Open Access

This doctoral thesis is concerned with two separate but intertwined topics in the field of financial econometrics: (i) the measurement and relevance of new sources of information on financial markets in the form of online investor sentiment and attention and (ii) nonlinearities in financial time series in the form of structural breaks. According to classical finance theory, competition among rational investors, often called arbitrageurs, leads to an equilibrium in which prices on capital markets equal the present value of expected future cash flows. Under this theoretical lens, the trading decisions of irrational investors have no significant impact on prices since their demands are offset by rational investors. However, the classical finance theory fails to fit the extreme levels of and changes in stock prices corresponding to events such as the Great Crash of 1929 or the Dot.com bubble of the 1990s, which are difficult to align with any rational explanation. Akin to the notion of "animal spirit" first coined by Keynes (1936), behavioral finance theory sets out to augment the classical model by explicitly taking into account two assumptions: Firstly, trading activities of investors are thought of to be partially influenced by subjective beliefs about investment risks and future cash flows, generally referred to as investor sentiment. Secondly, there are limits to arbitrage in the sense that betting against sentiment-driven investors is associated with higher risks and costs. Thus, inconsistent with predictions of the classical finance theory, arbitrageurs do not aggressively force prices to fundamentals. On this basis, irrational (collective) investor behavior has moved into the focus of modern finance theory and corresponding empirical applications. The widespread internet access and usage of social media platforms in recent years have led to new sources of information - and with them new sources and types of data that can be used by researchers and practitioners alike - pertaining to this collective investor behavior and corresponding financial market outcome: Short messages published on social media platforms such as Twitter or StockTwits on the one hand and online search queries on the other. The first part of this thesis makes use of such data in empirical financial applications, also from a high-frequency intraday perspective, in order to assess its impact on predictions of financial variables and to unravel new relationships. In general, it is reasonable to assume that many relationships in economics and finance are nonlinear. Thus, several kinds of nonlinearities can arise when considering financial markets and time series of financial variables that are not necessarily approximated well by simple linear models. Relating to the behavioral finance literature, the model of De Long et al. (1990) proposes that in the presence of sentiment-prone noise traders the price of a risky asset evolves as a nonlinear function of these noise traders' average bullishness (i.e., their mean misperception of the expected price) and its variance. Though being of a different philosophical nature than sentiment-induced noise trader theories, some other models of trade based on noninformational reasons, such as changes in risk aversion or liquidity needs, also involve nonlinear relations. The second part of the thesis focuses on one often overlooked kind of nonlinearity that entails potentially more severe implications, namely structural breaks in financial time series. Structural breaks, also referred to as change-points, in the data generating process underlying a given univariate time series do not only constitute a source of nonlinearity that can be modeled but also a more subtle source of nonstationarity. Given that endeavors of time series model building and prediction usually demand some stationarity assumption to be made, the latter poses a common problem in the analysis of univariate economic and financial time series. Matters are complicated by the fact that the exact number and timing of structural breaks are usually unknown ex-ante. Therefore, the consistent estimation of structural breaks, or change-points, has been studied extensively in the related literature. This thesis adds to the ongoing discussion by proposing a two-step model selection procedure for the detection and timing of change-points in structural break autoregressive models. A similar methodology is then used to investigate the effect of Box-Cox transforms on the estimation of structural breaks in realized volatility time series.