Professor Bergman’s “NFL Survivor Pool” Plan Has Ramifications for Other Businesses, Too
OPIM professor David Bergman loves football, predictions, and data analytics, so it is little surprise that he would use his knowledge to plan an NFL survival-pool strategy.
His research article titled, “Surviving a National Football League Survivor Pool” with co-author Jason Imbrogno, a professor at the University of North Alabama, appears in the current issue of the journal Operations Research.
But the tactics and calculations that the OPIM professors have created carry implications far beyond the betting world. They can be used in many different complex real-world applications, including screening at airports and how kidney transplant recipients are determined.
Below, he explains some of his research findings.
Q: Start by telling us, in general terms, how an NFL “survivor pool” works.
A: The NFL regular season consists of 256 games, with the 32 teams playing 16 games each over the course of 17 weeks. The season revolves around a weekly schedule, which begins on Thursday and ends on Monday. During the season, each team plays one game per week and receives one “bye week” in which it does not play a game. The organization of the season by weeks has resulted in the development of “survivor pools,” where betting does not involve point spreads, and stretches over multiple weeks of the season.
For each entry into an NFL survivor pool, a participant must pick a winning team for one of the games that are played in a given week. If this team loses, the participant is eliminated from the pool. If this team wins, the participant selects a team to win in the next week, but cannot choose any team previously selected by the participant. This decision-making process is complicated, even more so when participants control multiple independent entries in a pool, as is often the case. The goal of a participant is to select for one of his/her entries a longer sequence of winning teams, starting in the first week, than any other participants’ entries.
Suppose it is early in the season and the best team in the league is playing the worst team in the league. Selecting the team that is most likely to win – in this case the best team in the league – is not necessarily the best choice, because you may want to “save” that team until later in the season. Finding a strategic way to select teams can substantially increase the chances of success.
A survivor pool can have thousands of participants, and each participant wagers a small amount of money, so winning results in a tremendous payoff. The American Gaming Association estimates sports fans wagered over $95 billion on professional or collegiate football in 2015. The complexity of the optimization, along with the substantial sports-betting industry, makes survival pools an exciting problem to study.
“The American Gaming Association estimates sports fans wagered over $95 billion on professional or collegiate football in 2015.” -David Bergman
Q: What strategy did you find most effective?
A: We collected data from the past 12 NFL seasons, and, based on that data, using different metrics that are freely available online, we were able to build a predictive model to estimate the win probability for each team in each game – starting for any point of the season and lasting throughout the season.
In any given week, we can deduce a relatively accurate prediction for the probability that any team will win in that week. For future weeks, however, the probabilities will change as the season progresses, sometimes substantially.
For example, consider the game in the last week of the 2013 season between the Houston Texans and Tennessee Titans. The Texans finished the 2012 season with a record of 12 wins and 4 losses (12–4), the third-best record in the entire NFL, while the Titans went 6–10. Expectations in Week 1 of the 2013 season was that the Texans would win this game, but as the season progressed, with the Texans winning just 2 of 13 games before playing the Titans, probability estimates for the Texans winning that game decreased substantially. The Texans ended up losing this game. As a result, you don’t want to plan a strategy in the first week of the season that optimizes over the entire season.
The strategy that we found is most effective is to plan for the upcoming eight weeks. On week one, a participant should plan selections up-to-and-including week eight, to optimize the probability of selecting winning teams in each of the first eight weeks. The participant should then pick the team selected in the plan for week one. After making this choice, should the team selected win, the participant continues to week two and then re-optimizes the choices, now with updated probabilities up to week nine. A participant should make decisions based on a “rolling horizon” approach.
We found that if you plan the whole season in advance, you’re saving teams that might not have high probabilities of winning later in the season. If you employ a myopic policy of considering only the current week and picking the team with the highest win probability, later in the season, the remaining teams available to a participant may all have low probabilities of winning. The eight-week look-ahead strategy allows you to balance long-term and short-term planning, substantially increasing the probability that an entry survives the entire season.
“The strategy that we found is most effective is to plan for the upcoming eight weeks…The eight-week look-ahead strategy allows you to balance long-term and short-term planning, substantially increasing the probability that an entry survives the entire season.” -David Bergman
Q: Do your findings have useful applications beyond sports betting? If so, please explain.
A: Absolutely! My main body of research lies in computational optimization, where I develop algorithms that assist with automated decision making. The decision-making process in a survival pool is an example of a sequential stochastic assignment problem. What this means is that at any given point, you, as the decision maker, must make a choice which limits and/or impacts what choices are available later on.
Other applications of sequential stochastic assignment problems include kidney allocation, airport security screening, and real estate bidding markets. For example, in kidney exchanges, known transplant patients await kidney donations. The process starts with the arrival of the first kidney, and a decision must be made as to who receives this kidney, without knowing in advance how compatible future kidneys will be for the awaiting transplant patients. With the goal of maximizing total welfare, there may be a specific patient that is a perfect match for this kidney, but if there is a good chance that this patient will match later on, it may be better to assign the kidney to another patient whose kidney compatibility is less universal.
These examples, and others, exist in many real-world applications, and we are certainly looking to apply the algorithms we developed in this paper to other application areas.
Q: What got you originally interested in football analytics?
A: Optimization is the crux of my research. I focus on identifying innovative ways in which we can automate decision making, both in business, and in other exciting, and surprising, application areas.
About five years ago, my colleague told me about NFL survivor pools and I started thinking about the decision-making process one might employ. I read websites that proposed methods for making these decisions, and I got to thinking, “OK, we can do this using some optimization!” I started studying the problem and I thought, “Let’s dive a little bit deeper and see what we can come up with here.” And it turned out to be a really fascinating mathematical problem, with broader impacts to other problems as well. We gathered data, studied the problem, and devised successful strategies.
Q: So this is something that’s very unique – has there been much research in this area before?
A: There hasn’t been any academic research specifically designed for NFL survival pools. As I said, there are some websites that do some planning, but the way in which they make their decisions is not necessary “optimal.” There is some related work on NCAA basketball tournaments, but the structure of NFL survivor pools leads to a very different problem.
What’s nice about the specific algorithm we came up with is that the user themselves, the participant, can come up with the probability estimates however they choose. They don’t have to use our probability estimates – they use the algorithm we propose for making optimal choices based on the probability estimates they come up with.
Q: How big an impact do you expect this type of research to have on business practices down the road?
A: I think we’re just starting to explore the opportunities that exist. One promising direction of research is the integration of predictive modeling with optimization. Traditionally, data analytics has been siloed into descriptive analytics, predictive analytics and prescriptive analytics. In the predictive analytics domain, researchers have been making great strides in devising advanced algorithms for predicting future events. In the prescriptive analytics domain, a lot of the work explores how you can automate decision making given a setting without randomness.
My research integrates these two domains. How can we make optimal decisions based on advanced predictive models? Automated decision making in an NFL survivor pool is just one example of the integration of predictive models with optimization. But there’s a great deal of unexplored research here, and I think this is what we’re going to see in the upcoming years.
“Automated decision making in an NFL survivor pool is just one example of the integration of predictive models with optimization. But there’s a great deal of unexplored research here, and I think this is what we’re going to see in the upcoming years.” -David Bergman
This year I’m working with other faculty at UConn, as well as researchers outside the university, on various ways of integrating predictive and prescriptive analytics. Industry is demanding methodology of this sort, and I am confident that in the future we’re going to see a lot of research in this direction, namely the development of data-driven predictive-model-based optimization.