Ch 12 Intro to Predictive Modeling in the Life Insurance Industry

Question 1

4 Benefits of Predictive Modeling

Answer 1

1. Improved Mortality & more competitive pricing
2. Faster case processing & less invasive underwriting
3. Lower underwriting costs
4. Better underwriter utilization

Question 2

Challenges w/ Predictive Modeling for Underwriting

Answer 2

1. Data Availability: Need large number of death claims/ data from 20 yrs ago may no longer be relevant/ data must be accessible, electronic format/ modeling targets/ predictor variables
2. Data Quality: missing or corrupt data/ inconsistent formats/ technical issues

Question 3

underfitting

Answer 3

model insufficiently represents real-world phenomena w/ its predictors being too far from actual data to be considered useful. Caused by insufficient amount of data, missing 1 or more key factors, or utilizing a model form that is too simple. Miss important nuances in relationship, making it less useful.

Question 4

overfitting

Answer 4

Model is developed to accurately predict the target for particular dataset but its predictions don’t continue to hold into the future
Divide into build dataset to develop and validation dataset to test the model. Difference between two show power of model. Worse validation performance is possible overfitting.

Question 5

blind spot

Answer 5

Missing data results in model that isn’t predictive for certain age bands or other segments of population since model had no basis in its build data set.

Question 6

rules engine

Answer 6

software tools that automate decisions, critical for flagging situations that can be rare but concerning. Paired w/ predictive models.

Question 7

areas that use modeling targets

Answer 7

marketing, sales, UW, servicing, claims

Question 8

applications of predictive modeling

Answer 8

Application Triage & Requirement Selection
Propensity Scoring
Approve/decline decisions
Mortality scoring

Question 9

3 steps to develop models

Answer 9

feature engineering/ feature selection & model development/ model validation

Question 10

feature engineering

Answer 10

variable used in model, created from raw variables captured in data collection

Question 11

principal components analysis

Answer 11

reduces number of features when there are many correlated raw variables, that when combined, result in much more powerful feature better correlated w/ target

Question 12

outliers

Answer 12

points in data that require careful attention, they can have outsized impacts on model resulting in skewed or biased results. Exist due to data entry errors, measurement errors, data processing errors.

Question 13

select minimum number of features that retain most productive value but are least correlated w/ each other

Answer 13

accelerates model development
improves interpretability
reduces overfitting

Question 14

linear regression

Answer 14

uses various inputs to predict an outcome that is typically a continuous number, most frequently used. assumes linear relationship between each independent/dependent variable

Question 15

logistic regression

Answer 15

similar to linear regression, except target variable is binary in nature (meaning yes/no). form of this called survival modeling.
Cox proportional hazards model: most used statistical technique for estimating individual risk in studies of survival

Question 16

decision tree

Answer 16

divides dataset into progressively smaller sub segments using features in rules. simple to understand and interpret, may not generalize well from training data, resulting in overfitting

Question 17

random forests

Answer 17

uses multiple decision trees based on different subsets of data

Question 18

validation

Answer 18

verify model is robust, not overfitted, relied upon to maintain predictive power into future

Question 19

Model Implementation & Monitoring

Answer 19

integration & end-to-end testing
reason codes for decisions
monitoring
model hold-outs

Question 20

UW Role in Predictive Modeling

Answer 20

• help build detailed rule sets based on guidelines
• provide insight into feature selection and engineering
• stress test models by asking questions
• help maintain and update models due to changes in risk, guidelines, requirements.

Question 21

Ethical & Legal Concerns w/ Predictive Modeling

Answer 21

1. W/ new data sources concern about use of data in underwriting context
2. Ethical concerns w/ using certain variables in model: what data is acceptable for use
3. Usage of predictive models, potential effects on consumer: if triage only no negative impact. If assigns risk/pricing, model can have adverse impact
4. Increased liability concerns w/ new data and technology to assign risk/pricing. Continuous monitoring and updates needed so decisions do not become inaccurate or invalid over time due to changes that make the knowledge bases and algorithms obsolete. (legal liability)