Development of a multiperil damage platform
(1) Léa Boittin - Modelling, R&D Department, CCR
(2) Gilles Kieffer - FoxPlan
(3) Frédéric Drapeau - Data Science and Actuarial Department, CCR
(4) Jean-Philippe Naulin - Modelling, R&D Department, CCR
(5) Damien Dronsart - Information Systems Department, CCR
(6) David Moncoulon - Modelling, R&D Department, CCR
INTRODUCTION
To estimate its exposure to natural catastrophes, CCR needs to compute the costs to its portfolio of several types of disasters. CCR has chosen to build up its own damage model over time, thus enabling it to control all the chains of loss estimation following an event and thus has a tool that is perfectly adapted to the perils covered by the Nat Cat scheme in France.
As in most commercial Cat models, the calculation of insurance losses is based on three modules:
- a hazard module, in which the occurrence, location and intensity of the phenomenon are modelled using various approaches (hydrological, seismological, etc.);
- a vulnerability module, which characterises the properties (or insured assets) potentially exposed to the hazard. This involves knowing the location of the buildings, their type (residential, professional, agricultural), their nature (house, building, flat, etc.), their use (owneroccupied, non-owner-occupied, tenant, etc.) and their insured value;
- a damage calculation module, in which the hazard is cross-referenced with vulnerability data to calculate the amount of losses. Whether one is interested in a single real event, or an event from a probabilistic catalogue, the calculation method is the same. Classically, the method is based on damage curves, chosen according to the characteristics of the exposed buildings, which link the intensity of the physical phenomenon to a damage rate. CCR models also have a specific feature, that of using probability curves for the occurrence of a claim and for the recognitionof the state of natural disaster.
In 2022, CCR decided to rewrite its damage calculation solution to have a more efficient platform, which anticipates the integration of future perils and the consideration of joint perils. This new code is being written in the Matlab language. The multiperil platform focuses solely on the third module, which is damage calculation. At present, it covers the following hazards: flood, drought, earthquake and coastal flooding. Eventually, it will include new features such as automobile damage, business interruption and agricultural crop loss.
METHODOLOGY
A robust calculation kernel
To provide the most robust platform possible, it is preferable to separate as much as possible the phases involving the reading of data (e.g. portfolio and hazard data), the calculation of damage itself, and the writing of the results. The post-processing of results is also separate. The idea is therefore to have a cost function for each hazard that assigns a damage rate and a probability of loss to each location hazard. These are defined by their coordinates, type, nature, use and insured values. The work of constituting a portfolio in which each location is entered is done upstream and is not integrated in the calculation kernel. For example, it is upstream that the coordinates of the locations are defined for which the exact address is not known, but only the commune. In the same way, it is upstream that one or more values representing the hazard are associated with each location. Thus, the calculation kernel is only concerned with the calculation of the destruction rate and the probability of loss, regardless of how the portfolio has been constructed.
Figure 1 - Operating principle of the multiperil damage platform
Figure 2 - Opportunities to scale the platform up
Industrialisation for portability and scaling
The portability of the model is part of a wider project to industrialise the modelling activities of CCR. Portability is essential to run a model internally on a powerful machine, on an on-site or cloud-based cluster. The latter option offers infinite computing capacities with platform providers that have developed a high-quality HPC (High-performance computing) offering as standard. However, having a portable model requires several preconditions:
1) having code that runs on Linux platforms because Linux is the de facto standard for scientific computing.
2) making Make the models agnostic of their technical environment through the use of parameterisation, which makes it possible, for example, to stay away from imposed data locations. There should also be no constraint on a particular storage technology.
3) packaging Package the model’s input and output data so that it is not strongly tied to the company’s information system during the computation process.
The need for computing capacity in modelling can increase very sharply from time to time in response to an urgent need, such as the occurrence of an event. It is becoming increasingly difficult for organisations to maintain a large amount of computing capacity on a permanent basis without significant strain and cost. It is vital for organisations to develop business models that will exploit the capacity available in the cloud as standard or as a one-off need. An additional option to reduce costs consists in exploiting a supplier’s residual capacity i.e. to run computations based on the availability of machines. This is very advantageous financially but requires the ability to stop and restart a calculation automatically. Having a portable model will not only allow for significant additional computing capacity but will also reduce costs to an unprecedented degree.
# damage calculation
# industrialisation
# flexibility
# robustness
# portability
Data reading
Particular attention has been paid to the management of paths by the model. In the new version, the format of the data is pre-set, but not its location. The file defining the hazard must be in a certain format, but the hazard can be read anywhere. In the same vein, while the organisation of events within a probabilistic catalogue is constrained, the location of the catalogue is not. However, for data such as damage curves, there are default files that are read, but a user can very easily tell the code to read another file to perform a test. This potentially allows for easy testing of any hazard as well as modification of certain model settings. Finally, it is possible to read portfolio data from a database or from files.
Processing of results and traceability
The results are saved at commune level and by type, nature and use. This then allows several post-processing operations to be carried out according to the user’s choices: aggregation at the scale of the event, at department level, display by type, nature, use on each scale, calculation of the overall distribution of costs. These processing operations can be carried out immediately after the damage calculation, or at a later stage. This separation of the post-processing of the results from the rest of the code and the safeguarding of the results on a fine scale meet the requirement for result traceability. For better traceability, a system of parameter files makes it possible to keep a description of the parameters used to carry out a simulation. For very specific studies, and particularly in the case of ceding companies insuring large industrial locations, it is still possible to save the results at policy level.
Anticipation of future perils
Currently, for perils covered by the Nat Cat scheme, the core element is the insured policy and there is a linear damage model using damage rate and loss frequency. For different perils, outside the Nat Cat scheme, it is conceivable that damages will have to be calculated for different entities via methods without damage rates. This is the case, for example, with agricultural plots, where the surface nature of the plot must be considered when calculating crop losses. The design of the new damage platform foresees the future addition of potentially very different perils from those currently being dealt with.
THE PARTNERS
FoxPlan is a consulting company that supports the management of company projects. FoxPlan supports CCR in its project to industrialise its modelling activities
CONCLUSION
The current project for a multiperil damage calculation platform meets several requirements formulated by CCR: to be able to calculate the insurance damage of several types of catastrophes, in a robust and traceable way, anticipating the future addition of other perils. Today, damage estimation is based on scripts that are mainly run internally and modelling services (for an insurer, for example) have to be performed by CCR staff. The aim is to develop a user-friendly interface in 2023 that focuses exclusively on the multiperil damage platform and, from 2024 onwards, to provide direct access to this platform via API technology or containers. Thus, in the long run, it would allow the roll-out of modelling services to external users./
CITATION
Boittin et al, Development of a multiperil damage platform. In CCR 2022 Scientific Report; CCR, Paris, France, 2022, pp. 14-17