If a life insurer wants to build a predictive model, how should they go about it? In this article, we explore the factors that need to be considered before beginning actual model development. We will do this by using the example of predictive models for improving persistency. (Improving persistency for a life insurer means increasing the volume of business they retain.)
The most important thing to do before building a predictive model is to define the predictive problem. This generally means finding answers to the following questions.
1. What is the universe on which prediction is required?
2. What will the model predict?
3. At what point in the lifecycle is the prediction required?
4. How long will the prediction remain valid?
5. How often will the model be run?
6. How will the business use the model prediction?
Notice that
a) These questions are interrelated, and
b) They are not specific to persistency models.
The universe of prediction means the set of entities for which predictions are required. For life insurers, the universe of interest will generally be one of proposals, policies, customers or claims. It may be further narrowed by additional conditions, such as all term policies from the Central region or all active policies due in the next month.
For persistency models, the life insurer’s universe of interest will generally be their active and lapsed policy base. However, other possibilities exist. For example, it is possible for an insurer to make a policy persistency-related prediction at proposal submission and use it – along with other underwriting and risk parameters - to make a decision about policy issuance. Doing this will require the universe of prediction to be all submitted proposals.
For a good predictive model, the universe of prediction needs to be as homogeneous as possible. Sometimes the universe of interest from the business point of view may include differently behaving policy portfolios. In such cases, the universe needs to be separated into different homogeneous subsets, and these subsets need to be analyzed separately.
For example, the business need may be to create renewal predictions for all active policies. But because of the differing behavior of policies with monthly payment frequency, the universe of prediction may be split into two subsets – active policies with monthly payment frequency and active policies with non-monthly payment frequency – and separate predictive models may be created for them.
This is an event that can be measured using available data (more precisely, it is some characteristic of the event: its presence, absence or number of times it occurs, etc.). It is generally a good idea to define the event to be as granular as possible, e.g., predicting whether the next payment will be done by the due date is better than predicting whether next 3 payments will be done by the respective due dates.
In the present case, the life insurer wants to improve persistency. Persistency is not a single event. Improving persistency is a combination of many different things such as:
a) Improving same month premium collection
b) Improving premium collection within the grace period
c) Reducing lapses
d) Improving the revival rate of lapsed policies
e) Reducing surrenders
Each of these can be defined as events for building separate predictive models. So, in practice, the insurer will likely need to deploy multiple predictive models to help with persistency-related objectives.
To ultimately determine a target event, we also need to define the associated time limit, such as premium payment within 90 days of the due date or lapse revival within 6 months of prediction date.
In theory, the later a prediction is made, the more accurate it is likely to be – since more information is available at the point of prediction. However, in practice, the point of prediction should be chosen so that the prediction is available to the business when they can effectively use it.
There are 2 broad categories of predictive models based on the point of prediction:
Models in which the point of prediction is coincident with an event in the policy lifecycle.
For example, prediction at proposal submission, prediction at policy issuance or prediction at the completion of 1 year from issuance. This type of model is generally used to score (i.e., make a prediction) individually for each policy or proposal when the defined event occurs.
Models which may be run at a given point in time.
For example, a model that runs at the start of every month and makes renewal prediction for all policies falling due within 90 days from prediction date. This type of model is generally run periodically and makes a prediction about all policies or proposals that meet the universe condition.
A large proportion of persistency models fall in the second category. Some examples are renewal prediction models, surrender risk models, and lapse revival models. However, it is not unusual to find a model in the former category, e.g., a persistency-related prediction made at proposal submission.
Every prediction is valid for a finite length of time. Generally, the validity of prediction is directly determined by the point of prediction and the time limit associated with the target event. It is important, however, to understand the validity of each prediction in order to use it effectively.
Consider a renewal prediction model that works on the universe of policies having their due dates in the next 90 days and predicts whether the renewal payment will be made within the grace period. For every policy that is part of this universe, the prediction is valid only till the end of the grace period – which means that this prediction does not tell us anything about the policy’s behavior beyond the grace period.
For a lapse revival model that works on the population of all lapsed policies and predicts whether they will revive within 6 months, the validity of the prediction is 6 months from the date of prediction.
The validity limit also needs to tie in with historical data availability for model training. If a prediction needs to be valid for 1 year, one would need to go back more than 1 year in the past. For example, consider a predictive model for 13th month renewal that predicts at proposal submission whether the 13th month premium payment will be made within 90 days from the due date. This prediction is valid for 90 days from the 13th month due date for each proposal, which means the period of validity from the date of prediction is approximately 15 months (approximately one year from proposal submission + 90 days). To train this model, we would need to use proposals submitted in, say, 2016-17 and study their 13th month renewal behavior in 2017-18.
Note that validity defines a time limit beyond which the prediction is not valid; however, it is possible to update the prediction before the end of the validity period by using the latest information. For example, a renewal payment prediction made at the beginning of January for a policy that is due in February may be valid till March, but the prediction can be updated at the start of February by using updated information from events in January.
The question of frequency of prediction generally arises for models which are designed to run at a point in time. Frequency of prediction needs to synchronize with business operations. It often helps persistency teams to have renewal predictions generated at the start of every month.
As an example, a scoring cycle (i.e., prediction generation cycle) which runs at the start of every month and creates renewal predictions for policies coming due in the next 3-4 months is generally helpful for renewal collection efforts. Scoring at the start of January can work on the universe of policies with their due dates in January to March and predict whether each policy will make its renewal payment within its grace period. Scoring at the start of February will focus on policies having due dates in February to April – which means predictions for policies due in February and March will be updated at the start of February, and the cycle will repeat.
A similar but lower frequency cycle can be used for surrender risk predictions. Predictions may be generated at the start of every quarter for the entire policy base that can be surrendered. Predictions can remain valid for a year and may be updated every quarter.
However accurate a prediction is, it is not useful if the insurer is not able to act upon it and gain from it.
Renewal predictions are generally used by the business to divide due policies into high/medium/low payment propensity segments and then optimize the follow-up with these segments. High payment propensity segments require minimal follow-up - a timely reminder may be enough for them. The reminder may be timed so that every policyholder is reminded individually based on the history of their premium payment times. More follow up effort is required for medium and low propensity segments to generate higher renewals from them.
Lapse revival predictions allow insurers to identify a small set of policies in their entire lapsed policy base which are more likely to revive so that the revival effort can be focused on them for maximum benefit.
If the goal is to reduce lapses, the insurer needs improved renewal predictions, not improved lapse revival predictions. Lapse revival predictions are made on the lapsed policy base. They can increase the number of revivals but won’t reduce lapses. Renewal predictions – if acted upon in a timely manner - can help improve renewals and therefore reduce lapses.
Acting on surrender predictions is a little tricky since the follow-up to reduce surrenders cannot be as straightforward as sending a renewal reminder. In this case, the insurer needs to engage with risky policies subtly and convince them to continue rather than surrender. To make this task manageable, generally, the insurer prefers an earlier warning about policies at the risk of surrender. Surrender risk prediction models will generally need to have a longer prediction validity, about 6 to 12 months.
It is a good idea to start on the actual development of predictive models only when all the above points are thought out thoroughly. Having answers to these questions means that the predictive problem definition is ready. One can then proceed to extract suitable historical data and subsequently to model development.
Interested in learning how Aureus can help you use event-driven processes to understand your customer's journey? Click on the link below to get more information.