Massimo Cavadini, Davide Burlon and Gili Smadja discuss how automated machine learning is changing the role of the pricing actuary
Pricing and underwriting represent a key pillar of personal lines insurance. Developments in technology and digitalisation have taken pricing sophistication to the next level, enabling many insurers to transition to a digital ecosystem. The birth of the automated machine learning (autoML) field is changing the role of the pricing department, placing it at the centre of the decision-making process and impacting the governance of insurance pricing teams. (Hereafter, ‘autoML’ refers to automated generalised additive models, or GAMs.)
The evolution of the role
There are three elements to property and casualty insurance pricing: data, prediction and decision-making. The improved quantity, quality and granularity of data supports the use of new algorithms for improving predictions. As technological advances enhance predictions for decision-making, the pricing actuary’s role is evolving.
We could describe the evolution of pricing across three generations (Figure 1):
Generation 0: When data was scarce and computational power tiny, a good pricing actuary interpreted simple model outcomes, reading the main statistical key performance indicators. Building a single risk model could take one or two weeks due to the lack of tailor-made software and slow computational time. Strong statistical knowledge and the ability to manipulate data were mandatory for running the code and interpreting results. Model complexity – mainly generalised linear models (GLMs) – was rather simple.
Generation 1: Proprietary pricing software made modelling easier for experts. The visualisation of the main statistical parameters and the structured approach to modelling significantly improved the time allocated to the process, allowing for the exploration of different models and testing of diversified approaches. The adoption of tree-based algorithms gave new life to the pricing exercise, encouraging the pricing actuary to blend traditional techniques (such as GLM) with more innovative algorithms. However, feature engineering, model development and selection still required a substantial manual effort.
Generation 2: The latest development in machine learning, coupled with a cloud-based approach, allows the machine to run and test different models in a short timeframe without losing model interpretability. These technological advancements are at the base of the autoML field, which is changing the pricing expert’s role. AutoML commoditises the prediction, allowing the pricing actuary to focus on decision-making and implementation.
AutoML demystified
In recent years, autoML has become a trending topic in industry and academic artificial intelligence (AI) research. It shows great promise in providing explainable and implementable results and allowing for greater access to machine learning’s efficiencies and model predictiveness.
Currently, each step in the pricing development pipeline, such as data processing, feature engineering, and model development and selection, must be done manually by the pricing actuary. In comparison, the adoption of autoML enables a simpler development process in which a smaller time investment can generate the models needed to enable actionable steps as a result of the analysis.
While the most disciplined approach to autoML keeps the actuary at the centre of the analytical journey, it suggests a future in which more efficient machine learning is a reality. In more detail, how can autoML be better described in the insurance context?
Without autoML, a pricing actuary has to start with a hypothesis, gather the correct dataset, try data visualisation, engineer extra features to harness all signals available, train a model using stepwise regression, and combine, apply and test implementations.
AutoML automates the model building process. From the actuary, it only requires model selection, hyper tuning and judgment, focusing on suitability of the results for the market. Think of it as a generalised search concept for finding the optimal solutions in an output that is understood. AutoML shows great promise in providing solutions for AI by offering explainable and reproducible results.
AutoML reduces time spent carrying out manual interventions while maintaining and even enhancing models’ predictiveness.
The actuary remains a key part of the process, from data collection and feature engineering to model selection. Interpretability and impact analyses are still the actuary’s responsibility. However, actuaries benefit from the clarity of these features because they make the high-quality models more compelling and actionable.
Benefits of AutoML
On one hand, shifting prediction to off-the-shelf algorithms removes the control that actuaries usually prefer to keep under their remit; on the other hand, it frees up considerable time, which can be used to address the challenge at hand and interpret the results of the machine, considering business requirements. Eventually, the time invested in the whole process might not differ dramatically from previous generations (Gen 0 and Gen 1), but the following are benefits that naturally emerge in Gen 2 pricing:
Model interpretability: Interpretability of model results and trends, either factor by factor or as a whole, is possible given the GAM approach. Additionally, it is possible to test the models’ outputs on stratified policyholder datasets, to see the impact analysis of new models against the incumbent.
Sufficient model prediction: The typical off-the-shelf algorithms used in predictive modelling come from a fairly distant past, if we consider how quickly the machine learning landscape has evolved during the past decade. In fact, traditional GLM techniques date back to the 1970s. One of the supposed limitations of Gen 1 approaches is that non-linear relations between the covariates and the response variable are less likely to be captured, even when using different sets of smoothing functions (splines, polynomials, step functions and so on) as with GAMs. In fact, when moving from GLMs to GAMs, the risk of over-fitting needs to be accounted for through, for example, cross-validation. AutoML naturally includes data segmentation and customisable k-fold validation rounds, since the additional computational resources required for the multiple iterations are built into the process. Generating a range of GAMs, given the dataset provided, the prediction will be at least as good as those developed with a full team of actuaries working on manually building GLMs using the typical stepwise regression methodology.
Reproducibility: The models generated by some proprietary solutions take the form of the tried and tested GAMs that most pricing teams rely on for their pricing models. This means that comparing factor relativities or model predictions in R, Python or any other software is possible.
Model implementation: Insurers rely on well-known policy administration systems and rating engines to implement pricing models. As a result, model output needs to be in a format that can be easily digested by these systems, whether this be rating tables, formulae or application programming interfaces. The autoML approach, using GAMs as the algorithm of choice, enables easy implementation of models in a variety of constellations and for a variety of client sophistication levels. This ensures that insurers go from modelling to live pricing as seamlessly as possible.
It is thus important to use autoML software that is tailored for the insurance industry, as it will be developed with implementation in mind. Additionally, many have partnerships with large, well-known policy administration providers to enable even easier deployment of models.
Scalability: One actuary can now launch parallel risk pricing models. By means of example, motor third-party liability and motor own damage perils in motor insurance can be run in parallel so that when the prediction cycle is over, results can be aggregated to generate the full burning cost. Alternatively, a single actuary might want to explore many hundreds of interactions across many peril-level models at once. Most of these analyses will be pruned using business acumen or for lack of statistical significance, but exploration can enable interesting discoveries to be made.
Resource efficiency: Taking advantage of automated model development and increased development time, autoML allows businesses to accelerate their technical sophistication journeys. Businesses at the intermediate and advanced stages of their journeys stand to benefit the most, maintaining and enhancing sophistication while using fewer analytical resources. Companies at an earlier stage in their journey can benefit from the other features of autoML, despite not being able to reduce resources.
Striking a balance
The future of the personal lines insurance pricing journey will bring the best of advanced computing, digitalisation and statistical rigour, coupled with increased use of the pricing actuary’s domain knowledge. The goal should be to strike the correct balance between efficiency, predictiveness and scalability. However, interpretation and implementation should remain at the heart of the model-building exercise to ensure maximum end use within the business and return on investment.
Massimo Cavadini is senior executive partner and global leader of insurance solutions at Munich Re
Davide Burlon is principal in insurance solutions at Munich Re
Gili Smadja is principal in insurance solutions at Munich Re
Image Credit | Shutterstock
