[HIGHLY EXPERIMENTAL] 2076 Conference Games prediction model

[TheseBoots]

Time to go out on a limb and have a stab at turning all the numbers into something predictive. Highly experimental first attempt only, but here is the numbers numbers numbers derived preview of the 2076 Conference games!


Difficult but not impossible matchup for MUR made worse by WVA's homefield advantage. WVA's offense is elite in both phases, so MUR will need their defense to match its season performance for 4 quarters. If they do that the score should remain close enough to give them a solid chance as WVA's pass defense is good, but not elite. (FWIW I think MUR has a better chance than the prediction - like I say - experimental!!).


CAL has the more balanced offense and can attack both way, but TEX's defense is real, and if they play to the regular season standard they can make things uncomfortable for the Stamps. TEX needs a clean game and to win the turnovers to offset CAL's home advantage.

[TheseBoots]

OK, so, caveats time. Net EPA is predictive of wins, but definitely not perfect. The gap is explained by: small sample size, special teams, turnover luck, close game variance, red zone efficiency. The model has no knowledge of injuries, scheduling, or current form, rather it treats each team as the sum of their full-season numbers. All numbers are regular season only. I haven't evaluated whether the playoffs lead to different outcomes.

As to the methodology...

Each team gets a single quality score called Net EPA per play which is their average offensive EPA per play minus the average EPA they allow per play on defense. It's a one-number summary of how much better or worse than average a team is on both sides of the ball combined.

To predict a game, the model takes the difference between the two teams' Net EPA scores, always from the home team's perspective. If the home team scores +0.08 and the road team scores +0.02, the difference is +0.06. If the road team is better, that number is negative.

That difference gets scaled up by a factor learned from 21 seasons of FOF8 MP games, essentially asking how much does a given EPA gap translate to actual wins? The answer turned out to be a multiplier of about 8. So a difference of +0.06 becomes roughly +0.48 on an internal scoring scale.

Home field is then added as a fixed bonus of +0.28 on that same scale, regardless of who's playing. This was learned from the data too and reflects the observed ~56% home win rate across the training data.

The final step converts that internal score into a probability. The conversion follows an S-curve: scores near zero give probabilities near 50%, and as the score gets larger (in either direction) the probability pushes toward 100% or 0% but never quite reaches either.

[TheseBoots]

just realised I forgot to flip the neutral venue correction - will fix next time

[Roadhouse_band]

very cool. did you test this at all in the regular season? just curious since it seems to be producing pretty heavy favorites for a conference championship

[TheseBoots]

very cool. did you test this at all in the regular season? just curious since it seems to be producing pretty heavy favorites for a conference championship

Based on a season yes, the distribution is U-shaped, with ~70% of games falling in the middle third of probabilities. Across the IFL2076 season, for games where one team is rated =>70% chance to win, the model predicts that overall 82.6% of those will win. In reality it was 87%. So for games with a strong favourite the model actually slightly under-predicts - small sample size though. Would need to run it on a much larger set (and will at some point).

TEX at CAL seems reasonable. That is a 10-6 team with a negative points differential visiting a 14-2 team, so winning 1 in 3 seems OK. This game is rated in-game at CAL (-7)

The other one is interesting for sure. WVA's offense is extremely good, so that tracks, but MUR is also very good in both phases (basically the two best teams) and feel a bit underpowered in this current iteration. They still win 1 in 4 according to the model so it is far from impossible, but in-game this matchup is rated even (so MUR rated slightly stronger as WVA is at home). WVA's EPA/play on offense is driving the big differential - I don't really know how volatile that is when faced with a top 5 def. like MUR in a one-off game

[TheseBoots]

2/2 hehe

[joe the commissioner]

There's no favorite in the Bowl. I'm curious what your model says.

[Roadhouse_band]

Based on a season yes, the distribution is U-shaped, with ~70% of games falling in the middle third of probabilities. Across the IFL2076 season, for games where one team is rated =>70% chance to win, the model predicts that overall 82.6% of those will win. In reality it was 87%. So for games with a strong favourite the model actually slightly under-predicts - small sample size though. Would need to run it on a much larger set (and will at some point).

TEX at CAL seems reasonable. That is a 10-6 team with a negative points differential visiting a 14-2 team, so winning 1 in 3 seems OK. This game is rated in-game at CAL (-7)

The other one is interesting for sure. WVA's offense is extremely good, so that tracks, but MUR is also very good in both phases (basically the two best teams) and feel a bit underpowered in this current iteration. They still win 1 in 4 according to the model so it is far from impossible, but in-game this matchup is rated even (so MUR rated slightly stronger as WVA is at home). WVA's EPA/play on offense is driving the big differential - I don't really know how volatile that is when faced with a top 5 def. like MUR in a one-off game

awesome, thanks for sharing all this. Congrats on the 2/2 haha

[TheseBoots]

There's no favorite in the Bowl. I'm curious what your model says.

It is not pretty, LOL. Slightly worse odds than MUR even though it's a neutral venue. TBH WVA handled me pretty easily at their place in the regular season and it will not be easier now that I am down WR1 as well my QB. Will WVA missing their C hurt that nasty nasty running game? A click of the simulate week button will reveal all! WVA HEAVY favourites though. They could easily have been a 15-1 team. Only VAN really handled them, and I don't think CAL has the personnel to play a game like VAN does. I don't know, I usually get my defense playing better in the post-season, so I am probably slightly under-fancied here, but not by much.

[TheseBoots]

3/3

[Roadhouse_band]

All hail the model