Anticipatory modelling of insurance losses, an application of the PICS research project
(1) Jean-Philippe Naulin - Modelling, R&D Department, CCR
(2) Olivier Payrastre - Water Environment Laboratory, Dep. of Geotechnics Environment Natural Hazards and Earth Sciences, Gustave Eiffel University
INTRODUCTION
In the context of managing a flash flood event, the prediction and quality of available information play a key role in decision making. This is the issue addressed by the ANR PICS project on Integrated Immediate Prediction of Flash Flood Impacts. This project, funded by the French National Research Agency (ANR - Agence Nationale pour la Recherche), was able to deal with a number of aspects ranging from rainfall forecasting based on meteorological models, to the modelling of impacts on the ground, including rainfall-runoff hydrological modelling and hydraulic simulations to represent flooded areas. This project also involved a user group comprising several flood risk management stakeholders, such as the flood forecasting services, the Departmental Fire and Rescue Services (SDIS), insurers, the SNCF, etc. to assess the various tools developed.
Among the tools implemented in this project, CCR worked on the development of a damage model, adapted to flash floods. Unlike the proprietary models developed by CCR, this model was able to benefit from hazard data from the organisations involved in the project, including water depth estimates at 5m resolution [1], and radar rainfall estimates. The impact model was also used with precipitation forecast data from meteorological models to assess whether it was possible to foresee the amount of damage to come and to locate the areas most affected by an event
METHODOLOGY
Calibration of the damage mode
The damage model operates at insured property level. For each property, the damage is estimated as the product of the probability of the property being damaged, its estimated destruction rate and its insured value. The latter is data transmitted by insurers or estimated by CCR from available information such as the surface area of the building, its location or its use. The probability of loss and the destruction rate are derived from models, which are calibrated on the basis of the claims also given by insurers each year. These models relate eventdependent explanatory variables (e.g. cumulative rainfall, the ratio between cumulative rainfall and the ten-year rainfall, or the height of water) to event-independent variables such as the height of the building or the difference in altitude with the nearest watercourse.
Several models were tested such as multiple linear regressions, the Lasso method or random forests. Calibration took place on six events and validation on two others. In the end, the model selected is based on a calibration by multiple linear regression. The expertise focuses on differentiating between areas affected by river overflow and those affected by runoff, as well as on controlling trends so as not to keep counterintuitive models. The results obtained during this calibration, presented in Figure 1, are satisfactory. It appears that the damage model is sensitive to the simulated hazard: the 5m resolution of the overflow model provides better performance than the 25m resolution, as does the use of radar rainfall estimates instead of kriged rain. However, the model tends to overestimate losses on low intensity events.
Application from rainfall forecasts from the AROME PI model
Rainfall forecasts from the AROME PI model [2] were made available by MétéoFrance as part of the PICS project. This model allows the forecasting of future rainfall up to 6 hours ahead and at a time step of 15 minutes. Forecasts are updated every hour. Figure 2 shows the predicted cumulative rainfall at different time steps for the Languedoc event in 2018. At 0:00, the model predicted a moderate event, then from 6:00 onwards, the forecasts show an increasingly intense event. This data was implemented in the CCR modelling chain to model the river flow, then to simulate the overflow at a resolution of 25 m, and finally to simulate the damage.
RESULTS
The first result of this assessment is presented in Figure 3. This result shows that 10 hours before the peak of the flood in Carcassonne, the model could predict a significant event of more than €170M. In the following hours, the cost of the event was refined to come up to a figure of €291M on 16 October 2018. The actual amount of the event was subsequently estimated at €220M. This discrepancy between the forecast and the estimate at the end of the event is largely due to the rainfall forecasts which are less accurate than radar observations. However, the order of magnitude of the damage is respected and this result shows that rainfall forecasts can be useful in predicting future damage. Spatialised outputs of the damage model were also produced by aggregating the costs at a 250m grid. The damage maps thus obtained make it possible to predict the evolution of damage over time. From midnight onwards, the communes that will be hit hardest by the event are detected by the damage model.
The losses are confirmed on the following time steps. There is one exception, however, the town of Carcassonne, where the damage is overestimated. This discrepancy is not related to the rainfall forecast but rather to the model itself, which overestimates the areas affected by the flooding of the Aude River. This confirms that the model is very dependent on the quality of the simulated hazard. An advantage of the damage model is that it does not focus on the main water courses but will also consider areas subject to runoff. These maps could therefore be useful information for crisis managers, as they allow them to locate in advance the places most affected by an event by considering all the flooding processes and the location of the issues at stake.
# PICS
# damage model
# flooding
# crisis management
# loss forecasting
The PICS project has also worked on the representation of uncertainties in rainfall forecasts. For this purpose, new forecast products have been developed by MétéoFrance, i.e. forecasts that include a large number of scenarios. The latter, by feeding the damage model, made it possible to generate several damage scenarios with different levels of probability: a median scenario corresponding to the 50th percentile (50% chance of occurrence) and a more severe scenario corresponding to the 90th percentile (10% chance of occurrence). The damage maps associated with these forecast scenarios can also be updated on an hourly basis and provide information that considers the level of uncertainty in the forecasts.
Figure 1 - Comparison between observed and simulated losses on the risks individuals occupying a house (the events of 14/06/2010 and 09/10/2014 were used for model validation)
Figure 2 - Cumulative rainfall maps for the event from the Arome PI model forecasts at 6pm on 14/10/2018 (top left) then at 0:00, 06:00 and 12:00 on 15/10/2018.
Figure 3 - Flood hydrograph in Carcassonne layered with the losses simulated from the AROME PI rainfall forecasts
Figure 4 - Simulated losses in the Carcassonne area on a 250m grid for a policy sample on 15/10/2018 at midnight
THE PARTNERS
The PICS project is made possible by a grant from the ANR (No.ANR-17- CE03- 0011). We would like to thank the members of the PICS project and the user group who agreed to get involved in this project.
CONCLUSION
This study carried out by CCR as part of the PICS research project has made it possible to build a damage mitigation chain adapted to flash floods. This work has confirmed the importance of considering the overflows of very small rivers, runoff and rising networks together to model damage properly.
The damage model has also proven to be able to forecast future damage when fed with data from a weather forecasting model. Although these estimates are subject to significant uncertainties, these uncertainties can be represented, thus the estimates could forecast the impact of current or future events by giving an indication of the amounts involved and the areas affected.
This article presents only a small part of the overall results of the PICS project, which has led to significant advances in the fields of flash flood forecasting and impacts. The project has a website: https://pics.ifsttar.fr where it is possible to look at the main works and results obtained./
REFERENCES
1. Hocini N., Payrastre O., Bourgin F., Gaume E., Davy P., Lague D., Poinsignon L., and Pons F., 2021. Performance of automated methods for flash flood inundation mapping: a comparison of a digital terrain model (DTM) filling and two hydrodynamic methods, Hydrol. Earth Syst. Sci., 25, 2979–2995, https://doi.org/10.5194/hess-25- 2979-2021https://doi.org/10.5194/ hess-25-2979-2021
2. Lovat, A., Vincendon, B., and Ducrocq, V., 2020. Hydrometeorological evaluation of two nowcasting systems for Mediterranean heavy precipitation events with operational considerations, Hydrol. Earth Syst. Sci. Discuss. [preprint], https:// doi.org/10.5194/hess-2020-629, accepted for publication.
CITATION
Naulin et al, Anticipatory modelling of insurance losses, an application of the PICS project. In CCR 2022 Scientific Report; CCR, Paris, France, 2022, pp. 40-43