Predicting Wine Prices Based on the Weather: Bordeaux Vineyards in a Changing Climate

Abstract

In the twentieth century, high summer temperatures were beneficial to grape and wine quality in Bordeaux. However, owing to global warming gradually raising temperature closer to the optimum needed by the regional grape varieties, the positive effect of higher-than-usual summer temperatures has vanished over the last decades. Therefore, it is unknown whether any weather variable is still impactful enough to quantify future wine quality.Here we provide a predictive model of wine prices, based only on weather data. We establish that it predicts more accurately a vintage’s long-term quality than a world-class expert rating this same vintage in the year following its production.We first design a corpus of features suited to the grapevine lifecycle, so as to distinguish the grapevine’s contrasting needs across the different stages of its growth.  Using Bayesian inference with a Hamiltonian Monte Carlo (HMC) algorithm, we then select the most powerful drivers of wine quality. Finally, we build a predictive model that leverages Local Least Squares kernel regression (LLS) to allow model coefficients to change over the vintages, thus factoring in the time-varying nature of climate impact on the grapevine.Using the phenology-adapted features, and letting coefficients vary through time with LLS, both significantly improve performance. The proposed model thus achieves state-of-the-art predictive accuracy, and it even provides a better predictive ranking of successive vintages than the grades given by world-famous wine critic Robert Parker.This demonstrates that weather is still a very efficient predictor of wine quality in Bordeaux. This study provides strong support for the usage of weather-based models as auxiliaries in the pricing of premium agricultural products. The HMC algorithm, both flexible and robust to noise, can be used in the feature selection step of many modelling problems. Finally, the novel usage of a LLS architecture, allowing the input-output relationship to smoothly vary over time, would have exciting development in the modelling of other agricultural systems, in face of the changes introduced by global warming and adaptation of production methods.

Keywords: Climate change, Grapevine, machine learning, Hamiltonian Monte Carlo, Local Least Squares kernel regression