Using variance from 2021, compared to the change in (priced) win total in 2022 we find profitable opportunities among NFL Win Totals in 2022. Between our 3 different publications, over the last 6 years, we have published a painfully accurate, “NFL Going Up/Down” article every season. And with each passing year we have improved the model to really hone in on the relevant variables we’re trying to capture.
The short version for why this works is simple: The NFL plays only 17 games in a season (compared to 162 in MLB, 82 in NBA etc). Along with the chaotic nature of an actual football play between 22 different, independently thinking men (not to mention a handful plus of “third party” HUMANS that act as judge to ensure fair play); actual performance, defined by scoring more points at the end of regulation than your opponent is not entirely dictated by the “better team” (sticky/heavily correlated variable to predicting the outcome of a game), rather circumstances outside of a team’s control (variance). Leveraging this knowledge, you aggregate all the (measurable) variance as best you can objectively, and find teams that outperformed “how nature, in a vacuum would have intended” (and vice versa). Things like recovery of a fumble, how successful your opponent’s were at converting FGs or injuries are simply things teams can not directly control, and therefore are left by chance. And anything left to chance will EVENTUALLY tend back toward the mean (expected outcome).
This we have always known, and have leveraged in past “NFL Going Up/Down” articles. Improvements we made for this season are:
- Knowing alot of these cherry picked variables are not entirely independent of each other, which led to “double-dipping” (think of Pythagorean Wins and Yards per point), we excluded said datapoints
- Assuming the ACTUAL “problem” we are solving is predicting “teams going up/down”, we must make our considerations at the “derivative level”, not at face value (in other words, its not as much what a team’s strength of schedule is this season, BUT what it is this season VS. LAST SEASON)
We have made these minor changes, but still plan on running a regression to properly weight the magnitudes within our model, as they are currently (relatively) weighted evenly.
Nevertheless there is lots to glean from what the model has to say. MIND YOU: this model does NOT take into account roster changes from the last season (think of the Broncos assumed Vegas total being 3 games more than last season, which is mainly due to the acquisition of Russel Wilson). HENCE, as the model is agnostic to these movings, you must use it as a BASE, to then make manual adjustments needed to roster movement.
Aggregate Model
| Team | RZ O+ D | 3D O+ D | FG Luck | NET EPA gained fr TO | SOS Change | Net Rest Rank Yr.ly Change | AGL Lost | Var. Pts | RZ O+ D | 3D O+ D | FG Luck | NET EPA gained fr TO | SOS Change | Net Rest Rank Yr.ly Change | AGL Lost | Var Pts. Avg Rk |
| WAS | -2.35 | 14 | 26 | 3 | 26 | 12 | 26 | 56 | 20 | 2 | 8 | 5 | 2 | 7 | 7 | 7.3 |
| DET | 9.3 | 2 | 20 | 1 | 23 | 1 | 30 | 39 | 3 | 14 | 14 | 13 | 3 | 13 | 3 | 9.0 |
| CAR | 4.2 | 5 | 29 | 2 | 5 | -4 | 15 | 15 | 11 | 5 | 4 | 9 | 10 | 18 | 18 | 10.7 |
| CLE | 1.45 | 15 | 32 | 0 | 1 | -2 | 18 | 19 | 14 | 1 | 1 | 16 | 13 | 16 | 15 | 10.9 |
| CHI | 5.45 | 4 | 23 | -3 | 20 | 25 | 13 | 54 | 9 | 6 | 11 | 25 | 4 | 2 | 20 | 11.0 |
| LV | 12.5 | 8 | 8 | 1 | -5 | 0 | 24 | 19 | 1 | 3 | 25 | 13 | 20 | 14 | 9 | 12.1 |
| SF | -9 | 4 | 9 | 5 | 8 | 2 | 29 | 12 | 29 | 8 | 24 | 3 | 6 | 11 | 4 | 12.1 |
| JAX | -2.15 | -4 | 28 | 10 | 11 | -13 | 19 | 5 | 19 | 22 | 6 | 1 | 5 | 29 | 14 | 13.7 |
| NYJ | -3.85 | 2 | 30 | 5 | -7 | -12 | 31 | -12 | 22 | 16 | 3 | 3 | 24 | 28 | 2 | 14.0 |
| TB | 1.2 | 3 | 24 | -3 | -5 | 15 | 16 | 14 | 15 | 12 | 10 | 25 | 20 | 4 | 17 | 14.7 |
| NYG | -3.65 | -6 | 7 | -1 | 30 | 9 | 27 | 31 | 21 | 25 | 26 | 19 | 1 | 8 | 6 | 15.1 |
| MIN | 1.2 | -10 | 14 | 0 | 6 | 13 | 21 | 12 | 16 | 31 | 19 | 16 | 8 | 6 | 12 | 15.4 |
| MIA | -4.2 | 4 | 29 | 2 | -13 | 14 | 4 | 4 | 25 | 10 | 4 | 9 | 27 | 5 | 29 | 15.6 |
| BAL | -10.65 | 7 | 2 | 1 | 4 | -5 | 32 | -2 | 30 | 4 | 31 | 13 | 11 | 20 | 1 | 15.7 |
| IND | 10.25 | -6 | 13 | -4 | 8 | -3 | 23 | 7 | 2 | 28 | 20 | 28 | 6 | 17 | 10 | 15.9 |
| PHI | 8.25 | 4 | 11 | -1 | 3 | -13 | 12 | 2 | 4 | 9 | 22 | 19 | 12 | 29 | 21 | 16.6 |
| DEN | 5.7 | -2 | 10 | -2 | -13 | 2 | 28 | -7 | 8 | 20 | 23 | 22 | 27 | 11 | 5 | 16.6 |
| DAL | 5.8 | -1 | 17 | -7 | -3 | 17 | 8 | 13 | 7 | 18 | 17 | 31 | 19 | 3 | 25 | 17.1 |
| ARI | 5.85 | -12 | 19 | -5 | 6 | -5 | 22 | -8 | 6 | 32 | 15 | 29 | 8 | 20 | 11 | 17.3 |
| KC | 7.1 | -5 | 4 | 7 | -5 | 3 | 1 | 8 | 5 | 24 | 29 | 2 | 20 | 10 | 32 | 17.4 |
| PIT | -11.95 | -5 | 5 | 2 | 0 | 8 | 20 | -5 | 31 | 23 | 28 | 9 | 15 | 9 | 13 | 18.3 |
| GB | 4.15 | -8 | 31 | -3 | -1 | -11 | 14 | -16 | 12 | 30 | 2 | 25 | 16 | 27 | 19 | 18.7 |
| TEN | -4.75 | 0 | 16 | 3 | -5 | -6 | 10 | -12 | 26 | 17 | 18 | 5 | 20 | 23 | 23 | 18.9 |
| HOU | 3.55 | -6 | 21 | 0 | -2 | -26 | 17 | -28 | 13 | 26 | 13 | 16 | 17 | 32 | 16 | 19.0 |
| BUF | -1 | -2 | 25 | -11 | -17 | 27 | 2 | -2 | 17 | 19 | 9 | 32 | 30 | 1 | 31 | 19.9 |
| SEA | -6.4 | 4 | 22 | -2 | -10 | -1 | 6 | -14 | 27 | 11 | 12 | 22 | 25 | 15 | 27 | 19.9 |
| LAC | -1.35 | -6 | 27 | -2 | -2 | -10 | 9 | -19 | 18 | 27 | 7 | 22 | 17 | 25 | 24 | 20.0 |
| ATL | -4.1 | 3 | 12 | 3 | -10 | -15 | 3 | -23 | 23 | 13 | 21 | 5 | 25 | 31 | 30 | 21.1 |
| NE | -8.1 | 2 | 1 | 3 | -17 | -4 | 11 | -24 | 28 | 15 | 32 | 5 | 30 | 18 | 22 | 21.4 |
| CIN | 4.6 | -2 | 3 | 2 | -23 | -8 | 7 | -26 | 10 | 21 | 30 | 9 | 32 | 24 | 26 | 21.7 |
| NO | -12.9 | -7 | 18 | -5 | 1 | -10 | 25 | -31 | 32 | 29 | 16 | 29 | 13 | 25 | 8 | 21.7 |
| LAR | -4.15 | 4 | 6 | -1 | -14 | -5 | 5 | -19 | 24 | 7 | 27 | 19 | 29 | 20 | 28 | 22.0 |
*For simplicity sake, I set all numbers to essentially be “higher the number = better for the team improving in ____________________ variable” (this is opposite for the ranking)
Table Notes:
–RZ O + D: Redzone Offense AND Defense Rebound effect. This data point was inspired by FO’s “3rd down Rebound” (see indirect link in “3rd Down Rebound” below) with a similar logic: Teams that faired well over the first 80% of yards to a score, yet failed (RELATIVELY) in their EXPECTED performance in the last 20% of yards (and vice versa for the defense) are likely to see a “bounce-back effect”, as “performing well” in a larger sample (i.e. the first 80 yards) is stickier, yet “performing well in the redzone” (i.e. the last 20% of the field) gets all the points
–3D O+D: Similar to above, how a team performs OVERALL is more predicitive, and although 3rd downs are magnified in value, they are simply 1 of 3/4 downs. That means the data point is ripe for regression (either way). We thank Football Outsiders for the concept and data.
–FG Luck: Thanks to our friends at Sharp Football Analysis, this is a variable we’ve wanted to capture for sometime, and finally have the means to do so. It registers the net value of your teams expected FGs made vs. net value of your opponent’s FGs made.
–NET EPA gained fr TO: “Net EPA gained from Turnovers”. For longtime readers you’ll notice this as a (much improved) “True Turnover Rate”. With our old stat we tried to hone in on the turnovers, least in the control of a team (not interceptions or even fumbles, but fumbles recovered), but this datapoint is even stronger, because it teases through all the noise, and cuts into “how much unexpected value did a team NET from turnovers last year”
–SOS Change: Strength of schedule CHANGE. This is NOT SOS based on vegas win totals, instead a stickier formula based on Sharp Football Predictions, and NOT the raw number, but the ACTUAL SOS from last season (for a team) compared to their projected SOS THIS SESON
–Net Rest Rank Yr.ly Change: Another metric from our friends at Sharp, this metric is similar to strength of schedule, but instead is strength of rest. The value aggregates and compares the total NET rest a team has compared to its components, and again is not the raw figure, but a comparison between 2021 actual and projected 2022.
–AGL Lost: Football Outsiders records every injury and weights it by snap, and measures the total “Average Games Lost” from an assumed starter for a the season. This data point is highly volatile between seasons, and always a great indicator or luck (or lack thereof) that will likely change from one season to the next.
–Var. Pts: Simply a weighted, balanced aggregate of all scores combined
-NOTE: All columns to the right of the first yellow highlighted cell are the same measures but ranked, (lower the rank, the better for the team in said metric in 2022)
How to leverage the data
Again, as the point of this exercise is to compare last season (and its level of variance) to this season, without taking current Vegas win totals into account (compared to last season’s actual win totals) is not valuable. Hence, see the same visual with relevant current market conditions:
Note the column labeled “Assumed Change”. This is simply a team’s actual win total, compared to what Vegas is pricing that team’s 2022 PROJECTED win total. This is a key area to compare as it slices into our very investigation. Further to the right you will find “Net**”, this is simply a net comparison (ranked to evenly distribute the data points), between “what direction Vegas thinks a team is going” and our aggregate variance model’s assumptions. Hence, you can read this as “
Please email me directly if you would like a link to the data, and or supplemental sheets these figures were derived by.