This is an exercise in building a simple blind predictive model based on a team’s historical stats in 4 areas: pass and rush efficiency, and avg EPA for special teams and penalties. The season long experiment will track how well future games are determined by past statistical performance and compare that record against Vegas betting lines.
All data used originates from Pro Football Reference and nflFastR.
Before I jump into week 12 results, I want to point out some maintenance that I did:
- Week 12 forecasts were adjusted prior to kickoff for Denver, Chicago and Baltimore to account for starting QB changes in those games.
- Cleveland and Carolina forecasts were adjusted prior to games to account for math error in calculating QB stats.
- Vegas point spreads were updated from weeks 1 - 12 to capture the spread immediately prior to kickoff. This resulted in a decrease in weeks 1 - 11 Vegas predicted winners from 110 to 108, which ties the model predicted winners going in to week 12. This did not change the net total winners against the spread for the model, but it did cause one game that was previously listed as a push to become a loss for the model.
- Colts LO, MED and HI forecast probabilities were changed to reflect the removal of home field advantage. This was done in week 6 of the model, but I forgot to update that table.
RESULTS
After a poor week 11, the model rebounded in week 12, as it correctly predicted 2 of 3 upsets (WAS, TEN), going 12 for 16 straight up and 9 of 16 against the spread.
That brings the season predicted win totals to 120 (67.8%) and 1 game ahead of Vegas with a 53.4% win rate against the spread.
While those numbers are enough to turn earn a profit off your bookie, the incremental wins over Vegas are razor thin and thus can’t be differentiated from pure luck. The true success of this model is the ability to mimic the market forces around game spreads using only historical data and no human perception at all.
PREDICTIONS
For week 13, the model predicts 2 upsets: Cardinals over Rams and Falcons over Saints. I’m guessing the driver of the Atlanta upset is the QB change in NO but that should be countered with recent poor performances by Matty Ice. I’ll have to open the hood to see what is going on there. Same with ARI-LA as that deviation from Vegas is a head-scratcher.
COLTS SEASON
Using total expected win probability through week 12, the model predicted 6.4 wins for the Colts and 8 wins using binary win predictions. Reality split the difference and the Colts sit at 7 - 4.
Against Houston, the model gives the Colts a 62.1% chance, which equates to a 3.8 point margin.
The current predicted 9.43 season wins is almost exactly what the “HIGH” model forecasted prior to the season start. That assumed Rivers passing efficiency would be 0.16 EPA/db mirroring his cumulative 2017-19 efficiency. As of now, his weighted EPA/db is 0.144 so that all lines up really well.
The probabilities in that “HIGH” forecast were set before the year even began, but are still very close to actual Vegas spreads. For example, once Rivers was signed back in March, the model predicted that in week 13, the Colts would be 3.2 point favorites and today — 8 months later — the spread sits at Colts by 3. THAT’S CRAZY!