A credibility-based yield forecasting model for crop reinsurance pricing and weather risk management

Abstract

Purpose The purpose of this paper is to propose an improved reinsurance pricing framework, which includes a crop yield forecasting model that integrates weather variables and crop production information from different geographically correlated regions using a new credibility estimator, and closed form reinsurance pricing formulas. A yield restatement approach to account for changing crop mix through time is also demonstrated. Design/methodology/approach The new crop yield forecasting model is empirically analyzed based on detailed farm-level data from Manitoba, Canada, covering 216 crop varieties from 19,238 farms from 1996 to 2011. As well, corresponding weather data from 30 stations, including daily temperature and precipitation, are considered. Algorithms that combine screening regression, cross-validation and principal component analysis are evaluated for the purpose of achieving efficient dimension reduction and model selection. Findings The results show that the new yield forecasting model provides significant improvements over the classical regression model, both in terms of in-sample and out-of-sample forecasting abilities. Research limitations/implications The empirical analysis is limited to data from the province of Manitoba, Canada, and other regions may show different results. Practical implications This research is useful from a risk management perspective for insurers and reinsurers, and the framework may also be used to develop improved weather risk management strategies to help manage adverse weather events. Originality/value This is the first paper to integrate a credibility estimator for crop yield forecasting, and develop a closed form reinsurance pricing formula.

Deep Learning for Improved Agricultural Risk Management

“…Their findings show that assumed risk evolution (i.e., increasing/decreasing) in rate-making methodologies can be violated leading to highly biased rates. Understanding the dynamic interaction between crop yields and various weather risks is, therefore, critically important in basis risk estimation and prediction [11]. Furthermore, Ghari et.…”

“…The following benchmark models were selected as competitors for deep learning: GLM and its two modifications based on the model selection algorithms of screening regression (SR) and principal component analysis screening regression (PCASR) [11,12], gradient boosting (GB) [26] and random forest (RF) learning [27].…”

“…The province of Manitoba was selected as the study region based on availability of crop yield and climate information and for comparison with the results of [11]. Historical, farm-scale panel crop yield data (yield per acre and total acres, 1996-2011, where the units of yield are metric tonnes, mt) covering 99 regional municipalities, 75 crop types from 19,238 farms were used, publicly available online from the Manitoba Agricultural Services Corporation (MASC) MMPP Variety Yield Data Browser 2 .…”

“…Such normalization has two linked benefits: 1) comparatively assesses (risks of) low yields for different crops, and, 2) enables aggregation across different crops. A re-statement procedure was next applied to the re-scaled yields using the same procedure detailed in [11]. This aggregates yield estimates for each single crop as a weighted-average representative crop mix, so as to provide a better reflection of the current risk profile as crop selection differs from year to year.…”

“…Following [11], two matrices of weather indicesseparately based on the observed station and reanalysis datasets -were computed for all 16 years (1996-2011) and municipalities, i.e., annual indices, for each month in the growing season (May-October). Each of the matrices included up to 432 variables: night-and daygrowing degree low (NGDL, DGDL), night-and daygrowing degree high (NGDH, DGDH), precipitation low and high (PREH, PREL), etc.…”