Shaving Points
Projection Model Picks
Last season the Projection Model boasted gains of +15 Units. While we are ecstatic about the 2022 season results, this offseason we will be hard at work making adjustments to improve the projection model. Stay tuned, it’s getting spicy.
5 Models in 1
Methodology
DVOA Model
This model was created by Q back in 2019 and has been tweaked and adjusted every year since. It uses a combination of Football Outsiders offensive and defensive DVOA, points, turnovers, and time of possession stats on a per drive basis to calculate a projected score. Because this model projects a final score, it is one of the two models that we can use for spreads and team totals. However, we do not generate a win probability number with this model.
EPA Model
Created by Q in 2021 and added to the existing DVOA model for averaging and cross-checking purposes. We start by finding each team EPA per drop back, EPA per Rush, EPA Allowed per drop back, and EPA allowed per rush. Then, we use time of possession stats for each team to project the total expected drives and plays, weighting for run to pass ratios. From there, we calculated the difference between each teams expected EPA to project a spread. This model does not generate a win probability figure or final scores.
POISSON Model
We start out by collecting data for each teams’ points scored, points allowed, yards, yards allowed, first downs, and first downs allowed. Then, we compare each teams’ stats to the average to generate offensive strength and defensive strength. From there, we can generate expected points scored and expected points allowed. At this point, we create a Poisson Distribution table and use that to calculate the probability of any given final score. This model generates all three factors we are looking at: projected spread, win probability, and team totals.
Linear Regression Model
This model uses the averages of five different linear regression models that source historical data on each team for covering spreads, performance at home vs on the road, and as a favorite or underdog. The five linear regression models are made up of:
- The Bradley-Terry Model
- The Team OLS Optimized Rating Model
- The Game Scores Standard Deviation Model
- The Z Score Standard Deviation Model
- The Power Rank Points Model
WEIGHTED LINEAR REGRESSION Model
The Weighted Linear Regression Model uses all the same elements of the Linear Regression Model. However, it generates different results. This is because instead of simply averaging the results together, we use another logistical regression calculation to weigh each model by its correlation to success, adjusting for error.
History
In 2021, Q used the DVOA and EPA models in together (along with considerations from PFF and 538 models) to make all his NFL bets.
Last year, using the model as a major factor in his picks, Q went 88-93 (+7 units). That includes ATS, Money Line, Over/Under, Parlays, and Teasers. However, the models are highly volatile for the first four weeks of the season. This is because they rely on last season’s data at the start of the new season. The 2021 data is just not reflective of what NFL teams are putting on the field in 2022. In 2021, the picks obtained from the model went 75-71 after week 4 (+11.3 units).