Predictive Modelling for Transportation Security Administration Claims Data

Abstract

Click Here:Access Full Text

Travelers can file claims against Transportation Security Administration (TSA) if their baggage are damaged or lost during screening. After reviewing the claim, TSA will make the decision either approve or deny the claim. The data is published by TSA each year. It is an important data to understand the baggage damageloss, but it’s underused by both researchers and industry. This article explores the models with high accuracy and interpretability that can be used to predict whether a TSA claim will be approved or not. The data columns used in this research include claim type, site, claim amount, and disposition as well as airport code, airline name, etc. The clustering method is used to combine the levels in the factor variables such as airport, airline. We first used grid search and cross-validation methods to tune a single decision tree. Then a boosted tree is built. The generated linear models (GLM) with 3 different regularization methods are applied to predict the probability of claim approval: LASSO, Ridge and Elastic Net. The GLM with LASSO is chosen as the final model because of its great interpretability and high accuracy. The optimized cutoff probability to convert the GLM probability to claim approval/deny class is also discussed. This research is significant for insurance companies to develop travel insurance, for travelers to estimate their proper efforts to be invested in the claims, and for TSA to better understand the baggage loss and improve their management.

Keywords: GLM with regularization, Predictive modeling, Tree-based methods, TSA claims data.

View PDF