which that language is used. For example, the mention of a ‘‘dividend cut’’ may screen
as a negative using dictionary-based algorithms, though in previous research we have
found that during the financial crisis, such news was often taken positively by the
market if the cash was saved to pay down debt. To take into account the dynamic
nature of markets, we have since investigated alternative approaches including
pattern recognition algorithms and supervised learning techniques to discriminate
between positive and negative news sentiment. Other classification approaches used
by data vendors include training datasets on the results of financial analysts manually
tagging equity stories and market response classifiers trained on the markets’
immediate response to news.
.Media coverage An alternative method is to define event days according to
unusually high media coverage. An interesting approach along these lines was
adopted by Lumsdaine (2009) who gathered Bloomberg data on the count of investor
readership of news stories for the bank sector during the Credit Crisis. A simpler
approach could be to monitor abnormal trading volumes. The argument is that, when
investors pay greater attention to stocks, they are more likely to trade, thus generating
higher volumes. The attractiveness of using an abnormal trading volume factor is that
historical data are readily available across the global regions. However, the pitfalls of
using this factor to proxy attention are that it captures investors’ actions rather than
identifying the motivations to trade and trading activity around corporate actions
(such as share buybacks) or dividend announcements may cloud the relevance of
using trading volume to capture investor attention effects. Moreover, with the grow-
ing importance of liquidity dark pools, traditional databases are missing significant
information on trading volumes which may limit the efficacy of this factor.
.Market-based measures Finally, market-based measures can be used to define the
importance of news, measuring abnormal returns over a window around news
announcements and using a threshold to determine whether news is positive, neutral,
or negative.
Ultimately, there are pros and cons with each of these approaches. Market-based
measures take into account the markets’ expectations of the news flow prior to its
announcement so that returns measure the surprise element of the news announce-
ment. The drawback to classifying news in this way is that it can only be measured ex
post based on observed returns so strategies cannot be implemented immediately. On
the other hand, text-based classifying algorithms offer a more timely approach,
analysing sentiment without relying on the need for returns. The disadvantage is that
it becomes difficult to isolate the ‘‘surprise’’ element from what is ‘‘known’’ news
within an announcement.
8.3 Understanding news flow datasets
Having cleaned the dataset, we were left with a sample of 90,000 news announcements
from January 2001 onwards for companies within the S&P Large-Cap Europe universe
(around 450 stocks). Figure 8.3 plots the time-series of news categories for a large-cap
European universe. It shows that the majority of news items relate to company earnings
news or guidance with the remainder equally split across other news categories. Figure
8.4 shows that news releases are not equally distributed across different calendar
The impact of news flow on asset returns: An empirical study 217