If we winsorize the return at an absolute value of 20 percent, the bias is much reduced, in absolute
terms, as shown in Panel B. The biggest positive and negative reported returns are replaced by exactly
offsetting numbers (+0.2 and –0.2, here), so the net effect in those cases is that, on average, one misses the
expected return per period. If such large errors happens rather rarely, the effect on a longer-run time series
average is quite small. But occasionally only the ‘up’ observation gets winsorized while the corresponding
‘down’ is not, and in those cases the improvement is less radical.