clock menu more-arrow no yes

Filed under:

Why computer models favor FSU in 11 games but predict only 9 wins

New, 184 comments

Computer models, humans, and probability.

Jimbo USA Today Photo

The span between the National Championship the following season-openers is long in college football; over 60 percent of any given year is the CFB off-season. Yet, the following it maintains, even builds, is incredible. It’s akin to what I deem the Chick-fil-A effect: High quality consumption, and they deny you access for a period of time - which causes you to want it more because A) somehow your craving peaks when it is closed and 2) that memory carries forward subconsciously.

As such, Bill Connelly’s preview piece for FSU is our Chicken Biscuit dream on a Sunday afternoon. Bill does an excellent of previewing all 128 FBS teams, and his 2016 preview of our fine fellows does not disappoint. 128 previews, by the way, is an insane undertaking, and we all benefit from it.

Before you get lost in the abyss that is that article’s comment section (wave at some past TN members, y’all!), here are some additional numbers on Bill’s numbers.

Favored in 11, but likely to win just 8-9?

One of Bill’s greatest achievements is the development of an FBS-wide probability model that even predicts football game scores. It takes into account game states, offensive and defensive efficiency, explosiveness, and other latent metrics that determine football game outcomes. Be a Bill fan.

In his preview, he shows his win probability of each game on FSU’s schedule. For instance, he has the Labor Day opener against Ole Miss as essentially a coin-flip with FSU’s odds of winning at 52%.

In fact, FSU is favored in 11 of its 12 games. This is a testament to zero-time ACC Coach of the Year Jimbo Fisher’s recruiting, coaching, and hiring ability as he enters his 7th year at Florida State looking for his next National Championship.

But summing each game’s win probability to give an estimate of the total number of wins on the season tells us a different story: Bill’s computer model says FSU is projected for 8.5 wins.

Why so few projected wins for Florida State?

Clearly, FSU is facing a tough slate of opponents:

Their schedule features six projected top-30 teams, three tossups (in this case, games with win probability between 48 and 54 percent) and six not-guaranteed wins (games with win probability between 64 and 75 percent).

That’s easy enough. But we also know that 8.5 wins isn’t that useful of a number. Football games result in a win or loss. Well, unless you’re Florida up 31-3 in the 4th quarter at Doak Campbell Stadium.

What are the odds?

If the cumulative win prediction is 8.5, what are the percent odds of getting 8 or 9 wins? 12-0? Or what about specific season outcomes like 11-1 with a loss against Clemson? 1-11 with a win against Florida?

Fortunately for us, Bill’s model gives game-by-game odds that let us answer that question using some ideas about probability. FSU is 52% likely to win against Ole Miss according to this model. Said another way: If the game was played 100 times under the exact same conditions, FSU wins 52 of those. Other models disagree with this probability projection, but arguendo, this is easy enough to figure.

What’s the probability of FSU winning it’s first two games under the model?

We use Bill’s numbers and conditional probability to help answer the problem.

FSU’s odds of winning it’s second game (CSU) is 99%. Thus, the odds of FSU winning its first two games is 52% times 99% = 51.5%.

Using this approach, we can multiply all of FSU’s per-game win probabilities to find their chances of going 12-0 given the individual game chances. This comes out to 1.21%, or 83:1 against. Seems kind of remote.

Looking just at favorites, FSU’s most likely set of wins and losses is 11-1 with a home loss to Clemson. If you wanted to see the odds of this, you multiply together the most likely outcome percentages. You multiply all the odds of FSU winning the other 11 games along with the odds of losing the Clemson game. The odds of FSU going 11-1 is the most likely result at 1.31%. Just sub in 52% for 48% against Clemson. And so, we see the odds of going 12-0 (1.21%) isn’t that much different than going 11-1 with a loss to Clemson (1.31%).

But that doesn’t mean going 11-1 is only 0.1% more likely than going 12-0. That’s because FSU could go 11-1 a lot of different ways. You could win 11 yet lose to Clemson. Lose at Louisville. Lose to a team that can’t recruit, you name it. So you’d want to add together all the conditional probabilities of the ways you could go 11-1 to find out that number.

Or, another way, using sampling. Just for fun, I built a regular-season simulator based on Bill’s model probabilities using R & Shiny. It even has a button! Based on a random number generator from a binomial distribution, I simulate each game’s outcome based on Bill’s game odds. The odd of that game’s outcome (“Odds”) is then listed, along with cumulative record and cumulative probability of that record to that point (“Cumulative”). Also, it’ll show you there’s lots of ways to 6-6 - but not all routes are equally likely!

8-4 and 9-3 keep coming up a lot, don’t they.

I’ll leave you with a histogram of 100,000 simulations of the above “simulator” to get a feel for the percentage odds of FSU winning a particular number of games.

As you can see, FSU’s most likely season record by the computer model is 9-3 at just over a 1-in-4 chance. Additionally, according to the model, FSU is more likely to go 5-7 then 12-0 and 38% more likely to go 8-4 against its regular-season foes than 10-2. For the record, 8-4 would match FSU’s worst regular-season record under Jimbo Fisher (2011).

Of course, very few if any humans believe this is going to happen, including Vegas oddsmakers and Bill Connelly himself, who thinks FSU loses at most twice in the regular season.

Because of the consistent level of success, consistent quality in recruiting, and amount of returning offensive production, FSU is projected fifth in the country. I've seen the Seminoles ranked pretty close to that in a lot of preseason publications. It makes sense.

FSU's projected rankings makes me marvel that much more at the Seminoles' projected win total. Their schedule features six projected top-30 teams, three tossups (in this case, games with win probability between 48 and 54 percent) and six not-guaranteed wins (games with win probability between 64 and 75 percent).

Two of the three tossups take place in the first three weeks. The Noles get Ole Miss in Orlando and Louisville at Papa John's Stadium, and a stark narrative might have taken hold by the time they head for another tricky test at USF: either this team will be undefeated and primed for a national title run or will be 2-1 or 1-2 and playing the role of Clemson Spoiler for the rest of the season.

The Seminoles are a high-ceiling, low-variance squad, and despite win projections, I have a hard time imagining FSU losing more than twice in the regular season.