- We propose a machine learning tool to identify customers that are likely to opt for a renewable energy product

- We were able to double the response rate of a mailing campaign

In saturated markets, it is particularly difficult to acquire new customers. Consequently, a promising strategy is to increase the value of existing customers. We supported an up-selling campaign of an energy retailer in Switzerland in their attempt to turn customers of a conventional product to customers of a premium product. By using machine learning, we were able identify households with a high likelihood of switching from a standard gas tariff to a premium biogas tariff. A challenge in this project was that neither ground truth information about the purchase interest towards a biogas tariff, nor socioeconomic or demographic data was available for classifier training.

We used a semi-supervised learning approach to obtain up-selling scores for each standard gas customer as a proxy for the likelihood that this customer switches to the biogas tariff. The scores were obtained using Random Forest classification algorithms. For building the machine learning model, we first analyzed the available business data and developed empirical features in cooperation with data scientists and energy retail experts.

In a pilot campaign, the utility company sent postal mailings to 1,000 gas customers, 500 of which were selected with the machine learning model and 500 who were selected randomly. Among the customers selected by the model, twice as many responded to the offer compared to the randomly selected customers.