Abstract: While intercomparison of process-based crop models for projections under climate change is being intensively studied at European as well as at the global scale, little effort has been made for comparing statistical models. In this study, several regression techniques (ordinary least squares, stepwise, shrinkage methods, principle components and partial least squares) were combined with different types of climate input variables (with different temporal resolution) in order to define a large range of statistical models. Each model was fitted to winter oilseed rape data collected in 689, 325 and 173 field experiments carried out in Denmark, Germany, and Czech Republic, respectively. The fitted models were then used to predict yield of winter oilseed rape in the field experiments during more than 20 years, up to 2013.  Interpretability of the estimated climate variable effects and accuracy of yield predictions were both analysed. Results suggest that recent statistical methods (e.g., shrinkage methods) may have considerable capabilities to complement traditional statistical methods in yield prediction. The selection of the most influential variables was strongly influenced by the statistical method used to analyse the data. Among the most recent statistical methods, the uncertainties in projecting yield of winter oilseed rape under climate change were mainly due to residual errors and uncertainty in estimated parameter values, and not to model choice. No Label

Abstract: When evaluating the performances of simulation models, the perception of the quality of the outputs may depend on the statistics used to compare simulated and observed data. In order to have a comprehensive understanding of model performance, the use of a variety of metrics is generally advocated. However, since they may be correlated, the use of two or more metrics may convey the same information, leading to redundancy. This study intends to investigate the interrelationship between evaluation metrics, with the aim of identifying the most useful set of indicators, for assessing simulation performance. Our focus is on agro-ecological modelling. Twenty-three performance indicators were selected to compare simulated and observed data of four agronomic and meteorological variables: above-ground biomass, leaf area index, hourly air relative humidity and daily solar radiation. Indicators were calculated on large data sets, collected to effectively apply correlation analysis techniques. For each variable, the interrelationship between each pair of indicators was evaluated, by computing the Spearman’s rank correlation coefficient. A definition of “stable correlation” was proposed, based on the test of heterogeneity, allowing to assess whether two or more correlation coefficients are equal. An optimal subset of indicators was identified, striking a balance between number of indicators, amount of provided information and information redundancy. They are: Index of Agreement, Squared Bias, Root Mean Squared Relative Error, Pattern Index, Persistence Model Efficiency and Spearman’s Correlation Coefficient. The present study was carried out in the context of CropM-LiveM cross-cutting activities of MACSUR knowledge hub. No Label