At the heart of insurer activity, actuarial data is a valuable tool for risk prediction and price calculation for each of the risks covered. Can this data also help to anticipate precisely the insured life events, and thus feed insurers’ customer knowledge? Can risk prediction be transformed into customer prediction, relying on the latest data analysis technologies?
What are the types of data used by insurers?
Actuaries, the data and statistics specialists for insurers, use the internal customer data, that is to say the data that the insured provide them, especially at the time of the contract, on their vehicle, their place of residence, etc. They can also “draw” elsewhere to estimate the maximum amount of a claim; in the flood history; burglary data; or the weather. From all this data, actuaries will create models to predict accident, fraud and termination scores…
What is the purpose of analyzing actuarial data?
These data allow us to segment the portfolio, to find the right price variables, to create tariff classes in auto or home insurance, for example. They can also be used to find the right price to cover new risks. They finally meet the new demands of the insured. Before, with auto insurance for example, when your car broke down, you were reimbursed for the costs and that was enough. Now, customers want to be housed, have a replacement car… It is up to the actuary to evaluate the cost of these new services to include them in the various premiums offered by the insurer.
What difference do you make between forecast and prediction?
I would say that the forecast concerns time series, such as those with the forecasts of household consumption, which makes it possible to predict what will happen in the future. While the prediction touches on individual behaviours: “I predict for example that such person will have an accident at such a time”.
In a predictive model, you enter information about someone and you will try to predict something about that person … But the predictive models of insurers, in reality, are not intended to apply to individuals: we want to know for example the percentage of people who will die next year – but not who exactly! The advantage, with insurance, is that we have a lot of history, so we can quickly confront his model to reality.
What behaviours can actuarial data anticipate?
We can guess things, auto insurance in particular, withthe age of the vehicle. We can estimate when a person will want to change vehicles, pass a license – depending on the age and the number of children declared in the contract. From a marketing point of view, it can also be interesting to propose offers. But there is on one side the marketing approach, which will seek to individualize to offer personalized offers; on the other, that of the actuary who will, for its part, seek to find relevant categories by pooling.
In insurance, we need to reason about groups, not about individuals. We must go back to the very definition of what insurance is. It is based on the fundamental principle of risk pooling: we cannot insure a risk by ourselves; we have to insure it in a community. Otherwise it’s called a bet! In insurance, we cannot cover ourselves against an individual risk. For an insurer, it is therefore necessary to find the right balance between segmentation and pooling.
Machine learning and deep learning: what maturity for insurers, in the service of anticipating the behaviour of policyholders?
Actuaries have always “turned” algorithms, it’s not new. The problem is especially the culture of the one who builds and uses these models. Today, actuaries share the same culture. Result: regardless of insurers, prices are consistent, with “good risks” for just about everyone.
But machine learning, if misused, can have an influence on the segmentation of policyholders: rankings can change dramatically, leading to premium variations ranging from 1 to 100, with the risk that this could lead to a lack of insurance, particularly in the area of auto insurance. Fortunately, it remains very framed with the bonus-malus system.