The Future of Group Underwriting

The key objective is to identify new features that can segment claim incidence better than traditional variables alone and rapidly incorporate them into underwriting and pricing. We accomplished this by developing a predictive model using a subset of historical Group LTD reinsurance data enhanced with third-party data. Since there is no single industry incidence rating manual, a model built only on traditional variables – such as age, gender, occupation, etc. –  was our baseline for comparison against models built on traditional and third-party variables. This third-party data model aims to reliably predict claim incidence better than the model that uses only traditional variables.

For this case study, Munich Re matched employer census data and fifteen years of historical claims to a U.S. consumer information database that includes certain demographic attributes. Unique identifiers such as name, date of birth, and state of residence were used to perform the match for a final match rate of approximately 75 percent. The market offers many tools that can be leveraged to match individuals to third-party datasets, including those that return sensitive information in an aggregated, de-identified form, which can work seamlessly for the Group market.

Feature selection is the process of reducing a large number of variables available down to those few that are the most predictive for the goal of interest. For example, fields with many missing values are typically a challenge to use in any model. Users of the final model may also prefer predictor variables. The relationship of the variable to the outcome is easy to interpret, as this helps build confidence in the reliability of the model’s output. While the application of third-party data to predicting claims is new in the Group Underwriting space, variables found to be useful in other lines of business are an excellent place to start. Munich Re is constantly evaluating new data sources and their value for mortality and morbidity risk assessment and applying that knowledge across multiple lines of business.

Our goal for this case study was to understand additional drivers of claim incidence beyond the traditional variables. We emphasized retaining only a small number of novel predictor variables during feature selection. Working with a small subset of features facilitated understanding each feature's impact in isolation while also controlling current pricing factors. The novel features applied in this study include data that is generally not found in a census. The variables that prove most valuable will vary by carrier, target market, and distribution. 

When building predictive models, it is best practice to experiment with various models and configurations to determine the appropriate approach. We considered both parametric and tree-based regression methods for this study. Parametric models such as generalized linear models (GLMs) are based on distributional assumptions about the data. The modeler must predefine certain relationships expected in the data, such as variable interactions or non-linear relationships. Tree-based methods are helpful when there is no prior knowledge of variable relationships. These models learn patterns directly from the data, with the tradeoff generally requiring a more extensive training dataset for increased accuracy.¹  Since the goal is to incorporate novel features in claim incidence segmentation, we selected a tree-based method (“random forest”). However, we have also successfully used GLMs in similar modeling exercises.

We designed the traditional and third-party features in this case study as the model inputs and the claim incidence as the model output. Finally, 70 percent of the available data was randomly selected to train the random forest, and the remaining 30 percent was set aside for later use in model validation. 

Validating the predictive model on data that was not used to build the model is crucial in any analytic project. This gives the user confidence in how well the predictive model will perform on new cases. We assessed the models by comparing actual to predicted claim incidence across age, gender, and industry groups. Figures 1 and 2 below compare the mean absolute error of the traditional data model to the third-party data model across the different age bands for men vs. women. The third-party data model produces a lower error rate across most age bands for both men and women. Thus, our third-party data model met our goal and predicts claim incidence better than models with traditional variables alone. Using only the limited data available to a reinsurer and readily available external data, we built a model that better predicts disability incidence than traditional data. These insights can be used to more accurately estimate claims and set better premiums for key risks. 

A different way of visualizing this information is to compare the actual claim experience to the traditional vs. third-party data models' predicted claim incidence. The plots in Figures 3 and 4 compare the A/Es using the model with traditional variables as one expected basis to the model with traditional variables and third-party data as the other expected basis across different age bands for men vs. women. The A/Es from the third-party data model are closer to 100 than the A/Es from the traditional data model, which coincides with the lower error rates. On the other hand, the A/Es from the traditional data model are slightly closer to 100 for men under 40 and over 60 years old. Views such as these are crucial to understanding where the models perform well and where they do not in order to apply the insights appropriately and target groups properly.

A predictive model has limited effectiveness if it cannot fully integrate into existing workflows. Incorporating third-party variables can allow the carrier to identify specific cohorts that perform better or worse and price the case more accurately. An automated pricing solution developed to capture the employer census, append third-party data and perform calculations can return the appropriate load within minutes of the received census. Integrating with external vendors can take many forms. Munich Re can work with carriers on the easiest way to incorporate third-party data and model predictions into a pre-existing underwriting workbench. 

It’s important to monitor data hit rates and model performance over time to ensure the current behavior meets the expectations set during any training or proof of concept stages. Periodic re-evaluation is also necessary when consumer behavior or market conditions change to ensures consistent, high performance from the data and models employed.