Will Georgia Bulldogs Run It Back in 2022?
NCAA Week 0 - When last we left college football, Georgia Bulldogs defensive back Kelee Ringo picked off a pass floated by Heisman Trophy-winner Bryce Young in the waning minutes of a tightly contested National Championship Game in Indianapolis. Ringo cruised 79 yards down the sideline for a touchdown, an exclamation point to seal a 33-18 victory over the Alabama Crimson Tide and claim Georgia's first title in over 40 years. Eight players from Alabama and 15 players from Georgia were selected a few months later in the NFL draft, more than a full starting roster of talent lost between them. And yet, we head into the 2022 season confident that the Bulldogs and Crimson Tide are comfortably situated on the very short list of legitimate national championship contenders again this year.
Underclassmen Ringo and Young were not among those that joined the professional ranks this offseason, nor were Alabama all-everything outside linebacker Will Anderson and championship-winning Georgia quarterback Stetson Bennett. The rosters turned, but the cupboards are not bare.
The Ohio State Buckeyes are a third consensus contender coming into this fall, led by Heisman finalist quarterback CJ Stroud and featuring a wide receiver room that might make some NFL teams envious. And we're certainly not going out on any sort of limb with our forecast of these teams at the top of our projections. Vegas lines imply a more than 75% likelihood that the eventual national champion is among this group of three elite programs.
This confidence is certainly influenced by the accolades of specific individual players back on the field this fall, but it's the collective talent that truly separates these squads from the pack. Their rosters are loaded in spite of heavy turnover to the draft in any given season. They are the only three programs to sign a top-five class according to 247sports in each of the last three recruiting cycles (and, surprise, they're each on pace to make it four straight). And they have simply proven to be consistent winners, year in and year out, which means quite a bit in forecasting future outcomes.
It is tempting to think that college football, with almost constant roster turnover, might be susceptible to wild swings in performance among teams. But we know that's mostly not the case. How good your team was last year actually correlates very strongly with how good your team will be this year.
Previous year data is a solid starting point for projecting the season ahead, and data over several seasons is even better. As I do every offseason, I ran a recent analysis to understand the relationship between two-year data, three-year data, etc., and next-year FEI data. Though data from five years (and an entirely different roster) ago may not intuitively inform the projection model, I've found that including it does make the model stronger. Program FEI (PFEI) ratings are produced from the results of games and possessions played in each of the previous five seasons, significantly weighted for the most recent results. And with one notable season exception, PFEI ratings have consistently outperformed any other model projections I have tested in terms of correlation to next-season FEI ratings.
The 2020 season was an outlier for many reasons, but it remains a challenge to understand how best to use (or discard) the data from that season. The PFEI model I'm using this year doesn't remove the 2020 season entirely from each team's five-year history, but it does mute the results of 2020 more significantly than it had before. For many teams, performances in 2021 "bounced back" to levels more closely aligned with 2019 data than to the year in between. The 2020 "bad data problem" is somewhat attributable to the strange circumstances of the 2020 games themselves (COVID absences, crowdless stadiums, etc.) but I suspect much more attributable to an extremely disconnected network of teams, and the impact of that on effective opponent adjustments. I'm not going so far as to say that the 2020 FEI data is completely worthless, but I don't trust it to carry much weight in the projection formula going forward.
A projection model built exclusively from prior performance data admittedly has blind spots, and those blind spots might be more pronounced (or even severe) in 2022. NCAA transfer eligibility rules changes have created an environment of remarkable player movement that is unaccounted for in the PFEI formula. Many star players that popped at lower-tier programs in 2021 seized the opportunity to level up to traditional power programs this fall, and it remains to be seen how much of an impact such moves can make in one offseason. I'm certain that there will be extraordinary differences between projected ratings and final ratings for some teams, but I don't think that will be the case for many. And I'm optimistic that on the whole, the PFEI model will continue to be an effective starting point for the season ratings. From 2012 to 2021, PFEI ratings correctly projected the winner of 80.1% of all FBS vs. FBS games in Weeks 0 and 1; meanwhile, the closing line per covers.com correctly projected the winner of only 78.6% of those games. As in the past, preseason projected data will be filtered out as new game data rolls in.
Even with a model that reliably evaluates team strengths, projecting individual college football game results is hard, of course. FEI projections were one of only 10 college football rating systems tracked by ThePredictionTracker in 2021 to have finished the season with an FBS vs. FBS winning percentage against the midweek spread better than .500, and it barely cleared that very modest bar. Translating a retrodictive rating to a predicted score in a head-to-head matchup is not a simple exercise. And as was visually made clear to me in my major offseason project to design and publish 15-year game rating histories for all teams, very few programs actually play consistently week in and week out. One of the insights drawn out of that game ratings analysis is that defensive efficiency is a somewhat more reliable predictor game-to-game (and year-to-year) than offensive efficiency. That's true in part because the difference between a good and an elite offensive performance can be magnified in terms of multiple points produced per drive, but the difference between a good and an elite defensive performance is often more likely to produce differences of only fractions of points allowed per drive.
I have attempted to incorporate those insights into this year's projected wins model and individual FEI game projections. My preseason win totals do have some substantial differences from Vegas win totals, and I'll be very interested to track the new model's successes and failures throughout the year.
I think it is important to note that the range of season outcomes (i.e., the likelihood to go undefeated against the regular-season schedule, the likelihood to win six or more games, etc.) is not produced from a range of outcomes of the strength of each team. Instead, it is the range of outcomes for a team with a given rating to win or lose games against the slate of opponents with their given ratings.
For example, the projections don't indicate that the UCS Trojans (PFEI No. 57) have a 1.1% chance of being much better than projected and winning 10 games. Instead, they indicate that the No. 57 team has a 1.1% chance of winning 10 games against its schedule without being any better than an average team. The Trojans are the biggest outlier in the chart above, with a PFEI mean wins projection of only six games but coming in with a Vegas over/under of 9.5 wins. I will not be surprised if USC turns out to be PFEI's biggest blind spot in 2022. Led by new head coach Lincoln Riley and infused with significant transfer talent upgrades at key positions, this team could (and perhaps should) make an extraordinary leap forward in 2022. But we'll see if that leap is enough to launch the Trojans into immediate contender status or if more modest results are forecasted correctly.
The first games of the season kick off this weekend, and I will post updated FEI ratings and analysis weekly here and on my site throughout the year. Whether or not Georgia, Alabama, and Ohio State outpace the field all season long, or whether new challengers emerge from unexpected places, we can't wait to find out. The offseason projects are in the books, so let's bring on a new season of teams, players, possessions, and plays to reinforce or challenge our assumptions once again.
2022 FEI Ratings (Preseason Projections)
FEI ratings (FEI) are opponent-adjusted possession efficiency data representing the per-possession scoring advantage a team would be expected to have on a neutral field against an average opponent. Offense ratings (OFEI) and defense ratings (DFEI) are opponent-adjusted drive efficiency data representing the per-drive scoring advantage a team unit would be expected to have against an average opponent unit. Strength of schedule ratings represent the expected number of losses an elite team two standard deviations better than average would have against each team's regular season schedule (ELS), the expected number of losses a good team one standard deviation above average would have against the regular season schedule (GLS), and the expected number of losses an average team would have against the regular season schedule (ALS). Ratings for all 131 teams are found here, and expanded ratings including projected win totals are found here.
10 comments, Last at 26 Aug 2022, 7:01pm
#2 by Brian Fremeau // Aug 24, 2022 - 5:43pm
Defensive coaching staff stability is stronger than offensive coaching staff stability (hiring/firing, promotion cycles, etc)?
Top end defensive talent doesn't jump early to the NFL as often as top end offensive talent (incrementally better roster stability)?
I'd be curious to hear other suggestions.
#6 by Aaron Brooks G… // Aug 25, 2022 - 8:18am
You don’t take it seriously because it’s not identical to the pros?
Is it the absence of tanking that bothers you? Fewer ghoulish owners?
I wonder if the phenomenon is defense is a better indicator of roster talent at depth? Offense being easier, the really great teams can have great defenses, too. You see some defense-first teams (Iowa, traditionally LSU), but not a ton.
#7 by Drivster // Aug 25, 2022 - 8:34am
I would think a variant of B, specifically on QBs
Quality of QB play is a major driver of oDVOA. Good, and especially elite, NFL QBs tend to have high levels of year to year consistency and to stay on the same teams for a long time. This keeps the oDVOA stable, or at least the trends smooth. It also indirectly keeps low oDVOA teams low by restricting the market availability of elite QBs. So today's oDVOA is a good predictor of tomorrow's oDVOA.
Where quality of QB play step changes (eg Brady to Bucs, Josh Allen's development) the oDVOA step changes too, but in the NFL these are relatively rare events.
In college the best you can really hope for is two years of elite play and then you're onto the next guy. Even for programs with the recruiting power of Alabama, finding the next guy is harder, and less predictable, than keeping the guy you have. So today's OFEI is less predictable of tomorrow's OFEI.
If the stability of NFL teams' QB production over time is the driver, then the big difference should be seen in the predictive accuracy of oDVOA vs OFEI, rather than in dDVOA vs DFEI.
#8 by Aaron Brooks G… // Aug 25, 2022 - 9:02am
I'm curious how FEI and DVOA matchup on their correlations.
This suggests DVOA does pretty well to wins but less well to future success, but the comparisons seem apples and oranges.
FEI has the advantage that college has more huge talent mismatches. Even the Jaguars employ professional players.
#9 by Drivster // Aug 25, 2022 - 12:04pm
I've always assumed DVOA to be objectively the better methodology because it has per play granularity, whereas FEI only per possession. As you say, FEI has the advantage that college is probably inherently more predictable than NFL because of the relative absence of parity mechanisms.
Said that, the data exist (for someone less lazy than me) to apply FEI and DVOA to the same data sets (NFL seasons) and directly compare the correlation.
@Brian, has that article been written?
It'd be funny if DVOA turned out not be statistically significantly better.