The Role of Data-Driven Methodologies in Weather Index Insurance

Abstract

There are several index insurance methodologies. Most of them rely on linear piece-wise methods. Recently, there has been studies promoting the potential of data-driven methodologies in construction index insurance models due to their ability to capture intricate non-linear structures. However, these types of frameworks have mainly been implemented in high-income countries due to the large amounts of data and high-frequency requirements. This paper adapts a data-driven methodology based on high-frequency satellite-based climate indices to explain flood risk and agricultural losses in the Antioquia area (Colombia). We used flood records as a proxy of crop losses, while satellite data comprises run-off, soil moisture, and precipitation variables. We analyse the period between 3 June 2000 and 31 December 2021. We used a logistic regression model as a reference point to assess the performance of a deep neural network. The results show that a neural network performs better than traditional logistic regression models for the available loss event data on the selected performance metrics. Additionally, we obtained a utility measure to derive the costs associated for both parts involved including the policyholder and the insurance provider. When using neural networks, costs associated with the policyholder are lower for the majority of the range of cut-off values. This approach contributes to the future construction of weather insurance indexes for the region where a decrease in the base risk would be expected, thus, resulting in a reduction in insurance costs.

Keywords

• index insurance
• crop insurance
• machine learning
• neural networks
• satellite data
• google earth engine