A US market leader in direct-to-consumer clothing.
Each year, the company spends about US$200 million on marketing, mailing 140 million catalogues to existing and prospective customers. Declining response rates on catalogues was, in turn, leading to fewer sales, tighter margins and fewer customers. Faculty was commissioned to explore how the latest techniques in machine learning could be used to optimise the company’s catalogue marketing spend.
To optimise its marketing, the company only wanted to contact customers whose purchasing decision would be impacted by receiving a catalogue.
Previously, the company had used a propensity-based machine learning model, which identified the segments of customers that were most likely to buy. But this propensity modelling approach couldn’t indicate whether a customer’s purchase decision resulted from receiving a catalogue, or whether they would have purchased anyway, resulting in “over-mailing” to certain customers and so wasted marketing spend.
We replaced the propensity-based model with an ‘uplift’ model. By calculating the difference between two quantities – the likelihood that a customer will purchase if they receive a catalogue, and the likelihood that they will purchase if they do not receive one – the model can identify which customers are most likely to have their purchase decision influenced by receiving a catalogue.
Working closely with the client, we ran randomised controlled trials (RCTs) to collect the data required to train the model. We then moved to implement this pipeline as an automated weekly procedure, hosted in their AWS infrastructure, which now provides their catalogue mailing team with a ranked list of customers to mail each and every week.
Our model is being used as the leading strategy for the company’s largest brand, driving around a 5% increase in profitability compared to the company’s original propensity-based model. This is equivalent to $4 million in incremental profit in 2020 across the company’s entire customer base.
The solution we deployed also allows Faculty to help this company continuously run small-scale tests, using the results of these tests to refine and re-train our models. As a result, the model can respond to changing customer and market forces – a necessity in the highly dynamic clothing retail industry.