Modelling consumer choice in the wine market

For many people, choosing a wine can be extremely intimidating, so in a project for Laithwaites, the direct to consumer wine merchant, we developed a model to generate recommendations and help guide the decision.

Laithwaites has a range of several thousand wines, covering all major countries, styles and price points. To help existing customers to navigate the range, we had previously developed a tool which identified wines similar to one that a customer already liked. However, new customers do not have the ‘starting point’ for such a journey, so the team were keen to help them to find a wine or two that they had a good chance of liking. The business had already developed a framework to describe their wines based on a number of attributes, which was the foundation of the ‘similar wines’ model. They then developed a questionnaire about other food preferences based on the same framework, which they wished to use to help drive the recommendation. For example, there were questions about how the respondent liked their coffee/tea (with or without sugar, black or white) and what sort of chocolate they preferred.

We suggested that respondents should also be asked to taste and rate a number of the top selling wines in the range. By taking the most popular wines, we already know that many thousands of customers like them, which gives a high probability that they will be to the taste of new customers. We then analysed the data to see whether the responses provided significant predictive power of someone’s liking for each of the wines. Having found that there was reasonable predictive power in the data, it was possible to build a series of models and then rank all of the wines in order of predicted preference so as to recommend the top few for tasting.

Finally, we built a simple iPad app which could be deployed at tastings, shows and the company’s shops to demonstrate the model in practice. The app presented the questionnaire and immediately generated the recommendations.

The results, whilst not perfect, are a significant improvement on what would be expected if the recommendations had been generated at random: 75% of people who tasted the top recommended wine said they ‘liked’ or ‘loved’ it – significantly up on what would be expected from a random choice. However, the value of the model is not purely in its ability to run the analysis and produce a recommendation. The whole process is interesting, informative and highly engaging for the customer. It provides an entertaining entrée to the Laithwaites range and even if the subject doesn’t particularly like the first wine suggested, it is highly likely that they will like one of the top two or three recommendations. The overall experience will help to make them a much more engaged and loyal customer than if they had simply wandered into a shop, selected a bottle and walked straight out again.