Let’s say that, early in the week leading up to Nov. 21, you got a tip that the Pittsburgh Panthers would have multiple players out due to COVID-19 protocols in their upcoming game against Virginia Tech (something that’s been all too common during the 2020 season). The Hokies had just gone 1-3 in four straight one-score games, but by the adjusted scoring margin, they played better than average in all four of them. The Panthers had been blown out in two of their previous three.
Let’s say you took Virginia Tech as a 3-point road favorite based on that tip and waited for the news to break — and break it did. Shortly before the game, Pitt announced that it would play down seven starters and 16 players total. The line shot up to Virginia Tech -6.5 or -7, depending on the book. You would have probably felt secure in this bet cashing since you beat the line by a field goal or more of closing line value. Virginia Tech also had its own COVID-19 absences (four starters), but even before the line moved, the Hokies were projected by power ratings to cover the initial 3-point spread.1
Who’s the most chaotic fictional football coach? | FiveThirtyEight
And so you settled in with ticket in hand, only to watch the Panthers absolutely blow the doors off of the Hokies in a 47-14 outright win. Wait, what?
It was a quintessential 2020 college football betting story. During a season that we weren’t sure should have been played, things have been rendered even more complicated for people outside the game, whether they’re just betting for fun or rely on it to make a living. Results like this go beyond tough beats and into a realization that, in yet another way, this year is not normal. The sport has been uniquely difficult to predict for those trying to find a reliable edge.
Even people who know what they’re doing are having weird seasons
Simply put: We were behind the eight ball way before the season even started, according to Parker Fleming, a college football analyst and economist.
The difficulty in handicapping this season goes back months if you consider the type of offseason many teams had — or, more accurately, didn’t have — thanks to the pandemic. Out of 130 Football Bowl Subdivision teams, 111 didn’t even make it halfway through their 14 allotted spring practices, and 40 percent, including bluebloods Alabama and Texas, didn’t get in even one spring practice. The strange offseason was just the first shock to the prognosticating system.
“Every good prediction is going to involve some kind of priors — a prior being some sort of belief or assumption or distribution of the way things should be,” Fleming said. “We know continuity in offseason work really matters. So without spring and with guys opting out and transfers and all this sort of stuff going on that’s unique to this situation and the vast uncertainty, there’s a huge shock to priors. And so that makes things really really hard because you can’t really think about the season the same way you’d think about a normal year.”
[Why Are We Still Returning Kicks In College Football?]
If you want to get even semi-serious about betting college football, a great place to start is to build a power ranking. There are some fancy ways to do it, but you can also just average a team’s ratings in a few of the predictive models already out there to make spreads of your own. One you could include is the SP+ rankings system created by ESPN’s Bill Connelly.2
Models like Connelly’s factor in opponent adjustments and tempo and are weighted in the early season by preseason projections that phase out as teams play games, which keeps wonky early-results from skewing a team’s rating. But models are perhaps uniquely suited to miss this season.
Using The Prediction Tracker, we tracked the most accurate systems that put out predictions from 2015-2019 to compare how they’re doing in 2020. We judged accuracy using the lowest mean absolute error, which tells us how far off a prediction is from the final scoring margin of the game, with anything below 12.5 considered exceptional, according to Connelly. Combining absolute error and performance against the spread gives us a full picture of a model’s accuracy — doing well against the spread with a high absolute error signals that a model is lucky as opposed to actually good.
Even the best prediction systems are struggling
Average winning percentages (straight up and against the spread) and mean absolute error per game for the most accurate* college football computer rating systems from 2015-19
|2015-2019 Seasons||2020 Season|
|System||Straight up %||% vs. spread||Mean Error/G||Straight up %||% vs. spread||Mean Error/G|
|Moore Power Ratings||73.0||51.0||13.13||70.8||49.0||13.96|
|Average of top 10||73.9||50.5||12.98||72.1||51.4||13.53|
On average, the systems are picking games straight up less correctly, and their absolute error is 0.55 points worse per game. So some of the best tracked models in college football predictions are just fine against the spread, but when they miss against the actual scoring margin, they really miss.3
“If you’re a good stats-based handicapper, you’re gonna start with a model and you’re gonna make manual adjustments afterwards because you know this quarterback’s out or this team’s checked out or whatever the situation is,” Connelly said. “You’re going to be able to manually adjust for those things, and that’s not what a model technically does.”
One such handicapper is Bud Elliott, a college football writer for 247Sports.com. He uses a composite power rating and marries it with his own knowledge of teams and players. A team-based model would know that Team A beat Team B by 40 points last week, but it wouldn’t know that Team B had barely practiced in the preceding 10 days because contact-tracing protocols are forcing quarantines and on and on. You have to do that as the bettor, but how exactly do you calibrate it?
[The Secret To Iowa State’s Improbable Season? Balance.]
“If you have a good model, and you know for sure who is out and who they’re being replaced by and you have a feel for that, then you should be able to adjust your numbers appropriately,” Elliott said. “I think availability of information this year has been important. One thing I’m not sure how to grasp this year is how much missing practice either for a given week or for a team that is going virtual training for the week or whatever or whatnot is impacted. It’s hard to quantify that. You need to be able to quantify those things.”
Football and other sports like hockey are consistently opaque in the injury information they issue, a problem rooted in the perceived competitive advantage that secrecy can preserve. The NFL at least forces teams to produce injury reports several times a week, and the league’s COVID-19 protocol reporting is generally good. But in college football, there’s nothing close to standard injury reporting even outside of a pandemic. That’s one of the reasons official inactive reporting runs the gamut of styles from …
This (and it wasn’t made clear which of these absences were connected to COVID-19):
“I put that disclaimer” about the season’s high level of uncertainty “on my column [in Week 2] like, guys, I really don’t know how this is gonna go,” Elliott said. “It’s been a little disappointing in some regard.”
Connelly also has a best bets column, in which he selects from games where SP+’s projections disagree with the lines by more than 3 points and also have lines around key numbers (13.5, 2.5, etc.). He did much better than 50 percent in 2019. Not so much in 2020.
“There is no system this year,” Connelly said. “I’m under .500 for best bets this year, and I’ll look at trends and do I have a good read on these teams and all of a sudden: no. It just really is a lot harder to come up with any sort of patterns or any sort of comfort level as far as the types of picks you’re looking for.”
So, what about the books?
The lines themselves are also performing quite poorly when compared with the actual results of the games. In fact, the mean absolute prediction error of the line itself hasn’t been this high (12.96) since 2002.
When you’re trying to beat the books, one way to do it is to beat the line to where it’s going to end up. You can do that by generating a power rating yourself for each team and seeing how it measures up against the opening spread for each game. If you do this on Sunday and your power rating says Team A is -7 against Team B and the opening line has Team A at -2, then you’d think you have an inside track at an advantage, all things being equal (no injuries or other issues). If you take Team A at -2, you have a better number than where you believe the line will end up. This is the concept of chasing closing-line value: When you try to beat the line, you’re also trying to beat the book to where the line might move by kickoff six days later.
Ed Salmons, the vice president of risk management at the Westgate SuperBook, laid out his bookmaking philosophy pretty simply: The best bookmaking is about getting to the closing line first. That’s all well and good in a normal year, but this is not a normal year — and playing closing-line value this year is a dangerous game.
“It’s so hard to be betting early in the week when you’re fighting for a number that you think is good, and then your team’s got some guys at the important positions that have [COVID-19] and you’re kind of screwed,” Salmons said.
[How Coastal Carolina Invaded The Top 15 (And Our Hearts)]
The data backs that up to a staggering extent. From 2015 to 2019, we found that the line got just 0.04 points per game more accurate between the opening line and midweek, then improved another 0.04 points by the time it closed. But in 2020? The line’s accuracy has increased by 0.52 points per game between the opener and midweek.4
This all tracks with something Salmons has noticed: The big plays at his book are now coming later in the week, and high rollers are waiting just like the rest of us to learn what COVID-19 will do to a team’s roster. But as the season has gone on and games have been canceled more frequently (instead of just postponed), some big plays will roll in on games that are at risk of not being played because of COVID-19. If the game gets canceled, the bet just gets refunded.
While the line itself is missing at a higher rate than normal, favorites are covering at a rate consistent with a normal year. Through Dec. 15, college football favorites were 256-258-9 against the spread, while NFL favorites were 92-116. The former is normal, but the latter is not.5
Lines can adjust in ways that algorithms can’t, and the books are keeping tabs on things a little bit more closely than regular folks — even if bookmakers like Salmons are using the same tools as ordinary bettors.6 Salmons says the handle (a fancy betting term for cash) for NFL games is roughly where it should be because the league has brute-forced its way through the pandemic without any game cancellations. The college handle is lower, but that’s largely because of the volume of games lost to cancellation every week.
But when it comes down to the big plays in general, those are part of a broader story about how the pandemic has hurt Las Vegas’s bottom line. Salmons says he’s seeing fewer big players — and fewer huge plays.
“You just don’t have the tourism out here that you’d have in previous years,” Salmons said. “It’s such an ordeal now to get on a plane and come out here, and you got to wear the mask everywhere. It’s just harder to get those kinds of people into town”
It might be best to just write these oddities off as a 2020 thing
Even if you’re just betting for fun, it’s still frustrating to get to the end of a Saturday, having used a process you believe to be sound, only to be dealt another week of brutal results. Turning a profit — whether you’re trying to win real money or you just want to puff your chest about your record because you’re competitive — is hard, and this year made it much harder. But chin up: Maybe the tips will be worth something in 2021.
Neil Paine contributed research.