<p>The use of statistical models to study the impact of weather on crop yield has not ceased to increase. Unfortunately, this type of application is characterised by datasets with a very limited number of samples (typically one sample per year). In general, statistical inference uses three datasets: the training dataset to optimise the model parameters, the validation datasets to select the best model, and the testing dataset to evaluate the model generalisation ability. Splitting the overall database into three datasets is impossible in crop yield modelling. The leave-one-out cross-validation method or simply leave-one-out (LOO) has been introduced to facilitate statistical modelling when the database is limited. However, the model choice is made using the testing dataset, which can be misleading by favouring unnecessarily complex models. The nested cross-validation approach was introduced in machine learning to avoid this problem by truly utilising three datasets, especially problems with limited databases. In this study, we proposed one particular implementation of the nested cross-validation, called the leave-two-out method (LTO), to chose the best model with an optimal model complexity (using the validation dataset) and estimated the true model quality (using the testing dataset). Two applications are considered: Robusta coffee in Cu M'gar (Dak Lak, Vietnam) and grain maize over 96 French departments. In both cases, LOO is misleading by choosing too complex models; LTO indicates that simpler models actually perform better when a reliable generalisation test is considered. The simple models obtained using the LTO approach have reasonable yield anomaly forecasting skills in both study crops. This LTO approach can also be used in seasonal forecasting applications. We suggest that the LTO method should become a standard procedure for statistical crop modelling.</p>