Building A Winning Model
We are building a winning model but it will take some time. Instead of pretending to know it all right out of the gate, we want to utilize our unique data-driven approach to gain an edge on Vegas.
As we officially enter the 2024 season, we wanted to start a series of posts to track our progress and be very transparent in our starting point, our findings and tweaks throughout the process. It starts very arbitrary and over time we’ll mold it into something that can win (at least 54% of the time).
We are building a predictive model here at CFBDepth that we hope grows into the most dynamic and data-driven model available. And it won’t be that out of the box. We have a starting point for 2024. We will tweak the baselines and algorithms as we begin to see the results. We want to do this along with you.
If you take one thing from this introductory post, take this: Don’t bet your house on any of these suggestions (yet).
Keeping in mind that college football will always have a mind of its own and there will be results that simply do not make any sense. We all know this. It’s what makes our game great. We wouldn’t change that for the world (or, even a few bucks).
First, we need a name. The best we could come up with is “CFB Depth Model”. Sometimes simple is better. We do have a cool acronym though: (CFBDm).
And a logo:
What makes this model different?
This is a model driven primarily on our depth charts and playing time projections. That’s about as proprietary as it gets. These drive everything that we do at CFBDepth and this model will be no different.
The primary difference between other models is our reliance (or lack thereof) of recency bias. We won’t rely on past performance and focus on what is available at that time. The talent on the field. If a team’s talent level drops due to a key injury, we’ll have that accounted for in our talent ratings/rankings naturally from the depth charts. We are less reliant on an arbitrary hit to that particular squad - we can quantify that a little more effectively.
Key to this whole thing is our Player Ratings system. If you’d like to know more about that, we have you covered:
So, where do we start?
We mentioned arbitrary numbers. Well, we have to start with somewhat arbitrary numbers and multipliers in our algorithm. This is where we will make tweaks based on our performance. We are very transparent that our starting point is data-informed, but certainly not data-backed (yet). That’s what we will be doing throughout this season and documenting in this series. Maybe we crush it right out of the gate? Maybe not. Either way, this is our starting point.
Okay, data and depth chart-driven, so what else?
The additional outside factors that impact the result of a college football game are numerous. Some we are aware of (known), and others are very hidden (flu outbreaks). We’ll stick with the ones that we know and include them in our model:
Home field advantage
Weather
Travel times/distances
Look ahead spots
Hangover games
“Gauntlet” impacts
Preparation times (ie: short weeks or coming off a bye week)
Back to those arbitrary starting points. We have a decent idea of how “home field advantage” impacts a result. These other factors, well, we’ll just have to tweak and develop as we go along.
Okay, sounds cool. What else?
We will tweak for what we’ll call a “projected game script” for each game. We’ll tweak play calling, pace and other factors based on what we anticipate to happen.
Example: Georgia is playing UMass week 13. We will manually adjust the run/pass play-calling a pace for both teams based on our assumption that this has every makings of a complete blowout. We’ll do that on an as needed basis.Trench Focus. That’s right. Our model will take into account and added emphasis on the talent difference between the offensive line and defensive line. In college football, it is this area that can make or break a squad. This is often the “separator” between two very evenly-matched teams. So, we’ll use our compiled OL and DL ratings (again, based on playing time projections and their player ratings) and take that into account.
Specialists. We’ll also put a little bonus into the specialists involved. We track these very closely and we all know having a good kicker can make all the difference. Just ask Chuck Martin.
Similarly, a good punter and punt returner can affect the game with field position all day long. We need to take this into account. It is not the same as the “trench” difference; but it’s in there.
Note: We are only focusing on FBS vs. FBS games right now. We don’t have a great handle on the FCS and until we build out those full-on depth charts (2025?), we’ll stay away.
///////////// WEEK ZERO PROJECTIONS //////////
Now, that moment you’ve been waiting for. Our first official projections from the (CFBDm) are in. Reminder: Don’t put a lot of money on these. We have some work to do:
Florida State 27, Georgia Tech 20
Georgia Tech (+11.5) - 2 units /// UNDER 55.5 - 2 units
SMU (-27) - 3 units
We’ll only suggest plays based on how close we are to the totals. For example, we have the SMU/Nevada total at 57, just 1.5 points over the 55.5 total set. That’s not enough for a “suggested play.”
We will go up to five (5) units on our favorite plays.
Note: These are ALWAYS moving. They are very dynamic. If a player is deemed “out” you’ll see the impact in our projection in real-time (eventually, when we get this up on the website - hopefully by next week.)
Thank you.