Model Madness 2019 - Conference Championship Results, Simple Composite Model

Today is Selection Sunday. It is the most important day in college basketball and the day when we finally get a chance to look at the overall bracket and begin using the models that have been developed over the past couple of weeks to pick winners and losers. We are first going to go over the results of the predictions of the models and then describe a quick and dirty final model - a simple democratic composite model.

So, here are the results of all of the conference tournaments. I made projections for all 32 tourneys with the EDM and SPM models, with the MASPM and EDLM models coming in for the later tourneys since they were developed later.

Results of Projections:

The first model that I built (EDM) ended up being the best predictor of individual games being the only model winning more than twice as much as it lost (.692). The SPM had the most perfect brackets with 3. The MASPM was the only other model to get a perfect bracket having 1. The important thing to note is that these perfect tournaments were typically smaller tournaments with very few upsets. Even though the SPM had the most perfect brackets, it also had the second worst record by percentage (.657). Our newest model, the EDLM ended up performing the worst (.654).

Now that we have these results, we can design our final model: A composite model. This model is the most simple of them all. We let the other teams vote on which team they have winning at a particular point in the bracket and the team with the most votes for a game progresses in the bracket. Pretty simple, but there are a couple of edge cases to consider.

The first is what occurs during a tie. We have four models that are voting members, so how do we break ties? We could select the higher seed as a safe pick. We could select the lower seed as an upside pick. We could use a fifth model as reference like the Elo Model I described earlier in this series. In our case, we will make the EDM model the priority model. It performed best during the conference tournaments so we'll give it a little more weight in the decision making process for our composite model.

The other edge case happens with the EDLM. This model doesn't generate scores but probabilities. With the other models, we'll simply pick the team with the higher rating. But with the EDLM we can't do that. We should pick the team of those available with the highest probability to that point in the tournament which may not match the selection of the pure EDLM since each round is selected by the composite model. We also need to fix the probabilities to the results of prior rounds when considering games in later rounds.

We could treat games as independent of each other to simplify the complex probability trees, but as mentioned in the last post, games are entirely dependent on the prior game in the single elimination setting. In either case, if the other three models are in agreement, we don't have to generate probabilities and if the EDM and one other model is in agreement, a calculation can be skipped since the EDM has the tie-breaker priority.

So, that is all for this post. A short one today. We are done with the development phase and are now on to the trial phase. The big tournament for prediction purposes starts Thursday afternoon. I'll try to post each model's predictions before then, maybe once a day. Tonight, we'll know the teams. I will also provide some potential upset picks by each model to help anyone look for a trendy pick to help them in their brackets. Happy picking!