Computer Rankings
- Thread starter lunker35
- Start date
Good question. The only two decent teams Liberty has played, Bama and NU, they lost to. Somehow number of wins must be have a greater weighting than th ranking of the opponent.
Thankfully the computers don't pick the 68m tournament teams. And NU beating a Liberty team with a higher computer rating helps the 'Cats.
All the computer rankings have us set as a bubble team that is generally on the right side of the bubble at the moment, but close enough where that could go bad in a hurry. Is that not… exactly what NU is?
Liberty's non-conference schedule was tougher. Rated 110 in the country. We are at 275. In addition, they've blown out their opponents regularly, which counts in KenPom (winning by more than expected). They're only 5 ahead of us now, BTW,
If it was run after this game we would likely be moved up. Still need to win a few more games
He runs the rankings several times a night. They moved from 57 to 51 after the win.
Computers are fast!!!!
I did my PhD with a PDP-8 in the sub-basement in Tech. How times change.......
I can tell you what Jeff Sagarin does (and the other systems end up being pretty close).
He assigns something like 3 points for home court advantage.
Each game against Division 1 counts equally.
He uses the scores only.
He sets the team ratings so that the total error for all the games is minimized.
Error is the difference between the expected outcomes and the actual outcomes.
So his system gives a team just as much credit for winning by 30 when they're only supposed to win by 10 as it penalizes the same team for losing by 15 when it is expected to win by 5.
In essense, his program says Team A beat Team B by 10, but Team B beat team C by 3, so team A is 13 points better than team C. And he does that across all the games.
KenPom is more complicated and uses fancy stats and some proprietary stuff. I tend to default to Sagarin because I think it’s easier to read and eyeball, but KenPomm is slightly more accurate vs Vegas than Sagarin
Well, one problem in equally-weighting ALL the games is that when teams improve significantly, the early games hold the rating down.
And the non-conference results dictate which conferences are rated highest, with no possibility of change. So if the level of play improves in a conference during the season, Sagarin (and the others) will never catch it - like what happened with the ACC last year.
KenPom ratings don't take wins and losses into account at all. It's just a calculation of how good you are offensively and defensively per 100 possessions adjusted for the level of competition you've played and where you've played them. An overall KenPom rating is just the difference between offensive rating and defensive rating. It stands to reason that the wider that gap is (in the positive direction), the more likely you are to have a good record.
This is, of course, not always the case. If you mostly win by blowout but mostly lose close games your efficiency margin is going to look much better than your actual record (case in point, the team we just beat, which has the efficiency margin of a 16-win team but only 11 actual wins).
NET now has us at #49 — still behind Ohio State! 6 Quad 1 wins for the Cats — only 6 teams in the country have more than that. 6 road wins — only 2 P6 teams have more than that (Purdue and Alabama). That’s a pretty darn good resume.
Sagarin (or KenPom) wouldn’t actually predict NU to necessarily go 1-6 just because they’re only favored in one game. As small underdogs in most the remaining games, you’d have to convert all those games into some percentage of a win, likely a lot in the 35% area. If we average just 35% to win the games vs Iowa, MD, ILL, and IU, then Sagarin would actually be saying the expect NU to win at least 1 of those four games, on average (obviously actually like 1.4).
I understand here, and I hope you're right, but wins/losses are discrete events. 1.4 is nice, which means one is likely, but they still either get a win or loss, and these games are close enough that the luck could go the other way.
Already the Iowa game has flipped from a tiny loss to a tiny win, but not a whole lot has changed in the world, so your comment is still apt.
I started a new thread about this topic, by the way, so feel free to cut and paste there and let's get that conversation going!
This is not how projections work. An analysis that treats either Iowa or Northwestern with a full projected win due to a .03 point spread either way, giving the same credit to the favorite as Rutgers as an 8 point hike favorite over NU, is flatly incorrect statistical analysis.
Torvik's model has NU finishing 11-9 in the B1G, beating IU, Iowa and PSU.
Anyone know how or if they include injuries in their calculations? Like did NU’s computer expectations change just on Roper being out?
Neither KenPom nor Torvik (which uses KenPom's data and weights the last 10 games higher) account for that during the season. But you can eyeball it and see the impact of injuries. Rutgers losing Mag appears to be significant because he played substantial minutes and they don't go very deep.
KenPom isn't really complicated at all. It's a pretty simple concept - it just looks at margin of victory (or loss) adjusted for the quality of each opponent and compares that across all games (in a sense it's like a massive matrix). So winning and losing don't strictly matter, it's just the +/- differential. He also uses the same methodology to calculate measures for offensive and defensive efficiency (adjusted for opponent quality) and pace. He's always said it should not be used as a strict ranking system because he doesn't look at W/L, it is a model that is intended for future predictive purposes rather than to describe which teams have had the most successful seasons thus far.
You mentioned Liberty but skipped the even more interesting example - we jumped up to 51 with the win but OSU after the loss is still at 39... with an 11-13 record and 3-10 in conference (I think they've lost 9 of the last 10?). OSU has been a master of the "quality loss" this year against tough SoS and has some blowout wins to go with it. Their "luck" rating is -0.196, dead last of 363 D1 teams and well behind #362 (which is-0.152). That means that out of 24 games played, KenPom thinks that based on their overall margin of victory thus far (Pythagorean wins, if you know the concept) they "should have" won 4.7 more games, would be 15.7-8.3 if the outcomes of close games were random. After being chronically bad in close games for many years, thus far this year our "luck" is +0.066, 49th of 363 in the country, suggesting that we've added about 1.6 wins based on how we've finished out close games. Props to CC and the team for turning that around (thus far at least). Good FT shooting and defense, along with generally responsible ball handling down the stretch (last night excepted) contributes to that.
Btw if you want to predict how much we will move in KenPom after a given game it's relatively easy - take the difference in margin vs predicted (won by 6, vs expected to lose by ~5 pregame**) - in this case a differnence of 11 vs expected - and divide it by total number of games played. +11/24 means our rating (adjEM) goes up by about +0.46. Sure enough, we popped up roughly that amount, moved from 57 to 51. There will also be other smaller movements every day as other opponents of ours (and opponents of their opponents, etc etc) play and have positive or negative outcomes.
** To figure out the expected outcome, this isn't perfect as his actual projections vary slightly from using this simplified method (I assume bc the model isn't perfectly linear in incorporating opponent adjustments), but close enough for our purposes: (1) take the difference in team ratings per 100 possessions: OSU +15.65, NU +13.85 so 1.8, (2) multiply by the average of the team's pace ratings / 100: OSU 66.9, NU 65.3, so 1.8 * 66.2/100 = 1.2, (3) add home court advantage adjustment of 2.5-3.0. If we were to play them again tonight after the win, that would give you a spread of -4 (it opened at -5 yday and I think was -5.5 before tipoff, with the -5 being pretty much exactly what KenPom system would have predicted before last night's games).
I didn't consider what you wrote in the last two paragraphs. Super helpful, thanks.
I was under the impression that Sagarin operates exclusively this way but KenPom includes some adjustments for shooting, the luck number, etc. But I’m not a sharp, so I don’t really know, I just glance at them to understand point spreads.
The luck is essentially just the residual of Pythagorean wins (or expected wins vs your SoS) against actual wins.
It’s an output of the model’s process, not an input.
I don’t know enough to say exactly the differences between Sagarin Vs KenPom. But Sagarin blends 3 different methods - 2 different versions of pure points based and 1 that overweights recent outcomes. I think Sagarin doesn’t take into account number of possessions or offense/ defense splits? Also he displays his output in a different way - a score based rather than +/- from 0.
And my explanation above is a bit oversimplified. There are multiple ways to calculate a pure points based metric. Minimize linear errors, least squares errors, etc… Sagarin and KenPom are generally pretty similar though. Shooting %’s don’t have a direct impact on KenPom modeling (to my knowledge) - just indirectly through offense and defense points per possession for each game.
I’m pretty sure Sagarin is pure final points results while KenPom uses some other stuff to calculate a per possession based rating
I’m football terms, it’s like Fitz normally is, winning lots of one-possession games, and Frost, losing lots of one-possession games. NU until recent wasn’t as good as their record said, and Nebraska probably wasn’t as bad as their record said.
Ha did you read my explanation..? I think you just restated it in over simplified terms (and ignoring all nuance, which most of our world is prone to do these days). Thanks.
That article implies shooting percentage is part of it. A per possession metric to calculate values to compare vs a pure final score ELO type system is pretty different.
At any rate, I think we get it.
Yeah, I see the article says that but still don’t think shooting percentage is a direct input, that’s just an underlying piece of per possession outcomes for offense and defense? But I obviously haven’t seen the code so by no means am sure of that.
And yes haha. Sorry for the sarcasm last night!
Great thread. Right now, KenPom predicts that NU (8-5 at present) will win 2 more regular season games: (Iowa and Penn State), leaving us at 10-10. But, his prediction for us is to conclude the regular Big Ten season at 11-9. Why the discrepancy? Because he sets the sum of probabilities at 2.81 wins. The prediction for tomorrow: Purdue 68 - NU 63. Beat that spread and we will move up in kenPom. He now has us at #50 BTW.