The easily discernible pileup of observations at the zero thresh-
old for ∆EPS is confirmed by the τ-statistic of 6.61.^26 The value of
6.61 is the largest for all points in the neighborhood as well as
being very significant. These findings are also confirmed with the
subsamples of I/B/E/S and Q-Prime (unreported).
2.“Meet Analysts’ Expectations.”Figure 18.6 plots the empirical dis-
tribution of the forecast error, FERR (equal to EPS minus the ana-
lysts’ consensus EPS forecast) in 1-penny bins in a range around
zero, using quarterly observations over the 1974 to 1996 period.
Consistent with the notion that “making the forecast” is an im-
portant threshold for managers, the distribution of FERR drops
sharply below zero: we observe a smaller mass to the left of zero
compared to the right. (Note that in the histogram, the bin starting
with zero represents observations that are exactly zero.)
There is an extra pileup of observations at zero, although this is
hard to see for a distribution like FERR that is centered on zero it-
self. The pileup is confirmed by the τ-statistic of 5.63, which is very
significant.^27 This value exceeds the values of τfor all the neighbor-
ing points, none of which exceed 2.0 in absolute value (unreported).
Parallel to figure 18.3 (which, like FERR, had the latent distribu-
tion centered at the threshold), we show the shortfall in densityEARNINGS MANAGEMENT 651 Figure 18.6. Histogram of forecast error for earnings per share: exploring the
threshold of meeting analysts’ expectations.
(^26) In this case the likely threshold is not at the peak of the distribution although its neigh-
borhood includes the peak; see elaboration A1 discussed in the appendix.
(^27) In this case, we compute τfor the case in which the likely threshold is at the peak of the
distribution (see elaboration A2 in the appendix).