LetNdenote the universe of all news records from the RavenPack dataset. Fix a
companyCthat is mentioned within some news record fromNwithECðNÞandDCðNÞ
representing the Event Sentiment Score and Event Novelty Score of companyCfor
recordN2N, respectively.
Definition 5.1.LetUbe the universe of companies andCbe a company such that
C2U. Letpbe a time period denoting a certain number of days. LetPNbe the records
of all stories published withinpdays before publication of news recordNup to and
includingNsuch that 8 Ni 2 PN. For 8 Nithere exist someCsuch thatECðNÞ6¼;and
DCðNÞ¼100. In other words, every record inPNhas an Event Sentiment Score for
someC2Uand an Event Novelty Score of 100. Finally, letm¼jPNj. The trailing
sentiment index,IUðN;EÞ, forUis the quantity
IUðN;EÞ¼1
mXmi¼ 1ECðNiÞ: ð 5 : 1 ÞRemark 5.1.Theuniverse of companiesis meant here to represent the constituents of a
broader equity index (i.e., the S&P 500 or Russell 1000).
Considering the constituents of the S&P 500, I construct a US market-level sentiment
index applying a 90-day trailing window (P¼90). One of the advantages of aggregating
over such a period is that I capture an entire ‘‘quarterly season’’ in each trailing window,
thereby ensuring that similar news flow characteristics are represented. Equity news flow
is very much characterized by seasonality, where a quarterly pattern is evident and likely
caused by the repeated earnings reporting season (Hafez, 2009b). Figure 5.1 depicts the
US Market-Level Sentiment Index vs. S&P 500 cumulative index log returns covering
the period March 2005 through December 2009.
5.2.3 Strategy and empirical results
Under the assumption that market returns are likely to move in the same direction as
market sentiment, I base my trading decision on the sentiment index delta, (^4) t. Focusing
on the index delta will capture the sentiment of the most recent period (i.e., 1 month),
but also include the sentiment change from the previous period which may have similar
characteristics (i.e., as captured by the ‘‘same’’ month in the earnings season cycle).
Definition 5.2.LetItbe the trailingsentiment indexvalue at timet, then thesentiment
index delta, (^4) t, at timetis the quantity
(^4) t¼ItIt 1 : ð 5 : 2 Þ
In order to construct a simple news-based strategy, whenever (^4) t>0 I take a long
position in the S&P 500 in the following period. Likewise, when (^4) t<0 I take a short
position. More specifically, I decide at the end of each month the direction to take in the
S&P 500 in the following month. S&P 500 index returns are calculated as monthlyclose-
to-closelog returns.
In Figure 5.2, I include the cumulative return of a trading strategy based on the US
Market Sentiment Index with and without applying an Event Novelty Score filter.
Furthermore, I include the 1-month price momentum strategy return for benchmark
purposes.
How news events impact market sentiment 133