4.1. CEO overconfidence
We report the results of our regression estimation for Eq. (1) in Table 3, using ERR as the dependent variable. They
support for our overconfidence hypothesis (H1). The coefficient associated with STREAK is positive, which indicates that
managers who have experienced a series of accurate predictions subsequently issue less accurate forecasts. This effect is
statistically significant. The z-statistics (corrected for heteroskedasticity and the simultaneous clustering of observations by CEO and year) is 2.61. The effect is also economically significant. Increasing STREAK by one standard deviation increases ERR
by 14% of its median value.17 The R2 is 45.04%, which is explained in part by the inclusion of CEO fixed effects.18 Neither the
addition of fiscal quarter indicator variables nor a change in the magnitude of special items in the reported earnings (scaled by
total assets) affects our conclusions. The inclusion of yearly indicator variables materially increases the significance of the tests
(in this case, the z-statistic equals 4.56). Dropping all of the control variables (except for CEO fixed effects) and re-estimating
the regressions has no effect on our conclusions, which suggests that our results are not spuriously created by the inclusion of
irrelevant control variables (the results of these different robustness checks are untabulated).