Georgia Tech Football: Advanced Stats - 2021 Predictions

This week, we’ve walked through how to evaluate teams and coaches, and so now, it’s time to further explore something we touched on while discussing team evaluation: projecting team performance. How do we take some of our ideas on evaluation and maximize their predictiveness as a holistic composite of team strength?

I find it most useful to double back to the idea of the “Five Factors of football”, pioneered by friend-of-the-program Bill Connelly, who lays them out as so:

1. Efficiency: ...examine(s) your efficiency and consistency in staying on schedule and putting yourself in position to move the chains. In terms of projection, it is by far the most important of the factors.

2. Explosiveness: Presented through Isolated Points Per Play (IsoPPP, which is unadjusted), IsoPPP+ (adjusted), and Marginal Explosiveness (see entry below). IsoPPP is the Equivalent Points Per Play (PPP) average on only successful plays. This allows us to look at offense in two steps: How consistently successful were you, and when you were successful, how potent were you?

3. Field Position: Presented through average starting field position (unadjusted) and FP+ (adjusted). This is mostly self-explanatory, with one important note: You should remember to measure an offense by its defense’s starting field position, and vice versa. Special teams obviously play a large role in field position, but so do the effectiveness of your offense and defense.

4. Finishing Drives: ...how frequently you create scoring opportunities, but how you finish the ones you create.

5. Turnovers: ...how many turnovers you should have committed (on offense) or forced (on defense) and how many you actually did. This tells us a little bit about quality and a lot about the Turnovers Luck idea defined above.

In a piece last year, we talked about some of the stats that Connelly uses to elaborate on these pillars in his advanced box scores, but we can also take those stats to build our own game evaluation model. The full details can be found here, but in short, based on some of the principles laid out in Connelly’s 2013 book Study Hall, we can evaluate teams’ performances in a game based on the Five Factors (compiling a “Five Factors Rating”, or 5FR), compare those performances, and then, — given those performances in aggregate and linear regression — determine who should have won the game (IE: the post-game win expectancy of each team).

What can we do with this information? Well, we can make generalizations about team performances during the season: teams with more second-order wins (AKA: expected wins, or the sum of their post-game win expectancies) than actual wins underperformed, while the reverse would mean a team over-performed and may expect to regress to the mean in the next season.

But if we’re sitting on a bunch of this in-game performance evaluation data, we can also use it to project future performances. After all, if post-game win expectancies are generated by taking the difference between aforementioned 5FR ratings, passing them to a linear regression model, and getting a projected points differential (AKA: margin of victory), then it would follow that if we can iron out a way to calculate a team’s 5FR at any point during the season, then we can compare two such ratings to generate a projected point differential (and projected win expectancy). We can also adjust these ratings to account for strength of schedule, conference strength, and subdivision (IE: FBS versus FCS) and generate a more theoretically-accurate point spread.

You can find more details on all of these calculations and some older proofs of their validity (read: betting performance) here, but now we finally get to the main conceit of this piece: using all of this math to project Tech’s 2021 schedule. Here’s how things shake out:

Based on these numbers, Tech is projected to net 5.22 wins in 2021, which is right along the lines of my feelings on the schedule (IE: “there are five wins on the schedule if Tech can grab them”) and what most of our friends in the desert are looking at as well. Here are some other thoughts on these numbers:

1. The prediction algorithm (or model, or whatever you want to call it) considers all FCS teams to have a strength rating equivalent to the bottom 2% of FBS. This may undervalue Kennesaw State — a regular FCS playoff contender — a bit, but it also serves to underscore the notion that FBS teams should not be losing to FCS teams.
2. The algorithm gives Tech larger chances to fire missiles into the exhaust ports of the two Death Stars it faces than I would feel comfortable betting on.
3. Just based on gut feeling, Tech’s projected results against Duke, UVA, and VT seem too narrow. Those are all toss-up games where I think Tech has an advantage. However, the numbers in those cases do emphasize what we’ve been thinking for years: the Coastal is the Coastal, and more often than not, most (if not all) of the teams in the division are generally of similar quality across the board. Most of these divisional games are going to be toss-ups, and Tech’s success or lack there-of will come down to execution on day-of more often than talent differential.

Got thoughts on these projections? Sound off in the comments below!