Volume 11, Issue 1 - Special Issue: Prediction methodology for the NCAA men‘s basketball tournament

Computing and machine learning advancements have led to the creation of many cutting-edge predictive algorithms, some of which have been demonstrated to provide more accurate forecasts than traditional statistical tools. In this manuscript, we provide evidence that the combination of modest statistical methods with informative data can meet or exceed the accuracy of more complex models when it comes to predicting the NCAA men’s basketball tournament. First, we describe a prediction model that merges the point spreads set by Las Vegas sportsbooks with possession based team efficiency metrics by using logistic regressions. The set of probabilities generated from this model most accurately predicted the 2014 tournament, relative to approximately 400 competing submissions, as judged by the log loss function. Next, we attempt to quantify the degree to which luck played a role in the success of this model by simulating tournament outcomes under different sets of true underlying game probabilities. We estimate that under the most optimistic of game probability scenarios, our entry had roughly a 12% chance of outscoring all competing submissions and just less than a 50% chance of finishing with one of the ten best scores.

Predicting the outcome of a single sporting event is difficult; predicting all of the outcomes for an entire tournament is a monumental challenge. Despite the difficulties, millions of people compete each year to forecast the outcome of the NCAA men’s basketball tournament, which spans 63 games over 3 weeks. Statistical prediction of game outcomes involves a multitude of possible covariates and information sources, large performance variations from game to game, and a scarcity of detailed historical data. In this paper, we present the results of a team of modelers working together to forecast the 2014 NCAA men’s basketball tournament. We present not only the methods and data used, but also several novel ideas for post-processing statistical forecasts and decontaminating data sources. In particular, we highlight the difficulties in using publicly available data and suggest techniques for improving their relevance.

Recently, the surge of predictive analytics competitions has improved sports predictions by fostering data-driven inference and steering clear of human bias. This article details methods developed for Kaggle’s March Machine Learning Mania competition for the 2014 NCAA tournament. A submission to the competition consists of outcome probabilities for each potential matchup. Most predictive models are based entirely on measures of overall team strength, resulting in the unintended “transitive property.” These models are therefore unable to capture specific matchup tendencies. We introduce our novel nearest-neighbor matchup effects framework, which presents a flexible way to account for team characteristics above and beyond team strength that may influence game outcomes. In particular we develop a general framework that couples a model predicting a point spread with a clustering procedure that borrows strength from games similar to a current matchup. This results in a model capable of issuing predictions controlling for team strength and that capture specific matchup characteristics.

We show that a classical model for soccer can also provide competitive results in predicting basketball outcomes. We modify the classical model in two ways in order to capture both the specific behavior of each National collegiate athletic association (NCAA) conference and different strategies of teams and conferences. Through simulated bets on six online betting houses, we show that this extension leads to better predictive performance in terms of profit we make. We compare our estimates with the probabilities predicted by the winner of the recent Kaggle competition on the 2014 NCAA tournament, and conclude that our model tends to provide results that differ more from the implicit probabilities of the betting houses and, therefore, has the potential to provide higher benefits.

The widespread proliferation of and interest in bracket pools that accompany the National Collegiate Athletic Association Division I Men’s Basketball Tournament have created a need to produce a set of predicted winners for each tournament game by people without expert knowledge of college basketball. Previous research has addressed bracket prediction to some degree, but not nearly on the level of the popular interest in the topic. This paper reviews relevant previous research, and then introduces a rating system for teams using game data from that season prior to the tournament. The ratings from this system are used within a novel, four-predictor probability model to produce sets of bracket predictions for each tournament from 2009 to 2014. This dual-proportion probability model is built around the constraint of two teams with a combined 100% probability of winning a given game. This paper also performs Monte Carlo simulation to investigate whether modifications are necessary from an expected value-based prediction system such as the one introduced in the paper, in order to have the maximum bracket score within a defined group. The findings are that selecting one high-probability “upset” team for one to three late rounds games is likely to outperform other strategies, including one with no modifications to the expected value, as long as the upset choice overlaps a large minority of competing brackets while leaving the bracket some distinguishing characteristics in late rounds.