|In 2008, Davidson College seniors, Colin Stephenson and Neil Goodson, used math to fill in their bracket and ended up ranking in the 100th percentile at a rank position of 834 in ESPN’s Tournament Challenge. Read about their experience below.|
Q: What class were you taking when you created your brackets? How did the idea of creating brackets with math algorithms arise?
Neil: The original research project came out of an elective course I took that focused on topics in operations research, which is an area of mathematics that focuses on the application of mathematics to solving complex problems in the real world problems. The class was a small group of graduate and undergraduate students, and we were all guided by the professor, Amy Langville. Knowing that Colin and I had an interest in sports, Amy encouraged us to conduct our research for the class in the area of sports ranking. Amy had already put effort in this topic as well as previous students, so we had tremendous resources available to us and were able to hit the ground running.
Colin: Our assignment was to use algorithms to solve real world problems. Amy recommended sports ranking models to us. It sounded perfect to combine one of our favorite sports, college basketball, and math.
Q: How did you break down tasks in your work?
Neil: The research project started in January at the onset of the Spring semester, so we had just a few months before March Madness began. Our research process required us to study existing methods, apply them to various past seasons and the current one, discuss results with our class and see how we can improve upon existing methods. Colin and I quickly learned to divide tasks to our strengths. I would spend time coding certain methods, and Colin would backtest previous year’s data. Both of use would scramble to present results to our classmates and professor each week. The class was structured so we could all brainstorm collectively on where to head next and that helped us move forward with our project.
Colin: First we wanted to understand current ranking models. Some were already being used in sports and others were being used for ranking things other than sports. Neil and I also thought of factors we considered to favor teams to go further in the tournament. We wanted to find ways to incorporate our own ideas as amatuer braketologists into our models. We decided to focus on weighting win/loss records depending on when they were played before the tournament. We both feel strongly that wins and losses in late February and March mean much more than those in November through January. The “hot” teams going into the single elimination tournament usually seem to go further.
Q: Did you create one bracket or several?
Neil: We created several brackets. We wanted to test various weighting schemes for each rating method. For example, we had several variations of the Massey method and several others for the Markov method. In total, I believe we tested over 30 brackets for that tournament.
Colin: We created 30 or more brackets. We also tested them against the 4 previous years’ results in the NCAA tournament.
Q: Can you describe which methods were successful? Did you have a sense of which would be most successful?
Neil: The most successful results were the methods that placed more weight on games occurring later in the season. Most sports fans would agree that this is a no brainer. What is interesting though is that we found that you can place too much weight on the end of the season as well. If you were to emphasize the conference championships in a model for instance, you probably would not do very well. So there is a trade-off between teams that have played well consistently throughout the season and ones that have positive momentum going into the post-season play.
Colin: The Colley and Ken Massey models that we weighted logarithmically worked the best for us, exponential weighting also worked well. We thought those models would work well because they were already used in sport ranking. We also thought that log and exponential weight would be best because the games closer to the end of the season get gradually more important than the last. They also did the best while testing previous years.
Q: What data went into making your predictions? scores? dates? anything else?
Neil: Our rating methods took into account each head-to-head match-up in Division I basketball, the point spread for each of those games, and when they were played. Strength of schedule also played an important factor for some of the methods. The major differences arose between the mathematical techniques used to rank the teams given this vast web of conference and non-conference match-ups throughout the season.
Q: What kind of excitement did you experience during the tournament? Were you ever on a leaderboard? What did it feel like to be in such a high percentile?
Colin: The tournament excitement was awesome. After all our preparation and work we were able to sit back and watch basketball for a couple weeks. When Neil went on NPR the morning before the first games he told them a couple upsets our models were showing. I think all the ones he told ended up happening. It was also great to go on national live tv on the CBS Early Show. We were live on the Davidson campus the night they were playing Kansas. When we let them know we had Kansas winning it all, then we got boo’d out of the building. The best models were in the 100th percentile on espn.com. They were doing better than any bracket I had ever put together myself. They were also beating all of my friends, so I had bragging rights with them. Kansas ended up winning in the last seconds of the championship. Neil and I went crazy when it ended the way it did, one last second missed shot and we would have been well out of the 100th percentile.
Neil: I have always enjoyed March Madness every Spring, but working on this project brought the excitement to a whole new level. After spending so much time in the lab crunching the data, I couldn’t help but constantly check how each model was performing when each tournament game ended. Since we submitted all of our brackets to the ESPN Challenge, we could instantly get a sense of how each stood compared to the 4+ plus other brackets out there. For most of the tournament, our best models were consistently in the 95th percentile and we ultimately finished in the 99.9th percentile with our best models. For me, it felt great to see the long hours of writing code, crunching data, and presenting research results payoff with winning brackets, but honestly even if we hadn’t been as successful, I would have enjoyed the project just as much. In that case there would have been so much else to try. I might never have wanted to graduate!
Q: Were you surprised about anything in the tournament? Were you surprised by how well or poorly certain methods performed? Were you surprised by the media attention you got?
Neil: Every year there are always upsets in the tournament, so of course some of those came as a surprise to me. I was also surprised at how well we did in picking the upsets. My feeling on upsets is that there are two kinds. Some upsets happen truly because some teams are less recognized in their ability throughout the season. Maybe it is because they are in smaller divisions or had a few notable losses and the pundits wrote them off. Other upsets happen because the best team had a bad day, but if they were to play the same team again, would probably win. I think the algorithms do a good job handling the first type of upset. I am not sure anyone can do well consistently picking the second.
I was definitely surprised by the media attention. When I heard that there may be some media interest in our story, I was thinking we may get a write up in the local paper. I was shocked when I had a voicemail from a producer at NPR and then the CBS Early Show.
Q: So far, no one has ever submitted a perfect bracket to the ESPN Challenge. Do you think this is possible, at least for a math algorithm?
Colin: I bet someone will eventually get a perfect bracket one day. It would take a lot of luck for them. I would like to think we could use math to get a perfect bracket, but it would also take a lot of luck. A lot comes down to the fact the NCAA selection committee puts together the bracket on Selection Sunday. The rest is about the unpredictabilty of the human element. The unpredictability is what draws so many people to watch the tournament.
Neil: It is just as possible for an algorithm as it is for any human being. Without a doubt, it will take a tremendous amount of luck for either. That is what makes March Madness so much fun.
Q: Have you tried making brackets in subsequent years? How did the methods do? Did you make any changes?
Neil: I have continued to use the models in subsequent tournaments and they have continued to do well. Well enough to win a pool here and there. I have been using the same methods we used in 2008. I would love to continue to tinker with them, but there is never enough time.
Colin: We have used our best performers every year since then. The following year my dad, uncles, aunts, brothers, coworkers all wanted a copy of the magical bracket. Of course I gave them out, and of course it failed miserably. The next year I kept it to myself, and I won my office pool. Last year I gave it out to everyone who asked, and it bombed again. So this year, it will be kept a secret again.