Once an integral part of actuarial modelling, the process of model pointing has become less essential as computer processing power has increased.
However, with current trends in the industry seeing a move towards Cloud-based infrastructure, model pointing offers opportunities to increase efficiency and reduce running costs, whilst maintaining the accuracy of results.
Model Pointing – Past, Present & Future
Model Pointing is the process of compressing the details of multiple policyholders, with similar characteristics and expected future behaviour, into a single line before running their information through an actuarial model.
In the past, this was necessary as the hardware and models available did not have sufficient speed and processing power to run calculations on every individual policy. Over time these speeds have increased dramatically, allowing ever more complex runs to be completed, including those at a policyholder level, in acceptable timeframes. As such, over time there has been less emphasis placed on compressing policyholder data.
However, the advent of Cloud-based computing has placed renewed focus on the granularity of this data, since reducing data volumes can reduce costs in two main areas.
- Model run-times: In Cloud-based environments, charges are incurred based on model runtime. For large insurers, the annual cost of running these models can run into the tens of millions in operational expenditure.
- Results Analysis: there are two main methods for storing and analysing results from a Cloud-based model
a) Results stored in the Cloud: users are charged for Cloud storage space. The more compressed policyholder data becomes, the smaller the results files will be, therefore reducing costs.
b) Results downloaded from the Cloud to on-premise servers: the greater the granularity of the policyholder data, and therefore the results data, the longer this process will take. Users will be charged according to the time taken, and this can become a bottleneck in actuarial processes.
Benefits & Use-Cases of Model Pointing
By increasing the compression of the policyholder data i.e. performing a greater level of model pointing, this will result in a reduction in both run times and operating costs, particularly for Cloud-based models.
The level of compression performed can be varied according to the run purpose – common use-cases for model pointing include:
- Business Planning
- Capital Planning
- Duration Matching
- Investment Decisions
Model pointing can also be used in reserving & valuation exercises. For instance, the IFRS 17 regulations specify that individual policies can be aggregated into groups (“Units of Account”) according to product, profitability status and year of issue.
Can this be achieved whilst maintaining accuracy?
When model pointing, a balance must be struck between increasing efficiency, compression of the data and retaining the accuracy of the results. This is possible using the latest techniques, such as those described below. As an example, for a large North American life insurer MBE Consulting have been able to compress their existing model points by 40%, whilst retaining over 99% accuracy of total cashflows over a 75-year projection period.
Model Pointing Techniques
Alongside the well-established algorithmic approach to model pointing, new machine-learning techniques such as clustering have been used successfully in actuarial modelling in recent years.
Algorithmic Approach
This refers to a methodical, logical process that is followed to group model points. This approach relies on a finite sequence of well-defined, computer-implementable instructions.
Techniques which can be used to further improve the efficiency and accuracy of the algorithmic approach include:
- Intelligent grouping: This involves the re-engineering of informative features from the policy data that can be used to distinguish between different policyholder groups. This typically involves combining or transforming raw policy attributes to create a new feature that captures the underlying risk profiles or characteristics. An example would be to group smokers and non-smokers together, with the addition of a ‘proportion smoker’ field.
- Rationalising segment boundaries: combining non-material segments of a policyholder variable with material segments.
Such algorithms are predictable and repeatable but can lack flexibility and require deep understanding of the problem to implement effectively. In addition, these algorithms can result in under-grouping i.e. model points that are too narrowly defined, resulting in disparate groups with relatively few policies in each group. This can result in increased computational time and resource requirements without necessarily improving the accuracy of the modelling outcomes.
Clustering
This is a method used in data analysis and machine learning to organise a data set into “clusters”. Data within the same cluster is more similar to each other than to those in other clusters.
This approach is a form of unsupervised learning, as it does not rely on pre-labelled data to form the groups. Instead, it identifies patterns and similarities in the data to determine the grouping.
Clustering can therefore reveal natural groupings and patterns in the data that might not be apparent upfront, making it extremely flexible and useful in uncovering insights and enabling informed decision-making. It is also much less prone to under-grouping (see above).
However, choosing the parameters required for clustering introduces subjectivity and can affect the results significantly – finding the optimal parameters often requires trial and error. Furthermore, the resulting groupings may not be explainable using actuarial logic and thus can be difficult to justify to stakeholders and auditors.
Pseudo-Clustering
MBE Consulting have pioneered a pseudo-clustering technique, which combines the strengths of algorithms and clustering to maximise compression whilst maintaining accuracy and auditability.
After applying an algorithm, the under-grouped model points are then clustered together with other, more-compressed model points. New machine-learning techniques are used to identify the most influential characteristic of the portfolio of model points. These are then used by the algorithm to systematically allocate the under-grouped model points to the more-compressed model points.
This methodology offers the advantages of an algorithmic approach, whilst also removing under-grouping with minimal impacts on model accuracy. The clustering element is also fully controlled and repeatable, and users have the flexibility to quantify what they consider to be an under-grouped model point.
Contact Us
As the industry continues to move towards Cloud-based infrastructure, model pointing offers a wealth of benefits. At MBE Consulting we have experience of transforming the model pointing processes of multiple insurers worldwide. If you’re interested in how we can improve your efficiency and reduce costs, without impacting accuracy, contact us today.
