Wydobywanie wydźwięku
[Sentiment extraction]

Wolfgang Haerdle (Ladislaus von Bortkiewicz Chair of Statistic and C.A.S.E. 
- Center for Applied Statistics & Economics, Humboldt-Universität zu Berlin)


Financial analysts, portfolio managers, corporate managers and retail 
investors have been portrayed as being particularly afflicted by herding 
instincts.  Such sentiment effects are possibly most readily visible in 
finance, where volatility is ubiquitous and asset price movements of 5-10% in 
a day are not uncommon.  A prerequisite to assessing the impact of sentiments 
on decisions of individuals within interconnected groups is their unbiased 
measurement, as any research on the influence of subjective criteria on 
decisions in the large requires the means to measure the basis of these 
criteria in the small.  We analyze the outcome of such methods when applied 
to large text corpora representing a multitude of different agents when 
applied independently prior to a final aggregation stop, in a framework for 
joint inference, or in a hierarchical manner.  To enable such studies, we 
also will pursue research regarding the systematic gathering of appropriate 
texts from the web by means of focused web crawling, a technique which allows 
to quickly amass very large text corpora (up to several terabytes of HTML 
pages) that are specific – to a certain extent – for a given domain of 
study.  A particular problem when it comes to harvesting web content is that 
of the required currency of the data; here, retrospective studies require 
completely different approaches than those targeting real-time decision 
support which must be based on the most up-to-date texts available.  Our 
basic tool for such analysis will be a graph model of systems: Information 
flows within a group can be modeled as a multi-digraph, where nodes represent 
agents and arcs represent (directed) flows. As communication is dynamic, such 
models must also consider time, for instance in the form of time-labeled arcs 
or by using sequences of networks (dynamic networks), where each single 
network represents an aggregated snapshot of the flows in a given window of 
time. Our project will research novel methods to model herding using graphs, 
to build such networks for concrete scenarios, and to analyze such graphs to 
quantify different effects of herding. We will use article texts from the 
NASDAQ platform and subject them to advanced text mining to identify key 
phrases and to assess the sentiment associated to such phrases. 

This talk will study the following research questions: 
Q1. How can textual information be used to assess financial market 
developments in real time, e.g. the evolution of market volatility? 
Q2. Which network structure is capable of depicting the interplay among 
analysts and the stocks they recommend?
Q3. Can communities, their interconnectedness and their leaders be identified 
by the network constructed by analysts and recommended stocks?