Reducing customer churn in fast food restaurants chain
The team faced the task to cluster the customer base for pizza restaurants in order to improve product analytics, define long term marketing activities and reduce customer churn.
Despite the active expansion due to the opening of new pizzerias and a positive unit of the economics, the company had a number of difficulties in analyzing the customer base, forecasting revenue, and quick testing of marketing hypotheses.
At the introductory stage we carried out a preliminary analysis of data quality. The study showed that although there are certain inaccuracies in data, like a number of orders have errors in delivery time (zero or more than 24 hrs), the data we got from customer’s CRM was well-structured and sufficient so we could detect errors and correct the result.
We could also draw certain conclusions about the efficiency of promotional codes and gift certificates. Some hypotheses on possible reasons for customers’ leave were rejected (tastes for pizza may very much vary
After sharing the results with our client it was decided to implement a full-fledged project.
Before using any machine learning models, data preprocessing was carried out:
- canceled orders were sorted out,
- orders in the restaurant and those to be delivered were separated,
- some preliminary procedures were performed with data.
Synthetic parameters turned out to be crucial as in any issue solved by means of machine learning — parameters derived from original data (for example, simple ones — the average time between orders, the time elapsed since the last order, or even some really ‘exotic’ like the sine of the slope of the order frequency curve.
During the course of work with recommendation systems we had an opportunity to test a large variety of options for such parameters for different categories of business, therefore, within the Unchurn product, this task is solved much faster than done from scratch.
To present the results in the BI system Qlik Sense, a personal dashboard that reflects both general business parameters (revenue by pizzeria and types of orders) and sheets with customer clustering by probability of churn and for planning marketing activities was developed.
In addition to the propensity to leave, we have selected some more customers’ groups like:
- customers ordering pizza to office,
- customers ordering pizza home,
- promo-code users,
- certificates users,
This way of segmentation allows to make more personalized offers for customers.
In addition, heat maps were compiled for the geography of orders to identify places potentially favorable for opening new restaurants of the chain.
In the end, we added data derived from Google Analytics to the dashboards to analyze which advertising channels clients with high LTV prefer.
Checking out the quality of the churn forecast
To test the accuracy of determining the propensity of customers prone to leave, our client planned a test promo.
SMS messages with a promo code were sent to each of the clustered groups of customers — “already left”, “active”, “about to leave”.
As expected, the most activity was shown by the “active” group, since they would make a purchase anyway, the activity of the group “already left” was less than 5% percent, the activity of “leaving” was slightly lower than the active ones and they had a lower average check, but significantly higher than those who, in accordance with our algorithms, finally left.
Selective telephone interviews with the “about to leave” showed that these customers were generally happy with pizza, but either order from different places occasionally, or have a limited budget for ordering food in a restaurant. Thus, a small discount or other bonus, as expected, encouraged these people make an order.
In the first few months, we managed to reduce churn rate by 9%, and increased the average check by 15%. The increase in the average check turned out to be a positive side effect of Churn System implementation. This was a feature of customers’ behavior when receiving promo codes, which was revealed at the stage of initial verification of data and hypotheses.
Why is it important?
One of the important factors influencing the duration of a customer’s purchases is how long the customer’s been staying with the company. That is, one-time purchases after a short time turn into a habit (especially if it is supported by good loyalty programs).
Thus, if we push a hesitant customer to buy at the right time, this will not bring only a one-time purchase, but also make him/her more loyal and encourage to continue buying the product without additional stimulus.