Yield prediction is an essential component of the production chain of wineries. Accurately knowing, in advance, the amount of grapes being produced is crucial to establishing a proper logistic. Yield prediction models based on field and ancillary variables have been developed; predictions can be made by variety at the global or local (field) level. Segmenting the data sets into different groups and then running the corresponding regressions within each group may improve the quality of the predictions. The use of ancillary variables such as aerial or satellite imagery may facilitate data clustering. The present work had for objective to explore different mathematical models for early yield estimation of wine grape. Three-year data were used. Data consisted on the weight and number of bunches per meter row, taken at different times before harvest:> 90 days before harvest (DBH), 60-90 DBH, 30-60 DBH, and < 30 DBH. At each field, samples (< 20 per field) were collected in a systematic design, with three replications at each sampling point. Ancillary data consisted on a vegetation index (either PCD or NDVI) taken at veraison. Several mathematical models, using cluster regression as a base, were evaluated including: general (one variety at several farms), farm (one variety at each farm), and field (one variety at each field). Clusters were made using a hierarchical clustering algorithm. Results demonstrated that in general, local models performed better than the general ones and that the predictions were acceptable.   It is possible to predict yield as early as > 90 DBH.