Conrad Electronic is a European retailer of electronic products based in Germany. With 36 retail stores nationwide, the company wants to move towards innovation and digitalization, especially by improving its website offer.
To successfully represent these core values, Conrad wants to improve its customer’s digital experience. It is critical to detect when its customers are churning and to understand the causes by analyzing relevant data.
Conrad was exploring why their customers churn. Their previous approaches involved ad hoc and statistical implementations, which suffered from two main caveats: the relentless need for human input and the lack of generalization.
Although Conrad was also using Google Analytics 360 data which tracks user interactions on the website, they had made hardly any use of the stored data.
The challenge, thus, was fourfold:
- Clean and reformat Google Analytics 360 data.
- Extract relevant features from Google Analytics 360 data.
- Properly frame the machine learning task.
- Test the trained model – based on ensembles of decision trees – through diverse simulations.
Since Google Analytics 360 data already resided in Google BigQuery, the ETL pipeline for cleaning, reformatting, and obtaining informative features was also designed and implemented in BigQuery. As Google Analytics 360 data wasn’t structured to have the features directly, it took several transformations steps to engineer historical purchase patterns and session-level user behavior (search, time spent, etc.)
One significant key to the success of this project was the proper formulation of the machine learning task, More specifically, we fed the training pipeline with only the last snapshots of the session-level aggregated features. It would have been irrelevant to use all session-level points for the modeling phase.
In terms of machine learning evaluation, the trained model was performant. It still needed testing in real conditions, though, i.e., at any time during the customer’s session. We then created some simulations to run the model. Our simulation experiments were also promising and welcomed by the business side.
By implementing a propensity to churn model, Aliz has helped Conrad toward its ultimate goal of serving its customers better with a personalized experience offer. The churn model is just the first step of a longer journey including actionable insights to set out of the model’s predictions and full integration into the website.
We delivered the project in August 2019, so we don’t yet have enough data points to evaluate a significant business uplift.
However, a more important value came from the collaboration between Conrad and Aliz. Aliz contributed to motivating the Conrad Analytics team to take a more sustainable data-driven approach to analyzing their customers’ buying behaviors and provided hands-on experience in implementing it.