Over the past two years, PFF has used the machine learning capabilities of AWS to quantify the value gained and lost by each team in the offseason: the Improvement Index. The index is built on the rigorously researched PFF wins above replacement metric, which translates player values into intuitive, concrete and position-agnostic currency of wins. The index is roughly the percent gain (or loss) in WAR for each team during the offseason, beginning with the first player cuts in early February.
We debuted the Improvement Index last season, and the results were strong. The 10 teams that ended the free agency period with the highest Improvement Indexes won 2.7 more games on average than the previous year.
You can learn more about the specifics of PFF WAR here, but it’s most important to know that it’s based on our player-level grades, as PFF grades every player on every play. This wealth of data allowed our research and development team to build a model to simulate the season with and without players based on their grade levels and projected usage in different facets of the game: passing, rushing, receiving, pass blocking, pass rush, coverage, etc.
The model behind the Improvement Index projects the number of snaps and grades per snap for each player in every facet. These models were trained on years of historical data, going back to 2006 for PFF grades and WAR. The model incorporates player- and team-level assumptions based on each player’s prior performance, team coaching tendencies, forecasted opponents and championship odds. These features help the model estimate the likely allocation of snaps among each position group and the likely grades per snap for each player based on that usage and historical trends.
We'll update the Improvement Index plot below every day, adding short write-ups on the most recent transactions that give some context to their impact in terms of WAR for the teams involved and the changes in usage for the specific position groups players are leaving or joining.
Here is the movement in the index, generated by the confirmed transactions, through Sunday, Feb. 21. The numbers include the impact of 2020 opt-out players returning to their teams for the 2021 season.