Cumulative EPA -2021

Cumulative EPA (Expected Points Added) is something that I first saw briefly on Twitter back in the day and wanted to see if I could replicate.

The basics is just to take a team's EPA across all of their plays for the entire season and use that to see both trends up or down, as well as which teams are probably the best teams all together. Obviously, you get some level of selection bias for teams that make it further in the postseason (looking at you KC), but in general looking at the total season's trends can be useful and for comparisons just use the average or rate statistics. 

Another point I want to mention here is that I do adjustments for EPA in pretty much all of my calculations and have found it to be more stable and predictive of next year's success in terms of raw EPA, points scores, and wins. The adjustments I make are simple; adjust for strength of opponent, and adjust for the current game state in terms of Win Probability. 

For these adjustments I take a team's values in rushing and passing offense and defense compared to the league average for that given season and use that to scale up or down an opponent's EPA performance against them. For example, you shouldn't get as much credit for smoking a bad team, because everyone smokes that team. 

Finally, I downweight EPA in really high or low WP scenarios. If you are winning and just grinding out three yards and a cloud of dust, your EPA means less because you aren't trying to gain every extra yard. If you are down 3 scores and the game is pretty much over, the defense usually changes and allows a lot more hollow yards knowing that they only have to bend and not break because they are trying to force you to eat up clock on your drives. In both these cases, the predictive value of EPA diminishes the further the game gets away from an even matchup. I adjust both rushing and passing differently (see below) since teams address them differently as the game gets more solidified. 

In conclusion, I think that adjusting EPA into more specific situations instead of the raw values can help with the predictive power of the stat. Additionally, it is always a good idea to visualize your data because you can sometimes see trends that wouldn't be apparent with descriptive stats or when looking at a table.