Customer Transactional Activity: FMCD and MCG

Irina Skorynina27 May 2023

Problem

A bank wants to make personalized offers to its card customers. The team starts from RFM (recency, frequency, monetary) segmentation and needs more than it gives: how broad and how regular a customer's spending is, how the customer base shifts from month to month, and which spending category to offer next. The talk does not quantify the cost of the problem.

Example from the article: a customer made transactions worth more than half a million rubles, but only once or twice, in few categories. The score places this customer in the bottom segment.

Solution in brief

Extend RFM to four metrics (frequency, monetary, consistency, diversity), turn each into a decile score, and sum the scores into an engagement segment (top, middle, bottom). Track how customers move between scores month to month. Predict one to three "non-trivial" categories per customer, meaning categories unused in the last three months, with association rules over categories.

Questions this case answers
  • How to rank customers on several behavioural metrics relative to the whole base so that one large outlier transaction does not decide the segment
  • How to predict which unused category a customer is likely to adopt next using association rules over categories
  • How to segment bank customers by transactional activity
  • How to track month-to-month migration of the customer base between activity segments

Tools and methods

  • FMCD scoring — four metrics, ten buckets each, summed into a segment.
  • Association rules (support, confidence) — predict a new spending category from customers with similar behavior.
  • Apriori — the rule-mining algorithm that the team plans to replace with a faster one.
  • Python, Pandas, Dask — process parallelization over 16 cores.
Continue reading — it's free

Google or a one-time email code. No password.