and the lower- and higher-order statistics of the return distribution in the German stock market utilizing the market-wide CDAX stock index as an exemplary sample.

Since empirical studies on investor sentiment are conducted mainly with USmarket-based data, comparatively few academic contributions are made to the German investor sentiment literature. Moreover, previous findings for other countries cannot necessarily be generalized to Germany, especially as Germany appears to be mainly influenced by global trends and investor sentiments owing to its high dependence on foreign trade.

Consequently, the empirical evidence for Germany in this research domain, which includes both cross-sectional and longitudinal perspectives, is sparse. As various approaches exist to measure and assess the links between investor sentiment and capital market movements, a proprietarily defined investor sentiment categorization system is established, in which each investor sentiment indicator is assigned. This dissertation's underlying investor sentiment sample consists of all three categories of the dedicated categorization system for investor sentiment indicators and covers up to 20 years to 2021.

With regard to the thesis structure, a comprehensive overview of the literature and current research on market efficiency and investor sentiment is initially elaborated before the applied methodology and the evaluation results are analyzed. Of particular note is the three-stage empirical analysis conducted in this thesis: First, a principal component analysis-based investor sentiment risk factor is established to improve model performance in traditional cross-sectional multifactor models as measured by the corrected coefficient of determination and additional metrics.

Second, the application of Long Short-Term Memory (LSTM) artificial recurrent neural network architecture models to account for time-varying investor sentiment risk premia explaining and predicting the return distribution's lower- and higherorder statistics leads to notable findings.

A performant model for the German stock market results from fitting a deep neural network fed with 73 sentiment indicators without dimension reduction and performing out-of-sample tests.

Third, an insightful exploratory Twitter study of social investor sentiment in the German stock market in times of COVID-19-induced market turmoil is elaborated. The study investigates the impact of incorporating unstructured data into investor sentiment analysis to improve discriminatory power and predictive accuracy and employs elaborate processing techniques. The exploratory study is based on a unique hand-curated dataset of almost two million tweets on the German stock market, exclusively collected for this study. In this context, the importance of investor sentiment in social media for volatility in the German stock market is investigated and highlighted.

As a result, all three empirical studies address many vital matters, although new challenges worthy of investigation are as well raised and discussed in the final part of the thesis.