not all customers are equal
a subscription business treated every customer the same. the top 8% were paying for the other 92%.
who are our actually-valuable customers, what do they share, and how much are we spending to keep the unprofitable ones?
segmenting by behaviour, not demographics, will reveal a small high-value cluster whose retention economics justify entirely different service.
3 years of transactions, support contacts and product usage for 480k customers. enriched with a single customer view stitched from web, app and call-centre identifiers.
built an scv with probabilistic matching, then ran k-prototypes clustering on rfm + engagement features. validated cluster stability with bootstrap resampling and cross-checked clv with a bg/nbd + gamma-gamma model.
collapsed nine legacy personas into four behavioural segments. routed the top segment to a named account team, automated the bottom two, and stopped discounting the middle.
service costs fell 18% in the unprofitable segments. top-segment retention improved 6 points. marketing finally had a definition of "good customer" that finance agreed with.
the question
finance saw a flat margin. marketing saw growth. both were right — and both were missing that the average customer did not exist.
the hypothesis
behavioural segmentation would expose a small, profitable core and a long tail the business was actively subsidising.
the data
three years of transactions, support tickets and product events for 480k customers, stitched into a single customer view with probabilistic matching across web, app and call-centre ids.
the analysis
k-prototypes on rfm plus engagement features. bootstrap resampling for cluster stability. a parallel bg/nbd + gamma-gamma clv model as an economic sanity check on the cluster ordering.
the decision
four segments, not nine. named accounts for the top cluster. fully automated journeys for the bottom two. no more blanket discounts in the middle.
the outcome
cost-to-serve down 18% on unprofitable segments. retention up 6 points on the top one. a shared definition of "valuable customer" between marketing and finance — possibly the rarest deliverable in the deck.
tools used
python · scikit-learn · lifetimes · dbt · snowflake