I think there is a slight difference of "philosophy" as it were on how to compare teams. My goal is to use as few parameters as possible. I have no doubts a model with 27 fit parameters will fit the data better than a model with 3. I appreciate the limitations of using only 2+1 statistics (points per play scored, points per play given up, and losses), but in fact that is my goal. There is a strong correlation between the Pythagorean win % calculated with OE and DE and the actual win %.
Another limitation is the availability of statistics in computable form. Possessions are not an easy stat to extract for every team for every game, but I agree I would love to have it.