I can see both sides of this argument. The problem with allowing modifications is that this can be abused. If the predictor is allowed to continually extend or bend the window he can get away with all sorts of things, and it's hard to quantify how the odds have shifted as a result of this latitude.

The problem is inherent when prediction evaluation is done in what is essentially an adversarial manner of racking up hits and misses, instead of open presentation of data, in which case the reasons for needing to alter the prediction can also be evaluated, or at least discussed.