D/ST is where I started, due to a slew of bad picks from other rankings.  I thought rankers were assuming that certain statistics were predictive, without testing them.  I was especially frustrated about the inability to identify early-season D/ST break-outs, by intelligently using information from the new season.  I thought rankers weren't judging past defensive performances by the strength of their past opposing teams.

Verdict:  It turns out there was room for improvement, and my models’ successes have surprised even myself.  D/ST prediction can’t get a LOT better-- because the randomness will always overwhelm the trends.  Still, accounting for strength-of-opponent still form the basis of all my data processing and gives an extra advantage in forecasting.