The analytics debate in football has been raging for years and seems to escalate in tenor every season. To “traditionalist” football fans and media members, analytics strips out the “real world” elements of football, oversimplifying a complex game of emotional men and losing the essence of football in the process. To the more data-inclined, traditionalists just don’t want to give up the way the game is supposed to be played in the face of new evidence.
The Ravens have famously (or infamously, depending on who you ask) leaned hard into analytics in the last few years, to varying degrees of success. The second half of their 2021 season seemed at times to be specifically designed to reignite the analytics debate over and over again – from last second two point failures down one point (twice!) to early first quarter fourth and goal aggression. Fans have ranged from supportive to flummoxed to outraged at some of the high-profile “analytics or tradition” decisions the Ravens have made. The debate has been particularly fraught this season because the results of these decisions have on the whole not worked out as hoped, leaving the fans with the bitter taste of gambles that didn’t pay out.
Rather than wade into the debate, which is most certainly not lacking in takes and articles, I wanted to walk through some of the higher profile decisions from this year by examining what I see as fundamental differences in how analytics approaches football, as compared to traditionalists, that I think interferes with the ability of each side to productively debate key decisions.
To be aggressive or not to be aggressive?
First and foremost, why does a more analytics-focused football philosophy always lead to more gambles – more fourth down attempts, more two point conversions, fewer field goals and punts? The answer is that it doesn’t, per se – there’s nothing inherent to the process of building a win probability model that leads to a more aggressive strategy. Somewhere, on Earth 2, football traditionalists believe that punts and field goals are cowardly moves that make football more boring, and so the old school coaches go for it all the time – because no one in the league wants to be seen as the team coaching and playing scared.
That is, of course, not how traditionalists think here on Earth 1. Rather, for reasons we will explore below, traditionalist football often focuses more on minimizing risk and taking “sure things” (even if they’re not really so sure). Thus, analytics currently appears more aggressive because when traditional football and analytics differ, it happens to almost always be the case that models prefer the risk while traditional football prefers a more conservative strategy.
We can’t handle a loss!
Part of the reason analytics often seems more aggressive is that traditional football strategy tends to be willing to sacrifice some amount of win probability in service of other factors, particularly the experience of the game. Based on my conversations with traditionalist family and friends, one of the primary motivators that lead to sacrificing win probability is regret aversion. A well-documented phenomenon in behavioral economics, regret aversion bias involves making a decision based on a desire to avoid regretting a negative outcome from a different decision.
You could see this phenomenon play out in the discussion around the Ravens decision to go for two at the end of the Steelers game. Many fans would’ve preferred taking the extra point and seeing how things played out in overtime precisely to avoid what happened – the two point attempt failure essentially guaranteed the Ravens defeat. The prospect of failure is much more acute when it all comes down to one play. We, as human cognition machines, are naturally more sensitive to the prospect of losing than winning. Mathematical win-probability models, on the other hand, don’t share our sense of regret aversion – they value the benefits of success and the costs of failure equally. As such, these models will gladly embrace a decision that will lead to what we perceive as a potentially more painful loss if that loss is slightly less likely to occur.
So you’re saying there’s a chance…
Compounding regret aversion is the fact that humans just aren’t very good at thinking probabilistically. We have a very strong tendency to round probabilities to clean numbers – often 0%, 50%, or 100%. For instance, when discussing whether to kick an extra point, almost no one mentions the possibility that the kicker misses. The chance of that happening is, after all, quite low, particularly when you have the GOAT kicking for you.
Since the league moved the extra point back in 2015, the league-wide yearly failure rate has settled around 7%, while Tucker has missed less than 2% of extra points in that time frame.
For all practical purposes, Tucker is automatic on extra points, and we intuitively think that way (how often are you really paying attention when Tucker kicks an extra point). Yet, we also know from experience that even with Tucker, the chance is non-zero. Analytics factors in the possibility that a low probability event will happen in a way that humans don’t and won’t without extensive training and effort.
Thinking About What Happens Next
One final key difference has to do with how we think about future consequences of the decision. Humans have a tendency to think more discretely than models, experiencing football as a sequence of specific events rather than a continuous string of actions and reactions.
To illustrate this difference, let’s consider the decision by the Ravens to go for it on 4th and goal from the 3 yard line rather than kick a high probability (but not 100% guaranteed!) field goal. Many fans disagreed with the decision, given that it was early in the game and it wasn’t a particularly short fourth down.
When most fans consider whether to go for it or “take the points,” we think specifically about the possible outcomes of the play ahead – maybe we score a touchdown, or maybe we lose possession and the entire drive is wasted. We don’t want to risk the regret we’d feel from the latter, so we want the points. A win probability model, on the other hand, is able to factor in the possible outcomes farther into the future.
What happens if the Ravens do fail? Then Green Bay gets the ball backed up deep in their own territory, a position that doesn’t yield points as often. The model can account for the probability that Baltimore will get a stop and the ball back in prime scoring position once again – which is exactly what happened!
It’s worth mentioning that the models you see online don’t literally enumerate possible scenarios and assign probabilities to each of them. Rather, the models will in all likelihood rely on historical data in which such scenarios have actually played out, in order to account for how likely they might be to occur again.
The (data) revolution has arrived!
There are many other ways in which models approach the game of football differently from “traditional” game play, but the key difference that encapsulates everything discussed here is this: models don’t care about anything other than maximizing the long term chance of winning, and people instinctively do.
Analytics is also an acknowledgement that it’s not always easy to know how to maximize your chance of winning. Game management involves reacting to rapidly changing conditions to make critical decisions in a very short amount of time. Given how emotionally fraught these moments can often be, it can be quite the challenge for coaches to consistently make these decisions optimally. Teams that embrace analytics want to develop processes that make it easier for coaches to overcome common biases in decision making, such as regret aversion, during these moments.
Like it or not, data-driven decision making models are here to stay. Over time, the differences between what’s considered “traditional” football and what the models advise will slowly converge, as we’ve already started to see with fourth down decisions. Until then, at least we have something to argue about on Twitter!
