:format(webp)/cdn.vox-cdn.com/uploads/chorus_image/image/68634898/1294725363.0.jpg)
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.
RESULTS
After the worst performance of the year in week 16, the model turns it around and absolutely nails week 17.
It picked 13 of 16 straight up winners bringing the regular-season total to 66.4% correct. That is higher than I expected when I started this experiment, but unfortunately, it went 1-1 in upsets and stayed 3 games behind Vegas for the year.
Against the spread, it went 11 for 16. With teams resting starters, I was worried my spreads would be wildly biased, but instead, the model posted its best numbers of the year going almost 70% ATS. That brings the season total to 55.7% against the spread and I’m kind of in disbelief of that number.
:format(webp):no_upscale()/cdn.vox-cdn.com/uploads/chorus_asset/file/22219354/Untitled.png)
PREDICTIONS
In the first week of the playoffs, the model picks only 1 upset: CLE over PIT and that is only by 1/10th of a percentage point. It likes Buffalo to end the Colts season, but it is much more bullish on the Colts in the spread than Vegas, calling it only a 2 point game.
:format(webp):no_upscale()/cdn.vox-cdn.com/uploads/chorus_asset/file/22219357/Untitled.png)
COLTS SEASON
The final expected Colts regular-season win total based on the model probabilities was 9.86, so it was off by 1 game. In individual picks, it was only wrong on 4 Colts games, incorrectly picking wins against JAX, CLE, and PIT while expecting a loss against TEN in week 10. That’s 12 of 16 for 75% straight up accuracy.
:format(webp):no_upscale()/cdn.vox-cdn.com/uploads/chorus_asset/file/22219361/Untitled.png)
What’s even more impressive, is that before even a single snap of 2020 football happened, I correctly predicted that by the end of the year Rivers trailing EPA per dropback would return to his 3 yr average of 0.16 — wait, that’s not the impressive part — AND the “high forecast” model that incorporated that level of play picked 11 of 16 Colts game outcomes.
11 of 16! before the season even started! . . . before the strength of any team was really known . . . before any injuries happened . . . before people finally realized that Gardner Minshew is bad.
. . . why am I the only one impressed by that?