Table 1 illustrates the upward bias from errors, as well as the effects of filtering and winsorizing. In
the example we show a sequence of three true values Pt yielding a positive mean return (4.9%). We then
administer various errors to the middle price: two times a shift in the decimal point, and six examples
with errors equaling ±10, 20 and 50 percent. First consider, for all these cases, the raw returns without
any filtering, as shown in Panel A. Regardless of the sign and size of e, as illustrated in the simple examples
of Table 1, the mean is always up after adding noise to a price, and the more so the larger the noise. The
simple and inescapable fact behind this is of course that, if a price drop and a price rise are similar-sized
in terms of dollars, this is not true for the percentages: measured as a return, the spurious price drop that
the error generates is smaller than the corresponding spurious rise, which always is from a lower base.4