Within-industry timing of earnings warnings: do managers herd?

Our sample selection is consistent with the warning studies by Kasznik and Lev (1995), Soffer et al. (2000), and Atiase et al. (2006). These studies identify warnings as unfavorable earnings disclosures issued in the 60-day window before earnings announcement. Because earnings are typically announced 25–30 days after the fiscal quarter end, the starting point from when these studies collect warnings is close to the beginning of the third fiscal month.

On average, seven warnings are issued for an industry-quarter. If a uniform distribution is assumed, the average interval between successive warnings should be about 9 days.

The persistence of an earnings shortfall is likely to be unrelated to whether it is caused by internal or external factors. For example, an economic recession and a change in the management could both have long-lasting effects.

The accounting literature has a stream of information transfer studies. A firm’s earnings forecast conveys information that affects nonforecasting industry peers’ returns (Baginski 1987; Han et al. 1989). Early earnings announcements affect the stock returns of late announcers in the industry (Foster 1981; Clinch and Sinclair 1987; Han and Wild 1990; Freeman and Tse 1992; Lang and Lundholm 1996b). The major difference between managerial herding and information transfer is that the former focuses on managers’ decisions in a multi-firm setting, whereas the latter focuses on investors’ inferences in a multi-firm setting.

The exceptions are Chamley and Gale (1994) and Zhang (1997).

Evaluators use the “covariation” of behaviors from several individuals and sources to identify possible causes.

Our conjecture about managers’ behavior does not necessarily imply that boards of directors are misled. Managers and their evaluators may be linked by interlocking directorships (Brick et al. 2006) and social ties (Hwang and Kim 2008). Issuing a warning in a cluster would enable the evaluators to justify attributing an unwarranted proportion of the shortfall to external factors even though the evaluators can see through the manager’s timing strategy.

The term “duration model” is used in engineering for models that examine the timing of mechanical failures (Greene 2003, 790–791). Biomedical research has a longer history of using this method and refers to it as “survival analysis.” Duration models have been applied in economics to examine the length of unemployment spells and in finance and accounting to study the staging of venture capital (Gompers 1995), earnings management (Beatty et al. 2002), and the influence of investment banking ties on the speed of changes in analysts’ recommendations (O’Brien et al. 2005).

Right-censoring means that the measurement for duration does not end naturally but ends early because of either researchers’ constraints or an outside force. In our setting, warning duration is right-censored at the earnings announcement date (strictly speaking, 3 days before the earnings announcement date).

We assume that at the beginning of the third fiscal month, a firm has decided whether to warn and the only remaining decision is when to warn.

The model is \( h(t_{i} ) = h_{0} (t_{i} )\;\exp {\kern 1pt} (X_{i} \beta ) \). Suppose a set of sample firms in one industry issues warnings on K different dates (\( T_{1} ,T_{2} ,T_{3} , \ldots ,T_{k} \)) for a quarter and that there is at most one warning each day. Let \( R_{i} \) be the set of firms that have not yet issued a warning by day\( T_{i} \), where \( i = 1,2, \ldots ,k \). The probability that a firm warns on day \( T_{i} \) is \( {{\exp (X_{i} \beta )} \mathord{\left/ {\vphantom {{\exp (X_{i} \beta )} {\sum\nolimits_{{j \in R_{i} }} {\exp (X_{j} \beta )} }}} \right. \kern-\nulldelimiterspace} {\sum\nolimits_{{j \in R_{i} }} {\exp (X_{j} \beta )} }} \). Then, \( \;LnL(\beta |data) = \sum\limits_{i = 1}^{k} {\left[ {X_{i} \beta - \ln \left( {\sum\nolimits_{{j \in R_{i} }} {\exp (X_{j} \beta )} } \right)} \right]} \)is the partial log-likelihood of observing the data, and the Cox estimator for β maximizes this function (Cox 1975).

We obtain similar results when PeerWarn (t) is measured in a 3-, 7-, or 10-day interval.

Brown et al. (2006) use the cumulative proportion of industry peers that have disclosed by day t as the primary variable of interest. We use the number of recent warnings by peer firms in the preceding 5 days as our primary explanatory variable because the measure best reflects our expectation that managers seek to reduce the blame for poor performance by issuing warnings at about the same time as other firms with earnings disappointments.

Within-sample rankings would be coarse for industry quarters that have only a few warnings.

First Call uses analysts’ earnings expectations as the benchmark to code managers’ earnings guidance as positive, negative, or in-line guidance. Although the data manual does not specify how guidance with a range estimate is coded, we infer based on the comparison of the estimates and analysts’ consensus that CIG codes guidance as “negative” if the upper estimate is lower than analyst consensus, “positive” if the lower estimate is higher than analyst consensus, and “in-line” if analyst consensus is between the lower and upper estimates.

The period begins in 1996 because there are only a few observations (328 events) before 1996 and because the Private Securities Litigation Reform Act of 1995 substantially expanded the safe-harbor protection to firms for disclosing forward-looking information.

The warnings are issued by 1,817 unique firms. Among these firms, 919 (50.6%), 465 (25.6%), 225 (12.4%), 107 (5.9%), and 101 (5.5%) issued warnings once, twice, three, four, and five or more times, respectively.

We also estimate an OLS model with the logarithm of the actual duration as the dependent variable and PeerWarn(Actual) instead of PeerWarn(t) as the explanatory variable. The coefficient on PeerWarn(Actual) is −0.039 (t = −6.22), consistent with the Cox model results, and the model R ² is 21.2 % (recall that duration is inversely related to hazard rate).

We also measure bad news as the absolute price-deflated difference between the firm’s EPS estimate (or the midpoint for a range estimate) and the most recent consensus before the warning. We substitute the realized earnings for company estimate when the latter is not in point or range form. The coefficient on BadNews is significantly negative (coefficient = −0.016, z-statistic = −2.18). Of our sample firms, 28.6, 64.6, and 6.8% provide point, range, and other-form warnings, respectively.

We construct a variable that counts the number of warnings in the 5-day window before the sample’s warning issued by firms in industries other than the sample firm’s. We find no association between this variable and warning timeliness, whereas PeerWarn(t) remains significantly positively associated with warning timeliness. This result is consistent with our message of within-industry herding.

Some industry-dummy coefficients have different signs than those in Table 5. As is typical for panel data, our interest is in controlling for the fixed industry effects, not in making inferences from their coefficients.

The 25th, 50th, 75th, and 95th percentiles of the distribution of events per industry quarter for the bad-news sample are 3, 5, 8, and 20, whereas those for the good-news sample are 3, 4, 5, and 12. To match with the good-news sample distribution, we exclude from the bad-news sample the industry quarters with 12 or more warnings. The reduced bad-news sample contains 1,881 events from 449 industry quarters and the 25th, 50th, and 75th percentiles of the distribution of the number of events per industry quarter are comparable with those of the good-news sample. We then randomly choose 115 industry quarters to mimic the sample size of good news, leading to 500 warning followers.

We speculate that the reduced z statistic is due to less variation in PeerWarn(t) after the industry quarters with more than 12 events are excluded.

We find few good-news alerts in warning clusters or within 5 days after the last warning in a cluster (unreported), consistent with bad-news herding.

Our approach treats each industry as one entity and prevents firm-level ROA outliers from driving the industry-level ROA.

On the other hand, later warnings have larger earnings shortfalls than early ones (Row 2), so the total market reaction to later firms should be larger. To examine the aggregate return effect, we collect market-adjusted returns from the day after the analyst consensus for the leader to the sample firm’s own warning date (included). After controlling for earnings surprise, we find no difference in stock returns across firms of different warning sequence.

Although our primary goal is to document the existence of warning herding, we conduct preliminary tests to determine whether herding managers have stronger career concerns than nonherding managers. We use the CEO’s age as a proxy for career concerns because the younger a manager, the longer his future earnings stream, and thus the greater his concern about others’ perceptions of his ability (Gibbons and Murphy 1992; Garen 1994; Bryan et al. 2000). We obtain CEO age data for 2,977 of the 3,016 warning followers (98.7%). We retrieve data for 1,430 observations from Compustat’s executive compensation database and hand-collect data for the remaining 1,547 observations from DEF 14A proxy statements. This sample is comprised of 1,699 herders (those in the middle or at the tail of a warning cluster) and 1,278 nonherders. We find that the average CEO age is 53.6 for the herding group and 54.6 for the non herding group. The CEOs are younger in herding firms than in nonherding firms in both a t-test (t = 3.78) and a nonparametric Wilcoxon test (z = 3.70). This result is consistent with career concerns motivating warning herding.