This will be a more analytics focused post, although it is the basis for a prior post on why I think the Eagles will exceed their 2021 line of 6.5 wins (https://phillycovercorner.com/2021/07/why-the-2021-eagles-will-beat-their-6-5-win-projection/). I have gone through each team ahead of the 2021 season and created win projections using an AV-based model.
Background on Approximate Value (AV)
Approximate Value (AV) is a ProFootballReference metric that assigns a value to every player. The method is more complicated but the basis for AV is the ability to score AV is a measure of a player’s value and is at its core based on the ability to score more than expected (for offenses) or prevent scoring more than expected (for defenses).
AV, like other similar metrics, has its flaws, specifically for more difficult positions to assign contribution to like offensive and defensive line, while skill positions are much easier. At an individual player level, AV will have variance issues but in aggregate, a metric like AV becomes very useful. Below shows the relationship between a team’s aggregate AV and its win percentage:
The R2 of team AV and win percentage is 0.795, meaning almost 80% of a team’s win percentage is explained by the aggregate value of the players.
A more commonly used metric is Expected Points Added (EPA) which is available in the NFLFastR data set. EPA is similar – for every play in every game, that play is evaluated on how many expected points were added to a team. A run that picks up no gain will have an EPA loss while a plays with larger gains will have positive EPAs. The concept is similar to AV (value of scoring vs. preventing scores). But there are big differences and uses for each – EPA is available for every play, meaning for each player and play type and game situation, you get a value for the play vs. expected – AV does not do this as it is an aggregate number per player per season. But EPA applies primarily to skill positions and defense to a degree, there isn’t an EPA for an offensive lineman for example. This is one of the advantages of AV over EPA. Both are awesome, but I use them for different things.
But for these purposes, AV and EPA are remarkably consistent. Here is the same chart as above, but comparing EPA to win percentage, which has the same exact R2 of 0.795:
And if you compare a team’s AV to its EPA, you see that in aggregate, the two are effectively the same with an R2 of 0.943.
The Pythagorean expected wins model
One commonly used win prediction models in the NFL is the Pythagorean expected wins model which compares points scored by a team to their points scored plus points allowed:
Expected wins = Points Scored2.37 / (Points Scored2.37 + Points Allowed2.37) * 17 games
But the ability for the Pythagorean model to predict next season wins is poor as there is very little correlation below between expected next season wins and actual wins.
One of the big issues with it is it does not reflect personnel changes. Tampa Bay had a 2020 expected wins total of 8.2 but that was based on a 2019 season with Jameis Winston, not Tom Brady.
This model, though, is useful retrospectively to show which teams outperformed or underperformed what they should have and may have benefited from luck (excess turnovers, high percentage of one-score game wins).
Modeling the 2021 season
First, no prediction model is ever going to be highly accurate for such a complicated sport like football. Injuries, coaching, and “luck” always swing several teams fortunes each year and certain positions have an outsized impact on the team’s overall success. Losing a starting QB (unless you are the 2017 Eagles which is an even freakier year than it seems when you really inspect it…) is almost always devastating to a team. Really poor OL play has spillover effect that impacts the overall offense (pressure decreases offensive DVOA by over 100 points).
But, talent is talent and players, in aggregate, have a somewhat consistent career arc of improvement, plateau, and then degraded performance as they age. And while predicting individual players is almost impossible as there is so much variability at an individual level, aggregating a team makes it easier (central limit theorem and all…).
The AV-based model has the following main tenants:
- Predicting and summing AV for each of a team’s 22 starters (further explanation below as this is the key part of the model)
- Non-starters (bench) are not individually modeled but aggregated as an additional 27% of total starter AV (looking backwards, bench AV is relatively consistent across teams)
- Applying a strength of schedule modifier to total AV, ranging from the weakest SoS being worth +10 team AV and the hardest SoS being worth -10 AV.
Starter AV Modeling:
For non-rookie starters, AV starts with the player’s prior year 16 game AV pace (adjusts for players that played only a partial season and extrapolates out their full-year performance).
AV then has three modifiers:
- Age
- Early career growth
- And a subjective “other”
As mentioned above, players in aggregate have a somewhat predictable career value arc where they have consistent year-over-year AV improvements in their first few seasons, then are relatively flat year-over-year until they near 30 where they begin to show AV decreases.
To model this, I give players in their first 3 seasons a +25% AV adjustment and a -16% AV adjustment starting at age 29 and a -32% AV adjustment at 31.
On the last bullet above, I have not tried a pure formulaic model and believe some subjectivity needs to exist in the model. One current year example is Matthew Stafford as he moves from an awful team with no weapons to a team likely competing for the Super Bowl. In 2020, Stafford created 11 AV which is only a bit higher than Jared Goff’s. But the expectation is that Stafford will provide materially greater value moving to the Rams (going from the 31st best WR group to the 8th). Another example of a manual adjustment is Tom Brady – until he actually shows he isn’t impervious to aging, an age-related negative adjustment makes no sense and needs to be corrected. And QBs in general show less age-related decline than other positions (RB is very pronounced on the other end). So there is subjectivity in the model but for a game like the NFL, there needs to be.
The modeled starter AV uses a 16 game pace but we know many starters will not play all 16 games due to injuries and other reasons. They are modeled to have a full season, though, which will capture their replacement’s contribution.
Rookie Starters:
For rookie starters, predicted AV simply uses the average rookie season AV for past rookies of the same position and same draft round.
Bench:
A team’s overall performance is more than just their starters, but the starters are the more variable and impactful driver of team value. Total bench value does vary but is more consistent across teams – 75% of team seasons between 2015-2020 were all within a 5% range of bench to starter AV. To represent bench value, an additional 27% of total starter AV is added. There may be some small incremental value in estimating special teams or bench value more specifically but there is diminishing value here.
Strength of Schedule:
SoS has shown to have an impact on win totals, with the easiest schedule equaling +10AV and the hardest schedule -10AV, with a progressive range in between. Getting a SoS list gets complicated as many of the preseason lists are based off of the prior season opponent records or DVOA or other metric and do not reflect changing personnel.
For this, once all teams are modeled, I generate a ranking of team strength, from highest total AV to lowest and then use this to generate SoS for each team based on who they play. For 2021, my strength of schedule is the following (going from easiest schedule at the top to hardest at the bottom):
This differs from the SoS lists that use 2020 opponent records (this is one example from CBS sports https://www.cbssports.com/nfl/news/2021-nfl-strength-of-schedule-ranking-for-all-32-teams-heading-into-the-nfls-inaugural-17-game-season/), which had the Eagles with the easiest schedule, along with Dallas, Atlanta, and Miami. The Eagles do play the Lions, Jets, and have the NFC East, but also have the Chiefs, Tampa Bay, and 49ers as well as the Chargers and Broncos who I have rated stronger than pre-season lists.
Sharp Football Analysis has an updated SoS based on Vegas’ projected wins for each team which is more accurate than just using the prior season (https://www.sharpfootballanalysis.com/analysis/2021-nfl-regular-season-schedule-grid-strength-of-schedule/).
Calculating Win Totals:
Based on the regression shown above (the very first chart in the article), I use the following to calculate a projected win total from total team AV:
Predicted Win % = (0.646377 * TotalTeamAV) – 82.9537
Which is then multiplied by 17 this year to get a win total.
An Eagles back-test from 2020
I back-tested this on several teams and seasons, but wanted to share one on the 202 Eagles (since this is primarily and Eagles blog page). Below lists their projected 2020 starters at the season start – Desean is listed in although he ended up barely playing and the OL reflects JP at LT and Pryor at RG with Dillard and Brooks out before the season. I made no manual adjustments and the projected win total comes out at 4.1 wins which is right where the Eagles were.
A couple of things to note:
- Carson was one of the big swing points last year as he was historically bad and took the entire offense down. I projected him at his same 12 AV as the prior year but he only generated 5.3 AV in 2020. His replacement, Hurts, generated another 4 AV in the 4.5 games he played, giving the QB position 9.3 AV in total.
- The other big story last year was the OL injuries which I have written about before being such a statistical outlier. This model didn’t fully reflect the impact but because much of the line is older, the age penalty did bring the Eagles OL value down.
- There was good excitement on Nickell-Robey Coleman as a big nickel corner upgrade, but he had been between 1 and 3 AV per 16 games the past five seasons. His struggles reflected his historical value.
- Reagor was the rookie starter and 1st round WRs have historically generated 5.5AV over 16 games. Reagor, even with what most see as a disappointing rookie season, was right on forecast with 4AV over 11 games or 5.8AV over 16 games.
- There were several other players with errors but much of it evened out. For every Avonte Maddox (projected due to growth to be 6AV but was awful outside and generated 3.2 AV) and Miles Sanders (projected at 10AV but had a down year at 8AV), there is Alex Singleton (projected to only generate 1.8AV and ending up having a great year with 8AV) and Brandon Graham (doubling his projected AV with probably the best Eagles performance last year). And this is the point of aggregating team AV – I don’t think individual player predictions will ever be accurate, but aggregating evens out individual errors.
- And lastly, the overall Eagles projection took over a half-game hit due to a more difficult SoS.
This post was meant to explain the model and I will create another post with actual team predictions in order to keep the articles not novels.