Predicting Wine Prices from Wine Labels

Motivation: The wine label is an icon that unites powerful branding and fine design under a single image. With U.S. wine consumption predicted to increase to $23.8 billion by 2014, wine labels are used not only to convey a given wine’s type, producer, and origin, but also its emotional and cultural appeal. Thus, a wine label is carefully designed to convey the proper messaging of a particular bottle, whether it is for the casual consumer or the choosy collector. We have applied machine learning to the art of wine label design by learning the design features which most influence a bottle’s popularity and selling price. We have collected over 30,000 wine labels and associated data on price, type, origin, and other notes from the Web. From this data, we have developed an extensive training set by extracting particular graphical features from the labels and textual features from the metadata. By implementing a support vector machine (SVM) and a Naïve Bayes classifier, we have worked to develop a preliminary model for classifying wines, based on their labels, into specific price bins. However, our results demonstrate that the features we’ve chosen have little power in predicting the price (bin) of a wine bottle, especially at granularities that might be useful.