Statistical, data-driven methods are considered good alternatives to process-based models for the sub-national monitoring of cereal crop yields, since they can flexibly handle large datasets and can be calibrated simultaneously to different areas. Here, we assess the influence of several characteristics on the ability of these methods to forecast cereal yields at the local scale. We look at two diverse agro-climatic Italian regions and analyze the most relevant types of cereal crops produced (wheat, barley, maize and rice). Models of different complexity levels are built for all species by considering six meteorological and remote sensing indicators as candidate predictive variables. Yield data at three different spatial aggregation scales were retrieved from a comprehensive, farm-level dataset over the period 2001-2015. Overall, our results suggest the better predictability of summer crops compared to winter crops, irrespective of the model considered, reflecting a more intricate relationship among winter cereals, their physiology and weather patterns. At higher spatial resolutions, more sophisticated modelling techniques resting on feature selection from multiple indicators outperformed more parsimonious linear models. These gains, however, vanished as data were further aggregated spatially, with the predictive ability of all competing models converging at the agricultural district and province levels. Feature-selection models tended to elicit more satellite-based than meteorological indicators, with a preference for temperature indicators in summer crops, whereas variables describing the water content of the soil/plant were more often selected in winter crops. The selected features were, in general, equally distributed along the plant growing cycle.