This is a project focused on predicting the outcome of the NCAA Men's Basketball Tournament using past regular season data. See report.pdf for analysis of data and final prediction results.
See adaboost.py. The properties of the Adaboost algorithm allow us to easily determine the most
important features to prediction with this simple ensemble method.
See cov_plot.py. A simple exploration into the correlation of our feature set.
See ensemble_serial.py. This script uses the ensemble methods trained in adaboost.py and
random_forest.py to predict the 2017 NCAA Tournament in serial. Meaning, the games are predicted
in their tournament format as if we were building a bracket.
See pca.py. We compute and plot the two largest principal components.
Surprisingly, 99.8% of the variance in the data is account for by these two components.
See random_forest.py. With a bit more extra computation, the feature importances
can be calculated using the average out of bag error of each feature. Luckily,
Sklearn implements this functionality.
See rd_experiment.py. The Glicko-2 rating metric assumes a team's true skill is normally distributed.
In this experiment we explore how variable the average Glicko distribution becomes when considering
data from further back in time. We conclude that it is actually more likely using as much data as
possible---and therefore data from further in the past---is more beneficial.
See spending_experiment.py. A simple experiment to see how well correlated the
all-time rating (calculated with Glicko-2 beginning from 1985) of the top 25
teams is with Men's Basketball spending. We remove a few outliers through Tukey fences and show, unsurprisingly,
a correlation of r = 0.6058 between Glicko-2 and 2017 Men's Basketball spending.