Generalized Linear Model (GLM) is the standard insurance risk modeling technique. Due to its complete interpretability, it is widely adopted in the industry and by regulators in most of the markets. Understanding how a GLM works is crystal clear: you define the variables in the model and you choose the impact of each variable. Say you have built a GLM model using only the Age and the city as variables.
To estimate the risk of a 30 years old living in Paris, all you need is to look up for three numbers: the mean cost (say 500 €), the effect of Age at the value 30 (say -25%) , and the effect of the City in the value Paris (for example + 50%).
Then, the cost is given by the multiplication of these values: the final risk will be 600 €.
Without Akur8, GLMs can be automated but at the expense of their performances
On a competitive market, it is essential to rapidly capture emerging trends in risk. When the modeling process drains time and resources in your team, the need for automation is vital. At this moment, GLM automation is possible, but the variable effect will be purely linear or simple spline. Since the risk signal is often complex, this extreme approximation cannot capture the full risk signal, leading to weak market segmentation.
When we aim at a correct and powerful market description, the only choice is to manually build a GLM. Forcing your team to input each variables and its impact leads heavily iterative and error prone process. Deadlines are extended and models are slowly updated, reducing the possibility to capture market opportunities.
Usual machine learning techniques are black box, so useless.
The recent machine learning revolution seems to have both automated and increase the performance of risk predictions. However, all the trending techniques such as GBM (Gradient Boosting Machines) are black box models: interpretability is completely lost.
This loss of clarity has many hidden costs. First of all, you lose control of your risk model. The impossibility to incorporate actuarial knowledge in your model rises the anti-selection risks: a GBM can ignore many extreme sources of risk if they are under-represented in your (today's) portfolio.
The auditability of a model is also compromised: modeling errors are harder to spot, communication within your team more difficult.
Finally, reliably implement complex black-box models in your production pipeline can raise significative costs on your IT system.
Akur8 resolves the Automation + Performance + Interpretability Dilemma
Our team, coming from years of operative experience in the insurance market, believed that we can have Automation, Performance and Interpretability at the same time.
After years of R&D with top academical researchers, we are now ready to empower teams of actuaries with a full featured SAAS solution. Building a performant and understandable GLM is now as easy as choosing the range of number of variables in your model: our Artificial Intelligence will optimally choose the most important variables and automatically tune the effects of the variables.
Have you ever wondered if a more complex model will enhance your performances? Or if your costly external database is not capturing some signal you already have at your disposal?
Akur8 solution will show you these answers with just a click: you can now explore the complexity of your risk portfolio at an overview you will have never experienced.