While we consider our findings to be robust in several dimensions, we feel that future research can improve our current understanding of the role that media factors play in financial markets. First, there are multiple ways we can improve on the way to measure sentiment. To elaborate on this, we continue by describing the features that an ideal, ‘comprehensive sentiment model’, should possess to be able to detect sentiment in a human-like manner. Finally, we discuss the key research directions we see related to finance literature.
7.3.1 Sentiment methodology
While our sentiment estimation methodology is a significant improvement from prior simplistic word count methodology, our method has several caveats. One of the key difficulties in phrase-level sentiment analysis is the lack of deeper contextual information. Since each phrase has to be interpreted in isolation, a human reader would often prefer to see more context than one sentence in order to draw any conclusions. For instance, although acquisition events are commonly described in quite positive tone (e.g. “The acquisition of Sampo Bank makes strategic sense for DB"), it is quite uncertain whether they will add or destroy value. Therefore, human analysts tend to be generally skeptical about the positiveness of such news, and would require considerable amount of background information before wanting to judge statement positive or negative. In addition to the lack of prior knowledge on the companies, the LPS algorithms are unable to distinguish between advertising-like positiveness following from a company's own statements vs. independent reviews about the company. Also some events, such as nominations of new executives, are conventionally described in overly positive manner instead of reflecting the actual facts. An interesting
problem is also the detection of roles or perspective in a sentence (i.e. from whose viewpoint do we interpret the sentence when multiple parties are involved). For example, a sentence may talk about a company getting money from another company after a legal case. Good news for someone may well be bad for the other.
Clearly, there is still a number of ways to improve the performance of the sentiment models reported in this paper. In addition to the enrichment of lexicon with weights for different concepts and events, an important direction for future research will be to examine how phrase-level models can be merged with content models. We recognize specifically the following development areas for future models:
Developing a sentiment methodology able to identify the topic of a text, and to capture the relevance of that text. For example a study by Antweiler and Frank (2006), which uses an algorithm to identify news stories by their topic rather than their tone, does find some return predictability. As a further step, researchers could test for the impact that qualitative texts dealing with company’s key products have on sentiment, possibly looking at media that does not even mention the company’s name
Creating a method that is able to assess the credibility of a text source in order to establish a proper weighing score for the text
Studying alternative ways of aggregating sentiment besides the use of negative fractions: for instance, Das (2010) disagreement sentiment. Also, different “shades” of negativity, e.g. bad news of events that have happened vs. increase in risk could have a different impact on the stock performance
Assembling dictionaries in different languages, and researching whether or not a sentiment constructed from them has any difference to sentiment constructed from source texts written in English
Constructing a methodology able to differentiate between texts dealing with historical information, present information, and forecasted estimates. For example O’Hare (2009) points out that traditional media reports most likely news relating to a stock’s past performance but only few statements about a company’s future. However analysis about future performance of a stock should be the part that investors consider most
7.3.2 A comprehensive sentiment model
To foster future development efforts, we finish our discussion on sentiment methodology by describing what a future model for detecting sentiment could potentially look like. This
comprehensive sentiment model would take all the media on a given day, analyze this information in the context of previously publicly known information, and give an estimate of
What news this information says about the company’s fundamentals and how should this affect the company’s valuation (information)
How this information may impact different investors and thereby the demand for the company’s stock (tone)
For this goal, we describe the parts of that a comprehensive model could consist of. In addition to illustrating the purpose and workings of each part of the model, we also give our early hypotheses on how this model could be implemented, and potential pitfalls for each part. The proposed model is described in Table 23.
Table 23: Proposed comprehensive sentiment model
Model part Purpose Example Implementation Pitfalls
Word detection
Detect polarized parts in text, possibly also with sentiment strength
Detects word “well” in a sentence “We believe that Samsung will not do well” and assigns labels to it, e.g. “reversal” and “positive”
Word lists for keywords of different categories Use of flawed word lists Pattern matching for polarized words Translate a pattern of detected words to a sentiment, possibly with a score for strength
Based on training data, assigns label “negative to tags “reversal + positive”
Quasi compositional sequencing (SVM) with pattern compression
Training data that uses more knowledge than is available at this stage of the model Credibility filtering Assess the credibility of a sentence
Detects that the praising document is written by the company’s sales
department and reduces the strongly positive sentiment
Author recognition with entity detection, credibility assessment (entity-database from a Wikipedia-based ontology202 + database of historical source sentiments) Entity database lacking key entities, difficulty of assigning a credibility estimate to listed entities Topic detection
Detect the key topic(s) in a sentence
Detects from “Samsung’s new Galaxy is selling moderately well” that it relates to galaxy sales / smart phone sales
Topic recognition with a Wikipedia-based ontology Hierarchy of topics, colloquial language Impact assessment
Detect how much further impact will this media item have on the topic have when added on top of the existing information on the market
Detects whether a piece of information has been discussed before (Is it novel news/analysis?), confirms a previously uncertain information (Is this source more
credible?), or if it reaches now a new audience (Did this information previously reach only people who read the specialized industry magazine?)
Database on information by topic for comparing novelty,
impact factors for different sources based on reach, prestige etc.
Difficulty of novelty
assessment, lack of impact factor data for news
Topic relevancy filtering Assign a relevancy score per topic, highlighting the most important topics
Ranks e.g. the topic “mobile phones” higher than “laptops” for Nokia
A layered ontology of concepts: importance of concepts for all
companies, for different industries and for individual companies, based on metrics such as word-of-mouth in media, company’s own releases, and analyst reports
Difficulty of ranking concepts, hype of concepts vs. reality
A sentiment model of this kind would be significantly closer to mimicking the key stages of an analyst’s thoughts in assessing the impact of news on a company and could potentially form similar conclusions as financial analysts who study a company regularly. As for the moment, most studies have implemented only the first part of this model, which is likely to yield sentiment scores with significant biases. Though the proposed model includes some significant challenges in implementing, we find that models that go towards this direction could significantly improve the detected sentiment and make it more relevant in relation to stock performance.
7.3.3 Data and specifications
As with all studies, our study faced limitations in terms of data accessibility. Also, in light of our findings, there are areas that future study specifications should consider when conducting their research. We feel that the key areas of interest in the aforementioned areas are:
Analyzing sentiment using social media text sources, such as, i.e., Twitter
Studying the impact media factors have in Fringe markets (i.e. smaller countries such as Finland), and whether or not the impact differs from the findings of extant literature Using intra-day data when studying the impact media factors have on financial metrics Constructing trading strategies leveraging aggregate market news volume, and sentiment volatility forecasts, to measure economic feasibility of such trading schemes Studying the impact high firm specific news has on volatility, and whether or not that
translate into a rise in cost of capital as time passes on
To conclude, we have casted some additional light on the big picture of how media factors and sentiment link to each other. Future research could complement these findings in particular by further improving on the research, and by tapping to further data sources.