If rating were predictive of game outcomes, then the dotted line would
intersect the segments on the figure. With very few exceptions, the
confidence intervals computed from the observed data underestimate the
theoretical winning expectancy. Thus, lower-rated players are scoring
better than predicted by the ratings and the model, and that this
behavior is consistent across all rating differences.
Based on the poor fit to the winning expentancy formula, ...
これらの原因を解決するには、モデルを変更しないといけません。
実際、これらの原因の解決策として、Glickmanは代替となるモデルに関して
様々な研究を行い、それが、
Mark Glickman's Research Page
http://math.bu.edu/people/mg/research/index.html
にいくつか論文として紹介されています。タイトルを読むだけでも、
Glickmanの狙いが判ると思います。