sec 18.4 (donkey) small suggestion (sign change in definition)

[murphyk]

A small thing, but I find it confusing that error = (y-pred)/pred, instead of (pred-y)/pred, since one might expect an over estimate (which we want to penalized more) to be a positive error, not a negative error.