Predicting the daily sales of its e-commerce site through machine learning: this was the goal of the brand Sephora, which explained at the Big Data Paris show, on March 12 and 13, 2018, the benefits for marketing. “With an e-commerce site deployed in several countries, we needed, to drive our online business to make forecasts, to measure the reliability of daily sales forecasts,” says Clément Marchal, Data Manager Manager of the brand Sephora, on the occasion of the Paris Big Data Fair 2018.
Useful for sales, the forecast (the quality of commercial forecasts) is no less for marketing – to drive its marketing plan – or for customer service, to anticipate the volume of additional contacts in the process of contacting him. Managed manually by the marketing department of Sephora, the forecast represented a “heavy investment of time for the teams”, explains Clément Marchal.To enable marketing to fully focus on the core of its business, Sephora has embarked on a project of machine learning. And this, with several stakes: data (on which sources of data to rely?); machine learning (which algorithms to choose?) and process (how to give trades simple forecasts to take in hand?).
A project “data” and IA
Transversal, the project that mobilized the IT teams, data science, e-store, and the Avisia and Dataiku partners – to develop and industrialize the algorithms – is based on a good dose of data. The idea: to evaluate the impact of various factors in the prediction via the collection and analysis of data from the marketing plan (marketing planning, typology of offers, calendar of events and highlights) and navigation data on the site and the mobile application (visits, transactions, orders and amounts, in particular). These data are integrated into the data lake deployed by the brand, automatically, to D + 1. Then, structured, the data come to feed the forecast.
Once the type of data selected, which algorithms to choose to optimize the supply chain, from production to sale? Sales and order detail are the key factors for the e-commerce site. “We have uncovered algorithms integrating short, medium and long-term activity trends, and algorithms integrating the marketing context, namely the marketing plan and offer and the calendar of events,” says the Manager. Data Science of Sephora.
To make trades a simple forecast, the brand has deployed a process in several stages: in the first place, the business submits its marketing plan, the operational actions planned in a given time, on the data lake, connected to the algorithms of Dataiku … that scan the data over two years. The trade receives in return an email containing the forecast, with a forecast line, per day. “It is thus possible for the marketing to test several offers and several events”, reveals Clément Marchal, which notes a gain of autonomy of the marketing teams. Once the marketing plan is validated, sales forecasts are also validated. These are then disseminated to marketing, supply chain, and customer service. Advantage? “The job retains control of forecasts to ensure the consistency of its actions,” the professional adds.
Thanks to machine learning, Sephora claims an optimization of forecasts. The brand notes a gain of 15 points of precision brought by the machine learning between the forecasts and the realized turnover. And a process optimization for marketing, which frees itself from a workload to focus on its business.