The NFL constantly evolves, and so must our analytics. Every year, we update the historical Adjusted PFF Scores for each player (that qualifies), using the scores from the recent year (if they improve upon the existing score for that player). We then re-run our calculations to re-maximize our correlations, and sharpen our algorithms. This allows us to adjust for new trends in the NFL.
We also cycle out the data points for players whose rookie years are older than 8 years (unless they are still actively rostered). This allows us to keep the data from becoming stale.
2024 Additions
The most significant addition is that we have added an algorithm for safeties. We had not done so before because there was a recent shift in the NFL’s safety usage. The correlations before and after that shift conflicted with each other, preventing a unified safety model. However there were insufficient recent data points after the shift to feel confident in creating a model based only on prospects drafted after that shift. With the accumulation of another year’s worth of data PFF scores, there is sufficient data to draw correlations. You can see from the graphs on the front page that the Safety model (much like the Center model) contains fewer points than other models (such as WR). The accuracy of our correlation is likely not as robust as in other models, but it should be serviceable.
We have also changed the model for runningbacks, to exclude the short shuttle and 3-cone drills because the prospects do not appear to be running those drills. This likely degrades the robustness of its correlation, but the model had very few qualifying prospects without the change.
We now have estimated YAS ranges if a prospect completed all but two (or one) of the required events for a YAS score at that position. We assume they achieved times in the 25th and 75th percentile, and produce an estimated YAS score. These appear as ranges, instead of definite YAS scores (i.e., 6.3-9.7). These are less reliable, but allow us to do some level of analysis even if there is one event missing.