TOP
In life insurance, “the algorithm” can provide us with predictions about the probability of death and lapse rates, as well as the probability of disability – but when can we be confident that this prognosis is correct? How can we identify weaknesses? How do we build trust in outcomes we believe in, and how do we ensure that those outcomes are non‑discriminatory?
Modern analysis methods are becoming increasingly important in life insurance. Be it for analysing losses, predicting lapse rates or improving customer communication, many life insurers now employ special data scientists who have grown up IT savvy and with R or Python, and develop solutions using neural networks, lasso-generalised linear models (GLMs) or gradient-boosting methods.
At the same time, data scientists often lack the in‑depth product-specific knowledge to fully interpret data and results and to translate the results into action. The recipients of these analyses are often responsible actuaries and board members who have a deep, traditional understanding of insurance – but who lack access to new models.
This entails the risk of projects not having the right goals, being implemented incorrectly, or not using excellent results because decision-makers cannot be persuaded – as well as the risk of accepting results and drawing wrong conclusions based on exaggerated model faith. How is it possible to come to a common language in order to generate optimum results and find acceptance for them? While there are peculiarities in each individual case, we have gained excellent experiences across all projects with the following approaches.
The various disciplines affected should be represented in every complex data science project – depending on the project, this may be actuaries, underwriters or claims managers. Ideally, future decision-makers should also be involved at an early stage and on several occasions. In addition to a holistic understanding of the task to be solved, this leads to a common language and enables the concerns and perspectives of the various areas to be understood and taken into account at an early stage. The fact that decision-makers are given the feeling of being able to exert influence at an early stage helps to dispel reservations.
A detailed descriptive section provides all parties with an important overview of the basics and enables them to develop their ideas. This means that even colleagues who have problems with the more complex evaluations feel at ease, at least at the beginning, and build up less resistance.
A neural network may sound more impressive than a GLM and may or may not deliver better results depending on the case – but it is definitely more difficult tointerpret. Similarly, a forest of decision trees can deliver better results than a single tree, but the individual tree will be more vivid. In the final selection of the model used – of which different types are often tested in the course of a project – an increase in accuracy of the result must therefore be weighed against a loss of interpretability.
Data scientists check the plausibility of their results in a variety of ways. These tools and metrics, which are not limited to a specific question, are not familiar to all classic actuaries, underwriters or board members. However, they can be explained and give even those who are not involved in the evaluation the certainty that care was taken. An understanding of the quality tests used makes it possible to identify fields that are not suitable for testing.
Even if downstream decision-makers do not understand in detail how a model works, there are almost always opportunities to critically question the results on a case-by-case basis without falling into a vague rejection.
Suppose a model predicts mortality based on characteristics such as age, gender, pre-existing conditions and lifestyle characteristics. Are the values for larger subsets as one would expect? For example, do men have a higher mortality rate than women? Does mortality increase with age? Does pancreatic cancer have a greater impact than non-melanoma skin cancer?
How well are known special effects modelled: Is there an accident spike in the above example, and is this more pronounced in men than in women?
Which data were not considered, and why? Are there any outliers that would be of great importance in this particular context, e.g., in the case of extreme health insurance claims?
If the answers are satisfactory, this is a good confirmation. Unexpected effects can reveal potential model weaknesses or lead to interesting new insights.
There are certainly model questions that are so technical that laypersons are not much smarter after the answer than before: for example, how many layers a machine learning model consists of. But how the choice was made should be explained in a comprehensible way. This reflection can also be helpful for the data scientist.
Questions about the dataset should always be asked – if, for example, forecasts are derived from a very short data period for very long periods, this is at least a warning sign. When predicting the effects of diseases on mortality, for example, it would be useful to examine whether the prognosis for chronic progressive diseases such as diabetes is as good as for sudden events such as a heart attack.
Conversely, data specialists should often ask why certain assumptions are made, expectations exist, or results are considered (im)plausible in order to develop a better understanding of the relevant topic.
In the medium term, data science will become part of the natural toolbox of actuaries. At the same time, the pace of digitisation and change remains high, so there will always be techniques that only a handful of specialists in the company can master. In addition to technical skills, it is therefore all the more important to strengthen the communication culture within companies in order to be able to answer questions in an interdisciplinary and integrated fashion.
Are you looking for support for your project? Then don’t hesitate to get in touch with us!
