Research

Publications

1. Transfer Learning in the Actuarial Domain: Foundations and Applications
   Kim, Y. and Bauer, D.
   North American Actuarial Journal, Forthcoming
   
   Abstract

   With increasing data availability, the use of machine learning methods has gained popularity in insurance. Applications include novel areas for the use of models, for instance for automating business processes, as well as conventional actuarial prediction tasks such as claims or loss prediction. However, with the limited amount of labeled data due to claims being a rare occurrence, the superiority of advanced learners---particularly deep neural networks that have led to major advances in other domains---remains unclear. In other fields, *transfer learning* has been proposed as a potential solution in similar contexts. Transfer learning refers to taking the knowledge from one problem and applying it to a new but related problem, which can reduce the cost of collecting additional labeled data and improve the model performance. In this paper, various transfer learning approaches are introduced and applied to publicly available insurance data sets. The performance of each approach is evaluated compared to a baseline model in the context of predicting insurance claims. The results highlight transfer learning as a useful tool for the actuarial toolkit.

Working Papers

1. Insurance Agency Locations in Milwaukee: Insights from New Data and Agency Differentiation
   Kim, Y.
   Submitted for publication
   
   Abstract

   To assess the persistence of differential access to insurance for minority populations, this paper revisits the analysis of insurance agency locations in various neighborhoods of the Milwaukee metropolitan area, drawing on earlier work by Squires, Velez, and Taeuber (1991). Consistent with the previous study, we find a statistically significant relationship between the expected number of agencies and the proportion of the minority population when controlling for various factors. Additionally, the relationship between the minority population and the number of agencies is consistently negative and significant for both exclusive and independent agencies. Further analysis reveals no clear evidence that agencies are more likely to close in areas with high minority populations or to open in areas with low minority populations. Therefore, the negative relationship observed primarily results from the persistence of historical patterns. However, there is also no evidence that carriers are actively seeking to expand access in areas with high minority populations.

   
   

2. Cross-subsidization in the Personal Insurance Markets: Evidence from Different Business Lines
   Kim, Y. and Bauer, D.
   
   Abstract

   This study investigates the potential for cross-subsidization in the personal insurance markets, specifically between homeowners and auto insurance, in response to regulatory constraints and significant wildfire events. Utilizing a Difference-in-Differences approach, we analyze the impact of California's 2020 stringent regulatory policy and $1 billion wildfire events on auto insurance rates. Our findings indicate no statistically significant evidence supporting the hypothesis of cross-subsidization between homeowners and auto insurance rates. The results suggest that while higher homeowners insurance rates are associated with higher auto insurance rates, the regulatory policy implementation and significant wildfire events did not induce insurers to adjust auto insurance rates to compensate for constraints on homeowners insurance rates. These findings challenge the narrative of substantive climate-related cross-subsidization in the insurance industry and highlight the complexities of regulatory impacts on market dynamics.

Conference Presentations

2023

1. Persistence of Insurance Redlining
   At The 58th Actuarial Research Conference. Des Moines, Iowa.

2022

1. Transfer Learning in Actuarial Science
   At The 57th Actuarial Research Conference. Urbana-Champaign, Illinois.