Earlier this month, we introduced College Football Elo Ratings: our new take on the stat used to track the strength and head-to-head win probabilities of teams. Now we’ll use it to rank every team and give a prediction on how many wins they’ll achieve. Right now, we’re including teams that will not be playing in the fall, and we are looking at their schedule before the reduction in games. We will update this again once the schedules are updated in the database.
These 2020 preseason rankings can also be thought of as the 2019 postseason rankings, since the 2020 ratings are a regressed version of the final ratings of each team at the end of their 2019 season. Elo is a smart rating system, but not smart enough to know what’s happened in the offseason. That being said, if there were any major changes to a team’s roster or staff, it’ll take a few weeks for Elo to catch up and decide how that impacts their performance on the field. With that in mind, let’s take a look at the top 25.
2020 CFB Preseason Top 25 in Elo Ratings
Unsurpisingly, Clemson and LSU sit on top of the list going in to 2020. Clemson, even after their loss to LSU, were still rated higher than LSU at the end of the 2019 season.
The teams are ordered by their Elo, and as you go down the list, if you see a color difference between the Expected Wins column and it’s Elo Rating counterpart, it may indicate either a relatively easy schedule (light blue or orange) or a more difficult schedule (darker blue) compared to that team’s neighbors. For instance, Oklahoma has a lower Elo Rating than Alabama, but their Big-12 schedule results in 10.87 predicted wins, versus just 10.24 for the higher rated Alabama. This might tell us that Oklahoma has a slightly easier schedule, an unsurprising conclusion in this case.
On the other hand, USC were projected to struggle with their Pac-12 schedule (or what would’ve been) with just 6.92 projected wins while comparably-rated Washington has 7.95 expected wins, more than one extra win. Oklahoma State also leads USC by nearly one win with their Big-12 schedule, and UCF in the American is projected at a whopping 9.67 wins, despite having an Elo Rating nearly 50 points lower to start the season.
Non-Conference Strength of Schedule
Expected Wins are calculated simply by summing up each team’s win probabilities for all their regular season games. Once again, this was run with the old schedule, which is why you see Clemson at 11.54 Expected Wins, which is .54 more wins than games that they’ll be playing this season. As soon as the data is updated, we’ll rerun this and put out a new Expected Wins table. In the original schedule which included more out-of-conference games, the SEC had the highest average Expected Wins with 7.26 wins on average, followed by the Big 12 with 6.87, the Big Ten with 6.76, the ACC with 6.66, and the Pac-12 with 6.63. Their Elo Ratings also moved in the same order. The Mountain West was the top non-Power-5 conference in Expected Wins with 6.26. This can hint at two things: conference strength (SEC having more strong teams across the conference than the Pac-12 or ACC), and out-of-conference game difficulty (e.g. SEC scheduling easier non-conference games than the Pac-12). We can’t say for sure which is driving this value more, but both are plausible. Here’s a look at the non-conference opponent Elo ratings for each conference.
So it turns out the SEC does have a fairly easy non-conference schedule, while the Pac-12 have one of the most difficult ones.
Top Team’s Schedules and Win Probabilities
Looking at the schedules of the top teams, the toughest matchup by far for top-ranked Clemson was (and still will be with the updated schedules) against Notre Dame. This is their only game where they do not have greater than a 90% win probability. Instead, they have a 79.7% win probability according to preseason Elo Ratings. That game was originally slated to be played away and is still an away game right now.
LSU’s home game against Alabama shows them at a 65.5% win probability. This will likely change as the season plays out, as LSU is a big candidate to regress to the mean with Joe Burrow leaving for the NFL after a historic season in terms of QB stats. Remember, Elo doesn’t know that Joe Burrow left. Instead, Elo automatically reduced LSU’s preseason rating back down towards 1500 (by 5%) to account for any shakeup in the roster, coaching, and general year-to-year variation. LSU has three away games in the 80-percent win-probability range, with their toughest matchup coming against Florida in mid-October. Let’s look at the Top 25 with the Big Ten and Pac-12 pulled out.
Rankings without the Big Ten and Pac-12
So the top teams move up slightly, and some less-common preseason Top-25 teams make it in this year. Outside of the top 10 or so, the teams get fairly average, with Elo Ratings near the mean of 1500. This does not bode well for the competitiveness of this season outside of the top contenders. I wouldn’t be surprised if we end the season with several undefeated teams. And with a lack of non-conference games to weigh the relative strength of each conference head-to-head, it will be a difficult year to choose who gets left out of the playoff, even with the Pac-12 and Big Ten champions out of the picture.
A note on Elo Rating methodology
Since we posted the initial Elo article, we’ve made some slight improvements to the calculation that resulted in a Brier score of .172, compared to our previous score of .175. We’ve eliminated the tendency to under-predict actual results. We did this by adjusting teams’ Elo Ratings based on their conference. The Power-5 stays at an initial value of 1500, but the Group of 5 schools are now initialized at an Elo of 1200, and each season they revert to 1200 as their mean. All others, including other D-I and all D-II/D-III opponents, now get initialized at a value of 500. FBS Independents like Notre Dame get lumped in with the Power-5 at 1500 as well. We used 2019 as the baseline for what conference a team should fall into, since some schools have changed conferences since the start of the Elo Ratings in 2000. We’ll continue to analyze this and tinker with it to find the optimal balance when it improves predictive power.