Chapter 12: Introduction To Predictive Modeling in the Life Insurance Industry

Question 1

What is predictive modeling?

Answer 1

Uses statstics to predict outcomes using quantiative and qualitative data as inputs while the output is quantitative prediction.

Question 2

What are some qualitative factors that can be modeled?

Answer 2

1. Marketing
2. Sales
3. Underwriting
4. Servicing
5. Claims

Question 3

What is an example of a predictive model that we have used for years?

Answer 3

The Farmingham Model.

Question 4

What are major uses of predictive modeling for underwriters?

Answer 4

1. Application triage and requirement selection - To determine the type and degree of underwriting that is required for a particular case. Triage underwriting uses models akin to reflex testing in the context fof abnormal lab results. A model predicts what AAR’s are needed.
2. Propensity scoring - If more detailed info is collected then more complex models can be used. Uses large datsets in which labs are related to people with specific disease.
3. Approve/Decline Decisions - modles used to flag applications for immediate decline if they exceed a certain threshold
4. Mortality Scoring - Most difficult and complex. Application of models to directly predict mortality.

Question 5

What is the most common model used?

Answer 5

Triage Application.

Question 6

What are some of the benefits of predictive modeling?

Answer 6

1. Improved mortality and competitive pricing - Identify proposed insured who are worse risks than would be found through traditional underwriting.
2. Faster Case Processing - Low risk cases can be found early benfore more invasive and time consuming requirements are required. Can enable straight-through processing.
3. Lower Underwriting Costs - Identifying cases where requirements can be waived, cost savings can be obtained
4. Better Underwriting Utilization - Predictive modeling reduces the amount of UW requirements and take care of more cases of lower face amounts.

Question 7

What are some of the challenges of predictive modeling?

Answer 7

1. Data Availability - Building a model requires a large number of death claims. Using data from cases 20 years ago may not be relevent since the business mix has changed, UW guidelines have changed. Being able to find the data in an eccessible format will be challenging. You also need mortality outcomes on everyone (not just those applying for insurance)
2. Data Quality - Data can be missing or corrupt. Eliminating these will have a poor impact on the model. Missing data cases may be biased. If data is missing, you can use the mean, median, or mode.
3. Model Fitting and Subject Matter Expertise - Fitting the data occurs when a model is developed to accuratley predict the target for a particular data set, but its predictions do not continue to hold into the future. To show the data worked, it must be partitioned into 2 data sets - A build and validation

Question 8

What are “blind spots” in data?

Answer 8

Any use of a model in conditions outside of what is was buuilt upon should be done with caution, if used at all.

Question 9

What are the different areas of Predictive Model Development?

Answer 9

1. Data Sources for Underwriting - Data used in developing a model is usually identified and selected by considering several criteria (is there a logical hypothesis?, How ewell is the data correlated to the target? etc.)
2. Model Development Process - Once sufficent data has been obtained, the model can be developed.
3. Model Implementation and Monitoring

Question 10

What are the different steps to model development?

Answer 10

1. Feature Engineering - Variables are called features. Examples would be trasnforming DOB into age. A specialized approach to this is called principal components analysis which seeks to reduce the number of features when there are many correlated raw variables, that when comined are more powerful
2. Feature Selection and Model Development - A process to select the features to be used in the model construction is often done using a systematic approach that narrows it down to a manageable set. The data contains many features that are either reductant or irrelevant and can be removed without loss of information.
3. Model Evaluation and Validation - Once the model is selected, it needs to be evluated and validated. Concerned with determing the predictive power and relative error of a model.

Question 11

What are outliers?

Answer 11

Points in the data that require careful attention because they can have outsized impacts on a model resulting in skewed or biased results.

Question 12

Wha are common model types?

Answer 12

1. Linear regression - most frequently used to represetn business processes. Uses various inputs to predict an outocome that is continuous. Assumes there ia. linear relationship between each of the independent and dependent.
2. Logistic Regression - Also called survival modeling that results in a survival function
3. Decision Tree - divides a dataset into progressively smaller sub-segments
4. Random Forests - multiple decision trees based on different subsets of the data subsets of features, and outputs the mode of mean prediction of indiviudal trees.

Question 13

What are the different considerations of model implementation and monitoring?

Answer 13

1. Integration and end-to-end testing - Before a model is implemented it needs to be tested within the context of the broader framework to evaluat the “end-to-end” impacts that should be expected fomr its implementation
2. Reason codes for decisions - One of the challenges cited with predictive models is the “black box” nature of outputes. Reduced confidence by not knowin how the model arrived at a particular score.
3. Monitoring - Ongoing monitoring should be conducted to regulary anser questions.
4. Model hold-outs - Additional types of monitoring should be conducted if models are being used to waive requirements. This can result in anti-selection.