Please enable javascript in your browser to view this site!

QB Pressure Projection & "Outcome Residual" Consideration

Earlier this week we launched our “QB Pressure Projection” application. The goal of this app is to provide projections for the number of times a defense will hit the opposing quarterback - sack or not - conditional on a number matchup-specific variables. With DraftKings increased price variance at D/ST, I think it is a position we need to be carefully considering, where in the past it felt like a position where we were taking educated dart throws. Sharp DFS players know that when selecting D/ST, we want to place more weight on a D/ST’s probability of producing positive fantasy outcomes (through sacks, forced turnovers, and defensive touchdowns) than their ability to mitigate negative fantasy outcomes (through points allowed). Projecting QB pressure should be the first place we start when evaluating D/STs, because pass plays in which the quarterback is pressured have the highest probability of producing positive fantasy D/ST outcomes.

Projection Methodology

Our QB hits projection algorithm considers the following factors:

  • Spread

  • Game total

  • Offense (team & coaching staff) historic run/pass play & air yards tendencies

  • Defense (team & coaching staff) historic run/pass play & air yards defended tendencies

  • Offense historic scaled QB pressure allowed

  • Defense historic scaled QB pressure generated

The crux of our projection is estimating how good offenses are at mitigating QB pressure and how good defenses are at generating QB pressure. The easiest approach would be to just capture the average QB hits per game an offense allows or a defense generates. However, this a flawed approach for at a couple of reasons. First, the number of QB dropbacks (opportunities to hit the quarterback) are variable from game to game based on offensive tendency, game script, etc. Secondly, the probability of a defense getting to the quarterback is dependent upon a number of game conditions - it is far more likely for a defense to get to the quarterback on 3rd & long where they are expecting a pass than it is on 2nd & short, when the possibility of a run is more likely. Because of these two caveats, we estimate QB pressure on a per-dropback (instead of per-game) basis, and we attempt to scale QB pressure based on how likely it is that a QB hit is observed.

Our solution: “Net Pressure Rate” allowed (by offenses) and generated (by defenses). NPR is the average “net QB hits” allowed/generated per QB dropback. Unscaled, a QB hit is either a 1 or a 0, it’s a binary outcome. However, we scale all QB hits based on the probability of a QB hit on each play (as estimated by an internal machine learning algorithm), conditional on game factors like down, yards to go, yardline, time remaining, and score. In general, these factors contribute to the predicability of a pass play, which in general is correlated with the likelihood of the defense getting to the quarterback. As such, a defense that registers a QB hit on a dropback where the likelihood of a QB hit is less likely (say 2nd & 2, inside opponent territory, in the 4th quarter, with a 10-point lead) is attributed more pressure than a defense that registers a QB hit in a predictable pass situation (say 3rd & 12, around midfield, down two scores, late in the 4th quarter). You could make they case that they generated QB pressure at longer odds of doing so, thus, on this given play they should be attributed a higher pressure rating. The mean of all net qb hits produced by a defense or allowed by an offense is their net pressure rate generated or allowed, respectively. I think the best way to think about this metric is the increased or decreased QB hit probability that a offense allows or a defense generates on top of the baseline probability of such occurring given the condition of each play, or a games set of plays.

QB hit projections for Week 6, 2019

QB hit projections for Week 6, 2019

2019 QB pressure allowed and generated metrics through Week 5.

2019 QB pressure allowed and generated metrics through Week 5.

Outcome Residuals

I think in evaluating DFS outcomes, it is valuable to not only observe and evaluate outcomes simply as they are, but also consider if and by how much they deviate from expectation. Our Net Pressure Rate metric is what I’d call an “outcome residual”, which is something we try to focus on and quantify at ASA. As discussed above, we wanted to avoid looking at QB pressure simply as the summation of outcomes, and more precisely as and aggregation of how those outcomes deviated from the norm.

The same approach applies to our play count and air yards projection. We not only consider how often teams run run/pass plays, and how many air yards they average per play, but we also consider how those outcomes deviate from what is expected under the conditions that that play are run. An interesting example is the New England/Bill Belichick offense. Historically, they appear to be a very run-heavy team, they are perennially near the top of the run plays-per-game leader board. But they also have been in so many games in which they are leading, and a run-heavy approach is the norm. When we scale New England’s offensive tendencies to the conditions in which they operate, we actually can see that they are a fairly pass-heavy team relative to expectation. That being said, they will frequently find themselves in conditions where a run-first approach is the expectation, and even a skew towards passing will result in a still run-heavy offense. But by understanding their offensive approach as a residual, it allows us to flexibly project for them when conditions and game environment change.

In general, I think this is a really good approach to evaluating outcomes, and something we will strive for as we continue to grow ASA’s quantitative offering. Rather than saying “we’ve observed X, therefore in the future we can expect to observe some outcome that centers around X”, we want to strive for “we’ve observed X1, and given conditions A, the expectation was X2, therefore given the difference between X1 and X2, if we are projecting for an outcome under conditions B in which the expectation is Y1, we might be able to alter our expectation of Y1 by that X1-X2 difference”. Something to think about as you’re evaluating players, teams, games, or perhaps life in general.