Predicting Unemployment Status
My Winning Submission for the MEBDI 2022 Machine Learning Competition
I participated in the 2022 MEBDI Machine Learning Competition, and took 1st place.
Executive Summary for My Submission
I use a set of machine learning tools to predict next-year unemployment status for individuals in the Current Population Survey, using a subset of variables specified by the MEBDI competition judges. I use a Decision Tree algorithm, and two regularized linear regressions, and I improve the classification performance by combining these three classifiers using a hard-voting ensemble. This combined classifier is able to predict future unemployment in a holdout testing set with 70.95% accuracy and future non-unemployment (employment or not-in-labor-force) with 79.92% accuracy, for an average accuracy across classes of 75.44%.
The most important predictive features are a person’s current work/employment status. Essentially, the best signal that someone will be unemployed next year is that they don’t have a job this year.