![Table 7: Individual and Institutional Market Order Imbalances After News Table 7 presents daily cross-sectional Fama-MacBeth (1973) regressions of individual (IndBSI) and institutional market order imbalances (InstBSI) from days t + b through t + e on news and control variables. The three columns show [b, e] intervals of [0,0], [1,1], and [2,5]. The table reports the timeseries average of the regression coefficients, imbalances on days [b, e], and R2 statistics. The coefficient on the firm’s abnormal return on day t (AbRetit) measures aggressive momentum or contrarian trading. Interaction terms with AbRetit allow momentum trading to depend on the two textual staleness measures (stale1it and stale2it), newswire messages (Msgit), earnings news (Earnit), and firm size (MktCapi,t-1). The complete set of independent variables is the same as in Table 4 and Table 5. See these tables for details. The coefficient units are basis points per standard deviation of the independent variables. Newey-West (1987) standard errors robust to heteroskedasticity and five days of autocorrelation appear in parentheses. The symbols **, *, and + denote statistical significance at the 1%, 5%, and 10% levels, respectively.](/figures/table-7-individual-and-institutional-market-order-imbalances-2ds2kfn3.webp)
All the News That’s Fit to Reprint:
Do Investors React to Stale Information?
October 2010
Paul C. Tetlock
*
Columbia University
Abstract
This paper tests whether stock market investors appropriately distinguish new and old
information about firms. I define the staleness of a news story as its textual similarity to the
previous ten stories about the same firm. I find that firms’ stock returns respond less to stale
news. Even so, a firm’s return on the day of stale news negatively predicts its return in the
following week. Individual investors trade more aggressively on news when news is stale. The
subsequent return reversal is significantly larger in stocks with above-average individual investor
trading activity. These results are consistent with the idea that individual investors overreact to
stale information, leading to temporary movements in firms’ stock prices.
*
Acknowledgements: I thank Columbia University, Yale University, and the University of Texas at Austin for their
research support. I am grateful to Brad Barber, Nick Barberis, John Campbell, James Choi, Martijn Cremers, Amy
Dittmar, Simon Gervais, Mark Grinblatt (AFA discussant), Terry Murray, Terry Odean (NBER discussant), Chris
Parsons, Matt Spiegel, Laura Starks (editor), Jeremy Stein, Philip Tetlock, Sheridan Titman, Heather Tookes, and
two anonymous referees for helpful comments. I thank seminar participants at the AFA, Boston College, Columbia,
MIT, NBER, Notre Dame, Rice, UC Davis, Wharton, and Yale for valuable feedback. I thank Tanya Balsky, Adam
Laiacano, and the Program for Financial Studies at Columbia for excellent research assistance. All views and
opinions expressed in this paper are my own and are not endorsed by the retail order data provider.
1
“People everywhere confuse what they read in the newspaper with news.” – A.J. Liebling
This paper tests whether stock market investors appropriately distinguish new and old
information about public firms. In an efficient market where firms’ stock prices rapidly
incorporate all value-relevant signals, new information becomes stale information almost
instantly. Based on theory alone, the impact of redundant information on asset prices is unclear.
The proliferation of news increases the speed and quantity of information dissemination, which
could enhance informational efficiency. On the other hand, some readers of a news story may not
realize the extent to which other market participants have already traded on similar past
information, leading them to overreact to stale information in news.
Based on this latter argument, I hypothesize that investor overreaction to financial news
increases with the staleness or redundancy of information. The central contribution of this paper
is to use an extensive database on public news events to test this hypothesis and explore the
mechanism behind any observed overreaction. I gauge market overreaction by the extent to
which a firm’s initial daily return around a news event negatively predicts its return in the week
after the event. The staleness of a news story is its textual similarity to the previous ten stories
about the same firm. I focus on cross-sectional variation in return reversals because there are
many possible explanations for on-average return reversals.
The sequence of news events for Equitable Cos, an insurance firm with a market
capitalization of $15 billion, in March of 1999 illustrates the methodology. At 5:14pm on
Monday, March 1st, which I define as (one hour into) trading day t – 1, Dow Jones (DJ) releases
a newswire story about Equitable Cos. The story describes an SEC filing in which Equitable
proposes changing its name to the empty placeholder name of “( )” until shareholders adopt
2
another name at an upcoming meeting. On Tuesday night, part of trading day t, the news appears
again in a very similar format when the DJ newswire pre-releases selected stories from the
Wednesday morning Wall Street Journal (WSJ).
1
The headlines and lead paragraphs in the
original Monday DJ newswire and the Tuesday WSJ story appear below:
DJ: Equitable Proposes Changing Its Name To (...)
The artist formerly known as Prince chose a glyph to represent his new identity. Now the
insurance company about to be formerly known as the Equitable has done him one better. Until it
finalizes its new moniker, it apparently wants to be known as, well ... nothing.
WSJ: Can Equitable Find Any Better Name Than ‘( )’?
The Artist, formerly known as Prince, chose a glyph to represent his new identity. Now
the insurance company that soon may be formerly known as Equitable Cos. has gone a step
further. Until it finalizes a new name, the company apparently must make do with “( ).”
A simple [0,1] measure of the similarity between two texts, proposed by Jaccard (1901),
is the number of unique words present in the intersection of the two texts divided by the number
of unique words present in the union of the two texts. One can compute an analogous similarity
measure for unique adjacent word pairings, called bigrams, rather than unique single words.
2
The 23 unique single words in the first excerpt above include: “Equitable,” “propose,” “artist,”
“finalize,” and “moniker.” The 22 unique single words in the second passage include
“Equitable,” “find,” “better,” and “artist.” There are 28 unique single words in the union of the
two paragraphs and 16 common pairings in their intersection, implying that the single-word
similarity of the two first paragraphs is 16/28 = 55.2%. The rest of these two stories are even
1
The actual WSJ story occurs on Wednesday morning, which is also part of day 0, but this story is not contained in
the Dow Jones newswire archive.
2
Before identifying unique words and bigrams, I exclude a standard list of 119 extremely common words such as
“into,” “so,” and “that”; 42 common numbers (0 through 9 and 1978 through 2009); and 27 terms that are ubiquitous
in financial news stories, such as “Dow Jones,” “New York,” and “newswire.” I also use a standard word stemming
algorithm to equate all similar forms of a word—e.g., “changing,” and “changed” are both derivatives of “change.”
3
more similar, so that their overall single-word story similarity is 79.6%, which ranks above the
99th percentile of similarity on March 2nd, 1999.
3
Equitable’s abnormal return on trading day t of its highly stale newswire story is -2.99%.
Interestingly, Equitable’s abnormal returns on trading days t + 1 and t + 2 are 1.76% and 1.93%,
completely reversing the initial decline in its stock price. Although the return reversal after
the -2.99% reaction to the highly stale story suggests that the reaction was excessive, it is
difficult to draw accurate inferences based on the ex post performance of a single firm.
To systematically measure how markets respond to two or more possibly related news
events, I examine all DJ newswire stories from the DJ archive for public US firms from
November 1996 to October 2008. The market reaction to news about a firm is the firm’s stock
return on a trading day when DJ includes the firm’s ticker code in the news header. I analyze
whether and why the day-t market responses to these news events are partially reversed during
days t + 1 to t + 5. I use both Fama-MacBeth (1973) regression methods and calendar time
portfolios to assess whether the extent of reversal depends on news staleness.
A news story’s staleness is its average textual similarity to the previous ten news stories.
This measure identifies news stories that contain a greater proportion of textual information that
overlaps with previously known facts. Consistent with the idea that such news stories contain
less new information, firms’ stock return and volume reactions on news days with high average
staleness are significantly smaller than their return and volume reactions on other news days. The
market’s initial reaction to news in the bottom staleness decile exceeds its reaction to news in the
top staleness decile by 413 (75) basis points on an equal-weighted (value-weighted) basis.
3
Using bigrams, such as “Equitable propose” and “propose change,” instead of using single-words as the basis for
computing similarity, the bigram similarity in the two excerpts would be 12/37 = 32.4% and the overall story
similarity would be 63.5%. In general, the single-word and bigram similarity measures exhibit correlations
exceeding 0.8, where the single-word score is higher by construction.
4
The first hypothesis tested here is that market reactions to news are better negative
predictors of future returns when news is stale. The evidence supports this view. Equal-weighed
portfolios formed on news staleness exhibit return reversals that differ by 26 basis points in the
week after news. Comparing this difference in reversal to the difference in initial reaction of 413-
basis-points, the market’s initial distinction between stale and new news seems insufficient in the
sense that it revises its initial view by another 26 / 413 = 6.3% in the following week. The value-
weighted revision is larger in percentage terms (19.8%) but smaller in raw magnitude (15 bps).
In predictive regressions of firms’ returns on days t + 1 to t + 5 that control for
alternative explanations, the regression coefficient on a firm’s day-t return decreases with the
staleness of news on day t. The ability of day-t news staleness to predict post-news return
reversal remains significant after controlling for weekly return reversals (Jegadeesh (1990);
Lehmann (1990)), volume-induced return reversals (e.g., Campbell, Grossman, and Wang
(1993); Lee and Swaminathan (2000); Llorente et al. (2002)), and other variables that predict
high-frequency returns and return reversals.
To examine the mechanism behind these return reversals, I focus on a subset of investors
who may confuse new and old information and, as a result, actively trade on stale information.
Based on abundant prior evidence in papers such as Odean (1999), Barber and Odean (2000),
Barber and Odean (2008), and Barber, Odean, and Zhu (2008) that individual traders exhibit
behavioral biases, one hypothesis is that individual investors are more likely to react to stale
information than institutional investors. To measure the presence of both types of investors, I use
a database of individual and institutional trading orders routed through a large market center
from 2003 to 2007. Aggressive trading in each groups is the imbalance in buy and sell market
orders for that group. I show that individual investors increase their tendencies to trade