Transforming Underwriting in the Life Insurance Industry

Marc E. Maier, Hayley Carlotto, Freddie Sanchez, Sherriff Balogun, Sears A. MerrittAAAI 2019 · DOI

Abstract

MassMutual researchers have developed a high-resolution "life score" using Random Survival Forests (RSF) trained on a massive 15-year dataset of nearly one million applicants. This system significantly outperforms traditional manual underwriting by reducing claims by 6% while simultaneously slashing policy issuance time by over 25%. By shifting from binary classification to survival modeling, the system provides a more precise, automated, and actuarially sound approach to mortality risk estimation.

Introduction

Traditional life insurance underwriting has long been a bottleneck in the industry. For decades, the process has relied on human underwriters manually reviewing health, behavioral, and financial profiles to assign applicants to broad risk classes. This approach is often inconsistent, slow—sometimes taking weeks to issue a policy—and lacks the granularity needed for modern risk management. The core challenge lies in mortality risk estimation: accurately predicting the likelihood and timing of a policyholder's death to price products effectively. While insurance companies possess vast historical datasets, translating this data into precise models is complicated by "right-censored" data, where the time-to-event is unknown for the majority of individuals who are still alive.

Methodology: Beyond Binary Classification

To address these challenges, the researchers moved beyond simple binary classification (predicting if someone will die) to Survival Modeling. This approach models the rate of mortality over time, which aligns more closely with how insurance products are priced.

Random Survival Forests (RSF)

The team compared standard Cox Proportional Hazards models with Random Survival Forests. RSF was selected for its non-parametric nature and its ability to capture complex, non-linear interactions between health variables—such as the "U-shaped" risk curve associated with BMI—that linear models often miss. The model estimates a cumulative hazard function for each individual, which is then used to derive a "life score."

Survival probability curves for different age and sex strata over 15 years of exposure, illustrating the model's ability to differentiate risk over time.

The Life Score and Cohort Normalization

The resulting "life score" (ranging from 0 to 100) is based on the 10-year cumulative hazard. Crucially, the score is conditioned on demographic cohorts (age, sex, and smoking status). This normalization ensures that a score of 80 represents the same relative health standing for a 25-year-old woman as it does for a 60-year-old man, allowing for consistent risk stratification across the entire business.

Data and Feature Engineering

The study utilizes one of the largest and most comprehensive application datasets in the industry, encompassing approximately 1 million applicants from 1999 to 2014. This represents 9.16 million "exposure years" and 15.7k observed deaths. The model incorporates nearly 60 inputs, including demographics, biophysical measurements (blood pressure, build), lab results (cholesterol, creatinine), and health history questionnaires. Significant data engineering was required to digitize paper archives and standardize questions that evolved over the 15-year period.

Key Results and Business Impact

The deployment of the algorithmic underwriting system has yielded substantial improvements in both risk selection and operational efficiency:

  • Risk Selection: The RSF model achieved a 6% reduction in claims within the healthiest pool of applicants compared to traditional human underwriting.
  • Operational Efficiency: The time-to-issue for policies was reduced by more than 25%.
  • Customer Experience: Faster turnaround times led to a >30% increase in customer acceptance for automated offers.
  • Scale: The system now processes 90% of applications, driving decisions for tens of billions of dollars in insurance benefits.
Distribution of Life Scores and their relationship to demographic and health cohorts (Age and BMI), demonstrating effective risk stratification.

Discussion and Future Directions

To ensure regulatory compliance and transparency, the system utilizes model-agnostic techniques like Shapley values to provide "contribution factors" for human underwriters. The researchers also addressed "non-stationarity" in the data, such as the downward trend in average cholesterol levels over the last decade, ensuring the model distinguishes societal trends from individual health improvements.

Future work will focus on integrating high-value data sources like prescription drug histories and motor vehicle records. Additionally, the authors highlight the ongoing need for biannual model refreshing to incorporate the latest medical trends and monitor for potential societal biases in model inputs, ensuring the system remains equitable and actuarially sound.

Summary updated

BibTeX
@inproceedings{maier2019transforming,
  author = {Marc E. Maier and Hayley Carlotto and Freddie Sanchez and Sherriff Balogun and Sears A. Merritt},
  title = {{Transforming Underwriting in the Life Insurance Industry}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019373},
}
Figure