Actually, a 14 x 14 grid of random numbers (-1,0, and 1) could sometimes tend to match the spatial pattern of reality, and in an extremely rare instance could equal it exactly. 14 is a smaller number than over 200, so getting some very good scores and some very bad scores would be more likely than when using over 200 random numbers. I think in your last question you are asking what would happen if our "random" forecasts were taken from actual observations of other years, where all of the other years had equal chances of being assigned as the observations for the year in question. If there were enough years in the record so that very, very many possibilities were possible, the skill score should come out to be extremely close to zero. But with only 20 to 30 years in the record more luck becomes possible, and the baseline score could come out noticeably positive or negative by chance. Using only other observed years' observations limits the number of possibilities quite a bit, while using computer-generated random numbers make is more fair and less vulnerable to good or bad luck in sampling. Also, in a smaller sample of years it would matter whether the year in question is allowed to be selected, or not (it would result in a perfect skill score). The bottom line here is that the method used should be the one that ensures to the greatest degree that the assignments are truly random.