Who did you beat? How did you beat them? Does this team pass the eye test? All of these questions are a part of answering the much larger one…Does this team stand tall against the others when comparing resumes and deserve a spot in the College Football Playoff?

*If you are interested in the guts of our model, we have included that at the bottom in the section labeled “The SIS College Football Projection Model.”

SIS College Football Playoff Predictions

The SIS college football projections  model powers the SIS playoff predictions. We are able to simulate the remaining season, including conference championships, and tabulate a final standings at the theoretical end of the season. 

Conference championships are determined by the standings and tiebreakers that each conference utilizes, with some slight substitutes to specific ratings tiebreakers further down the list, instead using our own Total Points ratings.

Once the final standings are produced, we fill in the 5 automatic bids by ranking each conference winner by the most up to date CFP Rankings and Total Points rating if a team isn’t ranked. From there, we create an at-large pool that includes all 3-loss teams and better from the Big Ten and SEC and 2-loss teams and better from the Big 12, ACC, and independents. The conference championship loser is considered even if that loss pushes them over the edge of the bounds set prior. No other group of 5 teams are considered other than the 1 automatic bid.

Once the at-large pool is set, the teams are ranked based on the current CFP Rankings and are assigned a probability based on an exponential decay model where the remaining 7 entrants into the playoffs are sampled accordingly. This is to introduce variance when trying to predict what the CFP Committee will ultimately select. 

At the end of the day, the committee uses their set of criteria to determine the best teams without setting any rigid structure on ranking them. Therefore, it is nearly impossible to objectively capture the logic of the 12 individuals in a modeling capacity. This is why variance is introduced rather than standing firm on a rigid selection structure.

Additionally, with only one year in the books of the 12 team playoff, the sample size is too small in trying to capture past decisions made by the committee. That is why we didn’t try to model what the 12 teams look like at the end based on the final rankings last year.

The Lay of the Land Heading Into the Final Week

Ohio State and Indiana are on a collision course for the Big Ten Championship, with their two quarterbacks Julian Sayin and Fernando Mendoza staring at a Heisman dual in Indianapolis in a couple of weeks. Both of these quarterbacks are 1st and 2nd (Mendoza in 1st) in our Independent Quarterback Rating metric and On-Target Percentage in all of FBS.

Texas A&M and Georgia will both be tested against their respective rivals this weekend in Texas and Georgia Tech, respectively. Even with a loss these teams find themselves in prime position to make the playoffs, as they both have a greater than 99% chance of making it to the playoff.

The interest begins with North Texas, as we have them as our highest rated group-of-5 team ahead of Tulane, James Madison, and Navy. Drew Mestemaker has been nothing short of spectacular this season, leading all of FBS in passing yards and is 3rd in IQR. Our model predicts that more often than not this team will be in good shape to take the AAC crown.

The whirlwind ACC has come down to the wire, but the predictions like SMU to come out on top 63% of the time. A strong performance down the stretch has enticed the predictions to give them the best opportunity out of the ACC. Virginia and Pittsburgh both have a shot, but they will have to get right in some areas to have a legitimate shot at the playoff.

And then there are the Irish, which the predictions are lower on than most. Coming in at 9th in the CFP Rankings, the Irish need to hope that there aren’t any bid stealers (like BYU or Michigan) that come through and win their conference.

There is randomness that comes with a group of people coming together to select ‘the best’. However, we can use results on the field and the small set of criteria used to model the outcome of who they believe are the best teams. Besides, the speculation is what makes all of this fun, right?

The SIS College Football Projection Model

Our playoff predictions are rooted in our college football projection model, which outputs projected team totals, totals, spreads, and moneylines given various features from our robust data. These features include:

• Points For and Against Averages
• Team Snaps per Game For and Against Accounting for Garbage Time
• Offensive Penalty Yards Average
• Defensive Penalty Yards Average
• Home Team Designation
• Player Projected Snap Counts
• Passing Total Points Projection
• Rushing Total Points Projection
• Receiving Total Points Projection
• Pass Blocking Total Points Projection
• Run Blocking Total Points Projection
• Pass Coverage Total Points Projection
• Run Defense Total Points Projection
• Pass Rush Total Points Projection
• Kicking Total Points Projection

The team level metrics (Points For and Against Averages through Home Team Designation above) are averages based on the teams past 7 games and are recency weighted. The player level metrics (the remaining items from the list above) are per-snap averages of a player’s last 12 games and are also recency weighted. The projected snaps for a player in the game are then multiplied through a sample from a normal skewed distribution of each individual player’s projected Total Points value, and then aggregated at the team level before being inputted into the model.

The model is hierarchical, meaning that an output of one is an input into a proceeding one. All of the team based metrics are features for an XGBoost model that outputs a team level metric. The output is then used as an average and sampled via a normal distribution, with the residuals of the fitted model used to calculate the standard deviation of team performance. Once that number is outputted, it is then inputted as a feature into the next XGBoost model with all of the player level metrics, where the output is a team projected score for the game.

The opposing team’s defensive metrics are inputted as the defensive features of the model to balance the team’s offensive metrics.

Using Monte Carlo simulations, the games are simulated 1,000 times each. Therefore, a distribution is built with a team’s score, game spread, game total, and game outcome.