re: Week 8 Power Ratings and SEC slate

quote:

---

LSU fans still trying to pretend that ole miss game was close

Anyone with a brain knows that was an arse whooping

---

Ole Miss fans still trying to pretend the LSU game wasn't their Super Bowl.

It was not an "arse whooping".

quote:

---

in taht case UGA shouldnt' have moved at all since if our starting AA caliber safety had played (which he did not) we likely don't lose.

---

That’s not how power ratings work. It’s isn’t a poll where you get dinged for losing. It’s a sliding scale of strength. So even being in a close game with Auburn hurts you.

That said, UGA had a bunch of fumbles so probably dropped less as those aren’t considered predictive. LSU is higher in power ratings than rankings because UK hinged on a punt return + 77 yard run + 2 point conversion. None of which really predict future success as well as other stats.


The problem is that too much weight is being given to MOV against bad Clemson/Aggie teams and no adjust to 31 point margin against a team without a QB

quote:

---

Team A might pull its starters at halftime and only beat an FCS opponent 45-8, while Team B might leave theirs in the entire game and beat that same FCS opponent 70-0.

---

FCS games are completely ignored - the reason is that a random FCS team does play enough other FBS games to become an essential piece in the matrix of games. FCS games just create noise that makes the model more inaccurate - they are removed.

But your point still stands - you are misunderstanding the purpose of the model. The model is trying to predict scores and margins(spreads). If a team likes to keep their first string in and run up the score, then they'll probably do that to the next team they play, so that's baked into the model. So yes, if a team likes to run the score up on bad teams, then the model rewards them for that - HOWEVER, if that team is playing below-average teams, then the model picks up on that as well. Then, once the team plays an above-average team, model considers how they performed. It all works itself out over time. In other words, the model sees "you scored a lot of points against below-average teams, but you didn't do that to the top 25% of teams".

So to determine LSU's rating, it looks at LSU's performance against each team they played, and then each of those teams performances against their opponents, and then each of those teams performances against their opponents.. and so on. It's a very large system of equations.

quote:

---

Any power rating model needs to take into account that once you reach a certain discrepancy in talent, the actual outcome of the game doesn't really matter anymore.

---

Talent is only considered at the beginning of the season, it's your starting point. From there, on field results are the only thing that matters.

The best way to increase your rating is to outperform the model's expectation and for your opponents to outperform the model's expectations.

You can't just casually look at teams like LSU and say "they're just beating up on bad teams". That's simply not the case. There aren't 10 good teams and the rest are "bad". It's a gradient over 138 teams. The most average teams are Michigan St and Virginia. The model sees Clemson and Texas A&M as above average teams. And LSU outperformed the model's expectations in those two games - by a lot. LSU has beaten up on above-average teams.

quote:

---

Clemson and Texas A&M are both above-average teams.

---

Yeah, well that’s just like your opinion man


Above average among FBS teams? Sure. But bottom of the power conferences.

And a Kamario Taylor injury dropped the Vegas line 8 points. You don’t see that much impact from a non- QB injury. In fact, if Kamario comes back healthy it will make your ratings worse because LSU will get even more credit for beating an injured team

I’m not saying Ole Miss should be above LSU. I’m saying LSU should be in the same 17-18 point tier instead of way higher. Drop LSU 6 points and ND 5 points and I wouldn’t quibble with the rest

quote:

---

But your point still stands - you are misunderstanding the purpose of the model.

---

Look, you arrogant asshat, this is a power rating system for college football teams, not the operating manual for a nuclear power plant. Everyone here understands "the purpose of the model." I think our questions are more along the lines of, "If your model produces such obviously crap results, then what purpose does it really serve?"

quote:

---

The model is trying to predict scores and margins(spreads).

---

Ok. How is it doing at this? What is the model's record against the spread season-to-date? What is the model's record straight-up?

quote:

---

If a team likes to keep their first string in and run up the score, then they'll probably do that to the next team they play, so that's baked into the model. So yes, if a team likes to run the score up on bad teams, then the model rewards them for that - HOWEVER, if that team is playing below-average teams, then the model picks up on that as well.

---

This is a self-contradiction. So the model rewards running up the score on bad teams, but also at the same time penalizes teams for playing mostly below-average opponents? Ok. It seems obvious by looking at your results that you have weighted "running up the score on bad teams" far too heavily compared to whatever penalty is built in for "playing mostly below-average opponents."

quote:

---

The model sees Clemson and Texas A&M as above average teams.

---

Ah, so your model is drunk. Never mind, that explains it.

You get a lot of down votes, but I love these kinds of threads and I understand how models work.

Here's my local predictions using F+ power ratings.

Anyone who uses these to gamble is a fool, this is just for entertainment purposes only.

Rankings are my own, not AP or anything like that.

quote:

---

It is in CFB. I model it mostly retroactively to avoid giving a team an additional boost when they beat a team that had to suddenly play back-ups. The problem is that, with MSU @ LSU for example, it's just one game of data so you can't just punish LSU harshly for pummeling a disabled opponent. MSU's season EPA is what it is, you just have to hope you're applying the appropriate weight to recent game data.

And even then, you have games like Mizzou @ Ole Miss that statistically just make no sense. My models had Mizzou by 6-7 and Mizzou by 8-9 and they lose by 28. I'll nerd out for a sec if you'll humor me.

You have a team in Ole Miss that's bottom 10 in the country in Rush defense EPA and bottom half in rush offense EPA vs. a team that's top 10 in rush offense EPA and middle of the pack in rush defense EPA. All of these EPAs are opponent adjusted too. The result? Ole Miss outrushes Mizzou by 230 yards and wins by 28. We correctly predicted that a dominant rushing EPA advantage results in a big point swing. We could not have predicted it would be Ole Miss with the dominant advantage. You just gotta throw your hands up sometimes.

ETA: EPA is efficiency per attempt.

---

I haven't done power rankings directly in over 10 years. It seems impossible to me to truly take injuries into account. I consider that to be human input, so I typically avoid that.

How do you adjust for a lineman who generally doesn't have direct stats? It seems like some guess work, and that taints the stats for me. I think of doing the human adjustments in my head after. AKA if it predicted Miss St to win, I'd be thinking - well it's not accounting for the QB injury, I need to do that myself.

I'm just using F+ ratings for my power scores currently. It's not injury adjusted at all AFAIK. Maybe SP+ is doing it, and that makes up half of F+, I don't look at those directly anymore since they went to ESPN.

It'd be something like team x returns y OG this week. Compare team x's offensive EPA the first 3 weeks where y OG played vs. the two weeks he didn't and apply a weight.

The harder part is making that a reproducible prompt that, to your point, doesn't require weekly interjection but AI tools do make it a bit simpler. What makes injuries stupid to try and predict is the lack of official injury news. In the NFL, you know game participation lists are required an hour and a half before kickoff so you can schedule prompts for those slates to run. Half the time you're getting official word that college players are out literally minutes before kickoff.

Then you have situations where someone like Trammel Jones is starting over Aaron Philo. We should discount Florida some because they're missing a starting QB but how much? We have no data with Trammel at QB.

QB is really the only position I try to model for injury status as it's easily the most significant. Otherwise I think my model may take injury clusters in account IE starting RG and RT are out, multiple DB injuries, etc.