Combining engineering and data-driven approaches: Calibration of a generic fire risk model with data

• Engineering fire risk models can be improved by calibration with statistical data.
• The calibration must deal with incomplete data sets collected at portfolio level.
• The procedure is applied to a generic fire risk model for single family houses.
• The effect of the calibration is discussed using the single family house model.
• The model is validated by using only one part of the data for calibration.

Two general approaches may be followed for the development of a fire risk model: statistical models based on observed fire losses can support simple cost-benefit studies but are usually not detailed enough for engineering decision-making. Engineering models, on the other hand, require many assumptions that may result in a biased risk assessment. In two related papers we show how engineering and data-driven modelling can be combined by developing generic risk models that are calibrated to statistical data on observed fire events. The focus of the present paper is on the calibration procedure. A framework is developed that is able to deal with data collection in non-homogeneous portfolios of buildings. Also incomplete data sets containing only little information on each fire event can be used for model calibration. To illustrate the capabilities of the proposed framework, it is applied to the calibration of a generic fire risk model for single family houses to Swiss insurance data. The example demonstrates that the bias in the risk estimation can be strongly reduced by model calibration.