Simply enter your keyword and we will help you find what you need.
Who buys rookworst in the rain? How machine learning gives HEMA targeted insights into customer purchases
XomniaCasesWho buys rookworst in the rain? How machine learning gives HEMA targeted insights into customer purchases
"By delivering a number of state-of-the-art models and implementing our decision-making process, Xomnia made life easier and better for both our customers and HEMA. " Bas Karsemeijer, Head of Data at HEMA.
More about the case:
What they needed
What’s really driving, or stalling, sales? Like many retail stores, HEMA found it difficult to get a clear picture of the motivations behind customer purchases.
Category and store managers of the Dutch department store were still drawing on topline KPI’s, their own intuition and weather to make predictions or explanations. With 760 stores in nine countries, HEMA needed much more precise insights into the sales drivers to make data-driven decisions on stock availability, campaigns and promotions.
What we did
Together with HEMA, Xomnia developed an application on data including transactions, stores, customers, promotions, stock and even the weather and holidays for separate product categories and channels. The scores are shown to the business via a dashboard which allows for real-time tracking of actual sales volumes versus budget and predictions. A visualised salesbridge quantifies the financial impact per variable.
How it works
The application is written in Python and created for different models and different levels. For example, there are models specified for each product category in HEMA, and models for each channel (online divided into home delivery and click & collect, offline between franchise and own stores). Most of the features and data for those models are the same. However, there are slight differences. For example, offline shops are various sizes, but the online store has only one size. Xomnia used models on different levels to view the influence of features across the different circumstances. An example of this is that weather has a much greater impact on online sales than offline sales.
The weather data was collected by an API from nearly 30 weather stations and linked to the nearest stores. Then, we created additional features based on the data. Extreme weather is assumed if the wind speed is high, the temperature is above 25 celsius or below 0, snowy weather, etc. Because the weather can make a big impact on fashion choices, the model accounts for the first hot day, cold day of the season as a feature. We also added holiday data, including school holidays linked to the stores by school holiday region.
XGboost (Extreme Gradient Boosting) is used for the model. With XGboost, the model can take into account the influence of the features on each other. For example, a sunny day with a 25 degree temperature is different than a sunny day of 5 degrees.
What we achieved
The dashboard Xomnia built can revolutionise the company’s approach to sales forecasting and budgeting, stock management, in-store and online promotions, and performance management. HEMA’s management and specialists now have the tools to visualise and decipher the true impact of many internal and external, controllable and uncontrollable factors per product group, per store, per day.
Knowledge on this granular level reveals that a promotion with a huge value for the mens and kids departments may also deteriorate the margin on food. It can also show that, while rain has a negative impact on the general sales trends, some products like raincoats or even lingerie may benefit from a downpour.
Furthermore, collaboration between the business and development teams ensures that the model continues to leverage the company’s success through whatever events come their way – weather-related or otherwise.
“This communication is key because the financial, ecommerce, marketing and other sales teams must all be able to understand and explain how the model works,” explains Ilse, a Xomnia junior Data Scientist on the HEMA project. “It is not enough to develop a tool and then walk away. We are sitting together every week to fine-tune the model and to keep making it better.”