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.
Week 6 was a good week. I went 11 for 14 and erased the deficit against the betting lines as both Vegas and my blind model have now predicted 62 of 91 winners (68.1%).
Even more impressive (lucky) was that I went 10 for 14 against the spread, which brings me to a related issue: I am now measuring wins against the spread.
I have not done that previously, as the model does not actually predict spreads, rather it predicts the probability of a win. I then use a separate formula to derive a spread from that probability.
After week 5, I looked at the spread data and saw that it was significantly off from real results, which could be because of: bad model win probabilities, bad applied spread formula, or both. I believe it is both.
The spread formula I used was based on old data, which I updated and it gave slightly better results. However, the biggest impact was from the model itself, which utilizes a variable for home field advantage. This advantage translates to about 2.2 points for the home team (+6.7% win probability).
Before the season started, I knew that most games would be played in empty stadiums and that the traditional home field advantage may not apply, but I left the variable in anyway as I wanted to collect supporting data before removing it. Through 5 weeks, I found that the home teams were underperforming my spreads by about 2.1 points per game, almost exactly the added home field advantage amount.
Of course, that could be coincidental, so I looked at the home team scoring and spreads from the last 10 years and estimated the probability of a similar or greater disparity over a 5 week period (77 games) to be about 7.5%. That’s enough for me to conclude that I probably should remove the variable (in whole or in part).
When I re-calculated win probabilities with 0 home field advantage, my correct picks with the spread remained exactly the same, but my wins against the spread in the first 5 weeks increased from 37% to 51%, which is much more reasonable (50% is expected). Week 6 results boosted the season total against the spread to 55%, which is likely due for a regression.
There are lots of reasons to think a home field advantage should still apply even without fans (weather, field surface, no travel etc.) so this adjustment may be unwarranted. However, I go where the data tells me and there is good support for this change.
This season, I have disagreed with the predicted winner from the betting lines by at most 2 games in any given week.
So, is this the week I finally blow past Vegas or do I eat Sin City’s dust? I’m not feeling real good about these picks, but it’s not about feelings.
(EDIT: Updated table to reflect NFL schedule change.)
(EDIT: Most of this next 2 paragraphs makes no sense as I was looking at the pre-COVID schedule. So, there is no MIA-LAC game rather I have LAC over JAX and no BAL-PIT but TEN over PIT instead, both of which agree with Vegas.)
Hmmm, Miami over the Chargers. I think Vegas recognizes that Ryan Fitzpatrick was benched for an unknown Tua Tagovailoa, whereas my model doesn’t know that. I have toyed with the idea of using NFL averages for replacement QBs, but without actual testing the data, I’m not going to just “wing it”. Maybe I’ll figure something out before kick-off.
49ers over the Pats and Ravens over the Steelers I can see, but I think I’m going to eat it picking Arizona over Seattle. This will be the 4th time my model has picked the Seahawks to lose this year . . . the undefeated Seahawks. I’ve got to figure out what is driving that.
All win %’s have been altered to remove home field advantages. This changed the predicted wins, but not by a lot.
The Colts are currently sitting at a projected 9.3 wins on the season. Even with a bye week, however, that could change as future opponents will still add to their numbers this week.