Greetings and congratulations getting through the pre-championship bowl games. It’s been a wild ride. After starting a blistering 20-5, I limped into a still-not-too-bad 26-13 with ULL’s win over Miami of Ohio. There are two predictions left – James Madison and North Dakota State for the FCS title and two, perhaps unknown schools to college football fans, LSU and, let me check the spelling, Clemson, facing off for the FBS title. Clemson and North Dakota State are defending champions (and undefeated this season), but both are currently Vegas underdogs. Today’s Exploration is not about those games though. Look for two more G’s Expectations in the next few days.
…..
Before I explore the changes coming for the 2020-21 season let’s review the first 39 bowl games (apologies to NC A&T and Alcorn State the participants in the Celebration Bowl). The SEC has by far outperformed the other conferences and the Big 12 really stunk up the joint. Everybody else is bunched up in the middle. No other conference finished more than a single game above 500. No other conference finished worse than 2 games under 500. See the table below:
| Pos | League | Record | Pct | To Go | Teams |
| 1 | SEC | 7-2 | 778 | 1 | LSU |
| 2T | MTW | 4-3 | 571 | 0 | — |
| 2T | Pac-12 | 4-3 | 571 | 0 | — |
| 2T | AAC | 4-3 | 571 | 0 | — |
| 5 | Sun | 3-2 | 600 | 0 | — |
| 6 | Indies | 2-1 | 667 | 0 | — |
| 7 | Big 10 | 4-5 | 444 | 0 | — |
| 8 | MAC | 3-4 | 429 | 0 | — |
| 9 | ACC | 4-6 | 400 | 1 | Clemson |
| 10 | C-USA | 3-5 | 375 | 1 | — |
| 11 | Big 12 | 1-5 | 167 | 0 | — |
…..
This is the first of a number of pieces of analysis coming this offseason. I mentioned a number of times corrections in the formula for next year. The correction was to reduce an overstatement in the performance score. I also made an adjustment in the point differential score (completely hidden behind the scenes). If Ohio State beat Akron by 40 they would get the same benefit as Oregon beating Wisconsin by 40. Certainly there is an adjustment in SOS in those two fictitious examples, but the performance scores would be adjusted equally. I put in a modifier to lower the value when tromping on a weaker team and reduced the negative impact by being trounced by a superior team. Not a lot, but some The result of these changes, using 2019 data through championship week, was a re-shuffle of the rankings. Oklahoma moved from 21st to 4th, Clemson from 1st to 3rd. Most of the Group of 5 and FCS really good teams fell a few spots. Tough scheduled but bad performance teams dropped (example: #1 SOS South Carolina dropped from the low 60s to the upper 80s). I hope that made sense.
…..
The first cut of analysis on whether these changes were good or not had to do with the prediction model for the games. I looked at the first 1510 games of the season (high-level without splitting them into the decade ranges – that will be a future analysis and will include all 1555 games of the season, separated by Power 5 [at least one team], Group of 5 [no Power 5 team and at least one G5], and FCS [no FBS team], by each decade). Whew, that was a mouthful.
…..
This look compared the prediction model (same formula which uses the performance score as the major variable) for 2019 data, for this year’s and next year’s performance score formulas. Here is the table:
| Games | Correct | Incorrect | Pct | 90%+ Upsets | 80%+ Upsets | |
| 2019 | 1510 | 1133 | 377 | 750 | 1 | 7 |
| 2020 | 1510 | 1240 | 270 | 821 | 5 | 21 |
| Difference | 0 | 107 | -107 | 71 | 4 | 14 |
Using the same data, 2019’s formula generated a 75% accuracy score. That’s really good. Keep in mind, it’s also in hindsight. Overall, in real time games I had a .667 correct percentage (829 games) which is coincidentally my current bowl correct percentage. What I mean by hindsight, is, as regular readers may recall, the GCR doesn’t have a “history.” All games happen now. So, I just looked at them all. After the update, the correct rate jumped to over 82%, a significant increase. If I could pick games 82% of the time, life would be awesome. Again, this is hindsight. But, 7 more correct picks a week is a pretty strong argument that the updates improved the model. Keep in mind, I am a firm believer in the axiom that all models are wrong: some are useful. So, the first real test was the rankings made sense. The second test was the predictions improved. The other interesting piece of data was the impact on high percentage picks.
…..
In the 829 “live” predictions there were 124 that were 80% or higher. Of those 114 were correct. The 92% correct rate is good but while the high percentage games account for less than 14% of the total, the wins account for close to 21%. A full 64.5% of all games were less than 70% to win. I know there are a lot of close games, but almost 2/3 of them…in college football? I don’t buy it. My analysis by decades isn’t done yet, but there’s a bit of a preview in the table. In hindsight, there was only 1 game, that if played today (or after the 1510 games), the winner overcame a 90% or higher loss probability: South Carolina had a mere 9.2% chance to beat Georgia. That’s pretty small, but is there anyone who truly thinks this year’s Gamecocks would beat the Bulldogs one out of eleven times if they played again? On top of that, there were only 7 upsets in the 80% decade. That seems pretty conservative. Here are the games:
| 2019 | ||
| Winner | Loser | Pct to Win |
| South Carolina | Georgia | 9.2% |
| Georgia Tech | Miami | 10.9% |
| Cornell | Dartmouth | 11.5% |
| W Michigan | Ohio | 15.6% |
| Stanford | Oregon St | 15.0% |
| TCU | Texas Tech | 19.7% |
| Fresno St | Hawaii | 20.0% |
…..
With the new formula, the overly conservative predictions seem to be loosened. You can see now that there were 5 90% to lose wins including 3 listed as >99%. And there are now 21 upsets in the 80% decade. Sounds much more fun and realistic. Here are those games.
| 2020 | ||
| Winner | Loser | Pct to Win |
| South Carolina | Georgia | <1% |
| Valparaiso | Stetson | <1% |
| Cornell | Dartmouth | <1% |
| Delaware St | Bethune Cookman | 6.0% |
| Morgan St | NC A&T | 8.1% |
| Tennessee St | Austin Peay | 12.4% |
| S Alabama | Arkansas St | 12.9% |
| UNLV | Nevada | 13.1% |
| Marist | Davidson | 13.2% |
| Fresno St | Hawaii | 14.1% |
| Tennessee St | Tennesee Tech | 14.3% |
| Stony Brook | Villanova | 15.0% |
| Georgia Tech | Miami | 15.8% |
| Lamar | Sam Houston St | 15.8% |
| Jackson St | Prairie View | 15.8% |
| Northwestern | Indiana | 16.2% |
| Ga Southern | App State | 17.5% |
| UC Davis | San Diego | 18.0% |
| Tulsa | Central Florida | 18.3% |
| Abilene Christian | Nicholls State | 18.8% |
| Lafayette | Holy Cross | 19.6% |
…..
Some of the games in the first list aren’t even on the second. Maybe it’s just that I’m a geek, but this fascinates me. Clearly I have a lot more analysis to do before I’m completely convinced the changes are viable and improve the model. But so far, it’s 3 for 3. Much more to come.
…..
Please send comments, questions, and feedback. Thanks for reading and please continue to share with others. G
