Dynamics of investor attention on the social web

Abstract

Leveraging statistical analysis of "Big Data" relevant to real investors, this dissertation investigates micro-structures, temporal dynamics, and market impacts of investors' selective attention. We construct a novel and direct measure of modern investor attention, intertweet time, based on real-time and asset-related tweeting (microblogging) activities on the social web. We study a hierarchy of complex systems, ranging from individual investor's cognitive processes at the microscopic level to economic outcomes at macroscopic scale. The contribution of this dissertation is composed of three parts, each of which is summarized as follows.

Contribution I investigates mechanisms of cognitive control in individual investor's temporal selective attention. We develop formalisms of "cognitive niches", i.e., interplays between heuristics from adaptive cognitive control, to account for the selectivity of investor attention. Utilization of these cognitive niches is validated by empirical observations of investors' tweeting activities on assets. Such selective mechanisms are further shown to be contextual, depending on types of assets, investing experiences, and investing approaches. Embedded in a highly connected social environment, investor attention is found to employ the "social proof" heuristic, and the drawing power of the crowd in directing investor attention is significant and exceeds that of salient exogenous stimuli, especially when uncertainty in the financial market is high.

Contribution II characterizes the dynamical system of collective investor attention on the social web. We identify stylized facts of collective cognition in terms of fluctuation and memory persistence. Temporal fingerprints left by collective investor attention share several common properties with other complex systems with strong heterogeneity and interactions, such as clustering and memory persistence. In spite of scale-invariant fluctuations and long-range correlations identified from empirical tweeting activities, universality across assets was less supported than multi-scaling behaviors. To explicitly model the feedback mechanisms in collective investor attention, we propose a stochastic branching process as a coarse-grained generative model, which is shown to be an accurate representation of investors' tweeting behaviors on assets, especially during busy trading hours. Such results not only highlight significant endogeneity, or self-reflexivity, within the system of collective investor attention, but also provide more quantitative and real-time measurements of investor attention on the social web.

The World Wide Web has been revolutionizing how investors produce and consume information while participating in financial markets. Both the amount of information and the speed it flows around have achieved unprecedented magnitudes. The preeminent change is the growth of investor communities on the social web, which give rise to multidimensional information channels in real time. In achieving information processing so as to make investment decisions, what is immediately impacted is investor attention. Like other valuable resources in the economy, investor attention is limited. Therefore, it is crucial to understand how investors allocate their attention resources and the corresponding impacts for the financial markets.

Contribution III quantifies interactions between the dynamics of investor attention on the social web and price movements in the financial market. First, we show that these two systems are significantly correlated at a variety of timescales, with investors' tweeting activities carving out an attentional market. At microscopic timescales, we found feedback relationships between investors' tweeting activity and magnitudes of price movements, suggesting behavioral causes for the "volatility clustering" phenomenon. Furthermore, we demonstrate the disentangling of distinct patterns of volatilities with regard to both magnitudes and relaxation speed by using the nature of cognitive control to differentiate investor attention. At intermediate timescales, we identify bidirectional causal relationships between collective investor attention on the social web and trading activities on the market, including volatilities, returns, and trading volumes. Recognizing the different nature of investor attention allows us to observe the varying strengths of lead-lag relationships. A robustness check demonstrates that, as a social tape, dynamics of investor attention on the social web has its own information content, which has not been accounted for by known behavioral biases.