Accurate preharvest yield estimation is an important issue for agricultural planning purposes and precision farming. Machine learning (ML) based on readily obtained information on the cropping system, typically including spectral reflectance measurements, is an essential approach for achieving practical solutions. We tested in a 9-year soil compaction experiment the accuracy of ML-based yield predictions made up to 2 months before harvest from a Ratio Vegetation Index (RVI) and recordings of precipitation and reference evapotranspiration. The applied data set comprises 224 combinations of plots and years with measured grain yields in the range of 4.22–9.34 Mg/ha. The best ML model [i.e., with the smallest mean absolute error (MAE)] was selected automatically by the AutoML interface included in the R program package H₂O. Its cross-validated predictions made on June 30 more than 1 month before harvest showed an MAE of 0.38 Mg/ha when trained on all data from all years except the one under consideration. MAE increased to about 0.68 Mg/ha when determined 3 weeks earlier on June 10. MAE values in the range of 0.32–0.42 Mg/ha were obtained for predictions made on June 30 when based on data from at least six consecutive years; however, MAE showed no generally decreasing trend with the number of years. Yield estimations were robust towards a considerable soil variation observed within the experimental area due in part to the experimental treatments. The results show a potential of making yield predictions in barley 1–2 months before harvest, which, however, is not sufficiently early to support decisions on top-dress N fertilization.