← All work

Kvantum Inc · Aug 2017 – Mar 2021 · Delhi

Keeping customers from leaving

Segmentation and targeted intervention modelling that moved churn from 40% to 23%, on a base large enough that a point is a lot of people.

Role

Senior DS, model owner

Base

5M+ customers

Outcome

40% → 23% churn

The stakes

Churn prediction is the most over-solved problem in applied data science and the most under-delivered. Predicting who will leave is straightforward. Changing whether they leave, at a cost per contact that does not wipe out the value of the retention, is the actual job, and the two are not the same objective.

At five million customers, every intervention has a real budget attached. The first version of any churn programme spends most of its money on two useless groups: people who were never going to leave, and people nothing was going to save.

The system, in three stages

1 · Segment by reason

recency, frequency, value and support history, to find out why they churn

2 · Model persuadability

uplift: persuadable vs lost cause vs safe

3 · Match the cheapest fix

intervention per segment, with a permanent untreated holdout

Three decisions, and what I turned down

01

Target persuadability

The people most likely to churn are often the people least likely to be talked out of it. Ranking by uplift, the modelled change in retention if someone is contacted, moved the budget onto the segment where contact actually changes the outcome.

Rejected: contacting the top decile of churn risk. The default everywhere, and it happily spends the whole budget on people who have already decided.

02

Segment by reason, then match the cheapest intervention

Price-driven churn, engagement-driven churn and service-failure churn need different responses, and they cost different amounts to address. Segmenting by cause let us send a discount only where a discount was the answer. A nudge costs a fraction of the price and works better on the disengaged.

Rejected: one blanket retention offer for everyone at risk. Simple to run, and it pays margin to people who would have stayed anyway.

03

Keep a permanent holdout

A slice of every eligible segment was never contacted, permanently. It costs a little retention and it is the only reason the 40%→23% claim means anything. Without it, seasonality and a growing base will hand you a beautiful chart that has nothing to do with your model.

Rejected: measuring against the prior period. Free, and unfalsifiable, which is precisely why it is popular.

What came of it

40% → 23%

churn on a five-million-customer base, measured against a permanent untreated holdout rather than last quarter.

SHAP

used to explain segment drivers to marketing teams, which is how the model came to shape campaigns instead of only scoring them.

This sat inside four years of consulting across CPG, retail, pharma, FMCG and QSR for brands including Coca-Cola, Ross Stores, Abbott, Reynolds and Garden of Life, doing marketing mix modelling, attribution, cross-channel models and market-basket analysis. Different industries, the same discipline: decide what you would have to see to be wrong, before you build.

← All work Next: AI search relevance →