The RFM model, a cornerstone of marketing analytics, can be transformed into a decisive tool for predicting churn and sales. This blog post will explore how we achieved this transformation through a structured process involving scoring, segmentation, benchmarking, and prediction.
At its core, the RFM model assesses customer engagement based on three key metrics:
Recency (R): Time elapsed since the last purchase or visit
Frequency (F): Number of purchases made in a given period of time
Monetary Value (M): Total spending by the customer. In retail-models this mostly the value a customer spent on the platform, in subscription or SaaS-models this is mainly the addition of the monthly fees, in advertising-models, this value-metric is often replaced by pageviews (which indirectly lead to monetization via advertisements per pageview). A RFV model is equivalent to RFM in that perspective.
These metrics are combined to generate an engagement score for each customer, providing a holistic view of their purchasing behavior.
1. Scoring
The first step is to score customers based on their RFM values. Each metric is assigned a score, typically on a scale of 1 to 10, with higher scores indicating greater engagement. These individual scores are then aggregated to arrive at a composite RFM score. For instance, a customer who has made recent, frequent purchases with high monetary value would receive a high RFM score, signifying strong engagement.
In our model, we used a RFM model to calculate an engagement score for every individual customer ID. This allowed us to quickly assess the level of engagement for each customer.
2. Segmentation
Once customers are scored, they can be grouped into segments based on their RFM scores. This segmentation allows for targeted marketing and customer relationship management.
Common RFM segments look like below, but can be customized in function of specific use cases.
We usually segment our customers into four groups based on their RFM scores: sleeping (0-20), passive (20-40), active (40-60), and ambassador (60-100). More than 4 segments make it harder to visualize insights per segment in clean charts.
3. Benchmarking
Benchmarking involves comparing the performance of different customer segments over time. Techniques like cohort analysis can be employed to track the retention and churn rates of various segments. This analysis helps identify patterns and trends, enabling data-driven decision-making for customer retention strategies
A cohort-analysis based on RFM-scores show that after 3 months most 'new' customers became already 'passive'
This information highlighted the need for a retention strategy. Therefor it's useful to plot the RFM-segments into their proportion of churn: which segments lead to churn?
In addition to what could been assumed, not only the passive customers churn. Also 'active' customers need activation in order to keep them engaged.
On product level, now we search for patterns amongst ambassadors that might serve as benchmark for less active customers.
By identifying these "engaging products" and contrasting them with "churning products," we gain insights into product categories that foster customer loyalty and those that contribute to churn.
4. Predictive value of RFM
This stage leverages the insights gained from scoring, segmentation, and benchmarking to develop predictive models for mainly churn and sales.
Churn Prediction:
Machine learning algorithms can be trained on historical RFM data, along with other relevant variables, to predict which customers are likely to churn in the future (more to be found in the technical documentation on Kaggle).
The emphasis should be on maximizing recall, which measures the model's ability to correctly identify all churning customers, even if it leads to some false positives. The Naive Bayes-model obtained 0,91 recall. This approach ensures that proactive measures can be taken to retain at-risk customers and minimize potential revenue loss.
Sales Prediction:
Time series forecasting techniques, such as those employed in the Darts model, can utilize past sales data and RFM scores to predict future sales for individual sellers (with logistic regression, an accuracy of 0,78 got obtained, versus 0,63 fur the dummy classifier). More to be found in the technical documentation on Kaggle). Further training is desired but this model can already be used to optimize inventory management, resource allocation, budget forecast and marketing campaigns.
The blue line represents the predicted sales of an individual customer, while the black line shows the actual sales. The visualization helps evaluate the model's ability to capture the sales trends per single ID.
Finally, thanks to the feature importance plot, we found that RFM seemed to be a significant feature for predicting both churn and sales:
the combined metric RFM as such
the underlying individual parameters R,F, M . This suggests that engagement, as measured by RFM, is a key driver of customer behavior.
By leveraging RFM data in our predictive models, we were able to improve the accuracy of our predictions and make more informed business decisions.
5. Turning insights into actions
Beyond prediction, the true value of the RFM framework and associated models lies in their ability to drive actionable steps that enhance customer engagement and business outcomes. The sources provide examples of how the insights derived from RFM analysis can be translated into concrete strategies and activations.
The insights gained from the RFM analysis and predictive models can be operationalized through targeted activations tailored to specific customer segments.
Examples are very diverse & depend on specific use cases. Contact us to overview those in detail. Here are some basic concepts:
Targeted promotions and offers: For customers identified as "sleeping" or "passive," targeted promotions or personalized offers can be employed to re-engage them and stimulate purchasing activity. These offers could focus on products or categories that align with the benchmarks from ambassadors & the customer's past behavior or preferences.
Content recommendations: analyzing product categories and identifying those that drive higher engagement.
Optimizing payment options (payment installments, pausing subscriptions, adjusting payment plans, offering incentives for upfront payments)... potentially reducing churn rates and improving revenue.
By combining predictive modeling with tailored activations, businesses can proactively address potential churn, nurture customer relationships, and drive revenue growth. The activation strategies should be informed by the specific insights derived from the RFM analysis, ensuring they are relevant and impactful for each customer segment.
Conclusion
By following a structured approach to scoring, segmentation, benchmarking, and prediction, we successfully converted the RFM model into a powerful predictive tool. This transformation enabled us to gain valuable insights into customer behavior, identify churn risks, forecast sales, and develop data-driven strategies to enhance customer engagement and ultimately boost revenue.
The RFM model, when thoughtfully applied and analyzed, can be a valuable asset for any business looking to understand and predict customer behavior. By embracing this data-driven approach, companies can unlock the full potential of the RFM model and drive sustainable growth through enhanced customer engagement and retention.
(Contact us for more information on the setup, implementation & usecases of RFM, more technical documentation & code snippets will be published later on Kaggle)
Comments