Identifying indicators for extreme wheat and maize yield losses

•A method to investigate the accuracy of extreme yield loss indicators is proposed.•Best indicators identify extreme yield losses early in the growing season.•Some indicators outperform the other in specific crop–country combinations.•We develop graphs linking indicator values and the probability of extreme yield loss.•This study presents a roadmap for systematic evaluation of early-warning systems. Yield forecasts are generally based on a combination of expert knowledge, survey data, statistical analyses and model simulations. These forecasts, when public, influence crop prices and can be used to estimate end-of-season stocks. Thus, the skills and limitations of such products are important because they inform trade policies. In Europe, yield forecasts are made available to stakeholders throughout the growing season via the monthly MARS (Monitoring Agricultural ResourceS) Bulletin. The MARS Crop Yield Forecasting System relies on an in-depth analysis of past climate, short-term weather forecasts and crop growth simulations. In this paper, we focus on the occurrence of abnormally low yields and evaluate how accurately agro-climatic indicators and model outputs anticipate their occurrences for two crop species in two European countries of contrasted agroclimatic conditions. Importantly, the indicators considered here encompass a large range of complexity levels and several are used to inform the yield forecasts presented in the MARS Bulletin. Each indicator is independently used to predict the onset occurrence of an abnormal yield loss (henceforth named extreme) in France and in Spain for both winter wheat and non-irrigated grain maize for a period covering the 1976–2013 growing seasons. Indicators are ranked based on a score quantifying their ability to accurately separate extreme from non-extreme yield loss events. We provide and in-depth analysis of the robustness of our ranking to alternative definitions of extreme yield loss. No single indicator performs systematically well (e.g., whatever the country or crop species) but several show acceptable scores (e.g., averaged temperatures, vapor pressure deficit, precipitation, potential yield). We find no obvious relationship between the level of complexity of indicators and their accuracy. Single climate variables such as temperature or precipitation often perform as well as a crop model integrating temperatures and precipitation effects on crop growth. Namely, monthly averaged maximum temperatures and precip