Abstract:

Weather index insurance based on machine learning algorithms represents a significant innovation in agricultural insurance research. Since crop yields are primarily influenced by weather-related disasters, developing a robust data analysis model that accurately captures the relationship between yield losses and adverse weather conditions is crucial for pricing crop weather index insurance. This paper focuses on Qingyang apples in Gansu province, utilizing daily precipitation and temperature data during the growing season (April-October) and apple yield data from five counties (or districts) in Qingyang city spanning 1996–2020. Indices of low−temperature freezing, drought and continuous cloudy rainfall were constructed, and a regression model linking these indices to the meteorological yield reduction rate of apples was established using the XGBoost algorithm. The kernel density estimation method was applied to determine the pure rate of weather index insurance for apples in Qingyang. The findings of the study were as follows: (1) meteorological disasters caused significant fluctuations in the apple cimate yield reduction rates across counties (or districts) in Qingyang city. A nonlinear relationship was observed between the cimate yield reduction rate and seven types of apple disaster weather indices. (2) Regression models for the climate yield reduction rate-weather indices in Ning county, Qingcheng county, Zhengning county, Huan county, and Xifeng district (1996–2020) were constructed using the XGBoost algorithm. These models demonstrated superior fitting accuracy compared to multivariate stepwise regression models, with coefficients of determination (R²) improving by 0.157, 0.125, 0.190, 0.115 and 0.117, uhile root mean square errors (RMSE) decreasing by 0.045, 0.026, 0.335, 0.126, and 0.039 percentage points, respectively. (3) The climate yield reduction rate payout triggers for apple weather index insurance were 11.88%, 3.37%, 4.33%, 9.21%, and 17.70% in Ning county, Qingcheng county, Zhengning county, Huan county, and Xifeng district, respectively. The corresponding pure insurance rates were 4.00%, 3.64%, 4.91%, 1.94% and 4.98%.  

Key words: Weather index insurance, XGBoost algorithm, Kernel function, Pure rate determination