Beyond point forecasting: Probability density forecasting of corn yield based on quantile regression forest

Abstract Crop yield forecasting is crucial for global food security. In this paper, we go beyond traditional point forecasting to examine the probability density forecasting of corn yield using a quantile-based machine learning approach. Leveraging 36 years of county-level panel data that cover 1260 counties in China between 1984 and 2019, we develop a quantile regression forest model, enhancing the traditional random forest by integrating quantile regression, for probability density forecasting of corn yield. Quantile regression and quantile regression neural networks are selected as benchmarks. Thirteen meteorological variables are adopted as predictors, and LASSO is used to examine the importance of each variable. Our results show that all quantile-based models produce good point forecasts, prediction intervals, and probability density curves; in general, we find that quantile regression forest with LASSO is best. We also find that the quantile regression neural network does not perform better than the traditional quantile regression, and LASSO doesn’t improve prediction models much in this context. This paper demonstrates that quantile-based machine learning is a promising tool for probability density forecasting of corn yield — but attention must be paid to the selection of machine learning algorithms and their black box properties.