Improved crop forecasts for the Australian macadamia industry from ensemble models

Abstract

Annual forecasts for the Australian macadamia crop have been issued since 2001, with varying (and not always improving) degrees of accuracy. Regression models using climate variables have formed the basis for these forecasts, with general linear model (GLM) ensembles being adopted more recently. This has proven to be a challenging task, as there are only a small number of observations (18) combined with a large number (90+) of independent variables – these being different climate measures for different times of the year (representing ‘key physiological periods for macadamia trees’). Also, these ‘assumedly-independent’ variables contain various degrees of correlation. This study uses cross-validation, with the most recent data for the two dominant production regions of Australia (Lismore and Bundaberg), to investigate the relative performance of alternate modelling methods. These modelling methods were GLMs, partial least squares (PLS) regression and LASSO (least absolute shrinkage and selection operator) penalised regression. Model ensembles, which have been shown to be beneficial in many alternate disciplines, are used to advantage. Both GLMs and PLS produced quite-disappointing results, failing to meet the project's benchmarked accuracy of ±10% error. The optimal LASSO models performed notably better, with a further improvement when ensembles were incorporated. The lowest mean absolute error (MAE) rates here were 9.0% for Lismore and 5.9% for Bundaberg. Hence LASSO ensembles will be adopted for future forecasts of the Australian macadamia crop.