GLM.ModelBuilding

Reading: Goldburd, M.; Khare, A.; Tevet, D.; and Guller, D., "Generalized Linear Models for Insurance Rating,", CAS Monograph #5, 2nd ed., Chapter 3.

Synopsis: This is a short read that goes over the life cycle of a predictive model.

Study Tips

Your goal here is to understand the life cycle of a predictive model and be able to briefly describe any of the stages. Make sure you review this material occasionally so you don't miss out on easy points.

Estimated study time: 1 Hour (not including subsequent review time)

BattleTable

Based on past exams, the main things you need to know (in rough order of importance) are:

Describe how to specify the form of a predictive model (usually a GLM).
Describe how to validate a predictive model.
Describe the stages of the life-cycle of a predictive model from beginning to build it through to monitoring and maintaining its performance.

reference	part (a)	part (b)	part (c)	part (d)
Currently no exam questions for this reading

Full BattleQuiz

Excel Files

Forum

You must be logged in or this will not work.

In Plain English!

The following steps should be performed in any predictive modelling project:

Set the goals and objectives
Communicate with key stakeholders
Collect and process the data needed for the analysis
Conduct an exploratory data analysis
Specify the form of the predictive model
Evaluate the model output
Validate the model
Translate the model into a product
Maintain/rebuild the model

Each point may be further broken down into sub-tasks.

1.) Set the goals and objectives

a.) Given the goals, what data is required and how readily available is it?

b.) What are the key risks which may occur and how can they be prevented?

2.) Communicate with key stakeholders

a.) Regulators/Legal Compliance: The modelling team needs to understand which variables are allowed to be used in each state.

b.) IT: Programming constraints and budgets may limit what can actually be implemented.

c.) Agents: It's important to make sure the new product is understood by those who will use it.

3.) Collect and process the data

a.) Often the most time consuming part of the project. Data is messy, it's necessary to spend time cleaning it, calculating implied values for missing data (if allowable), adding additional variables from other data sets.

b.) Split the data set into at least two sub data sets so you can test the model on data that wasn't used to build it.

c.) Define a model validation process.

4.) Exploratory Data Analysis

a.) Plot each predictor variable against the target variable to see if any relationship exists. These plots can help suggest appropriate data transformations.

b.) Examine the correlations between predictors. This may be as straightforward as plotting pairs of continuous variables against each other.

5.) Specify the form of the model

a.) What type of model is best? A GLM isn't always the answer.

b.) What is the target variable?

c.) What transformations should be applied to the variables?

d.) What link function is appropriate?

6.) Evaluate the model output

a.) Assess the overall fit of the model and identify areas for improvement.

b.) Assess the significance of each predictor variable in the model. Remove or transform if necessary.

c.) Compare the lift of the iteration to previous versions of the model.

7.) Validate the model using the test/holdout data sets.

a.) See GLM.Validation for more information on this.

8.) Translate the model into a product

a.) Are there any ambiguities in the risk classification?

i.) Can someone/something fall into two categories for instance?

b.) Are there any illogical scenarios? For instance, the insured increases their deductible (to retain more risk) yet their rates go up...

c.) Apply judgmental adjustments to account for factors included in the plan that weren't included in the product. Such as adjusting for a new discount or deductible which had no data available in the model data set.

9.) Maintain and rebuild the model

a.) Accuracy decreases over time as processes and behaviours change. The data used to build the model gets old.

i.) Can refresh the model with an updated data set (same variables) to reflect newer experience.

ii.) Can completely re-model from scratch on a new data set.