Ready to Move on from RFM Models? ​​

1990s handset and dinosaur

Marketing teams are often tasked with identifying the right audience for each of their campaigns and the usual reflex is to use some version of RFM-based segmentation. Popularized in the 1990s, RFM models target users based on the recency, frequency, and monetary values of past transactions or subscriptions. 

While RFM models were a useful shortcut when working with the limited data of the 90s, modern marketers have at their disposal much richer data about their users. Organizations not taking advantage of that data richness with modern machine learning algorithms are leaving a lot on the table.


6 Limitations of RFM

RFM models suffer from following known limitations:

  1. Limited data: Most RFM based models or segmentation logic use just one or two past behaviors and not the full interaction set.
  2. Missing data: RFM models ignore missing data instead of imputing or generalizing based on rest of the known data.
  3. Noisy data: RFM models, even those with quintiles or percentiles, suffer from anomalies in data due to rigid or limited binning buckets.
  4. Non numerical data: RFM models generally cannot take advantage of categorical data unless coerced into bins or levels that do not provide  insight.
  5. Non explainability: RFM models do not have explanatory power or insight into specific user behaviors and features that drive the predictions.
  6. Bias: RFM models suffer from bias in datasets and human errors, and cannot measure or correct for them.


Fuel Predictive Models With ML Algorithms

At Blueshift we empower marketers to combine multiple data streams into rich user profiles that are then input to modern machine learning algorithms. Specifically we auto compute several sets of rich feature sets per user that are then made available for all predictive models inside the platform. Sample feature sets include the following:

  • User Profile Features: User’s location, timezone and language attributes, content preferences, subscription or loyalty tiers, and custom attributes.
  • User Activity Features: User’s prior activity and derived aggregates on not just final orders or transactions but full funnel events leading up to those transactions.
  • User Campaign Features: User’s engagement with prior messaging campaigns and promotions across all channels and devices.
  • Item Catalog Features: Users’ prior  browse or bought items metadata like category, tags, title, brand, ratings, reviews, audio-visual features, etc.
  • Item Activity Features: Item popularity, category’s popularity, brand’s popularity, etc., of interacted or transacted items.


Why Rich User Data Is The Best

Modern machine learning algorithms like gradient boosted trees or multi layer neural networks are proven to be better at exploiting these rich feature sets and build propensity models that are much more precise while generalizing beyond linear boundaries. For example a user’s propensity to churn or to purchase again is influenced by several factors including hidden signals embedded in the data beyond the reach of RFM models.

It’s time for the marketing teams to move from the 1990s into the 2020s and take advantage of these modern algorithms and rich user data.