Flagship course
Cohort Retention Lab
An eight-week working lab in Churn Prediction Analytics. You leave with a window diagram finance will sign, a leakage audit trail, and a one-page monitoring sheet — not a certificate wallpaper.
Learning outcomes
- Separate voluntary cancel, involuntary non-pay, and dormancy without collapsing them for convenience.
- Draw feature windows that follow invoices or top-up cycles, and explain the choice to a non-DS stakeholder.
- Run a leakage audit that catches future-known discounts, save-desk outcomes, and post-label ticket counts.
- Stress a first model on Thai calendar shocks and a simulated save campaign.
- Set a threshold against real queue capacity and write a monitoring sheet a CRM lead will actually keep.
Modules
-
Cohort windows and billing alignment
You map every candidate feature to a bill cycle or top-up cadence. Calendar-month defaults are treated as a smell until justified.
-
Survival curves without the academic fog
Time-to-exit for mixed pause and grace behaviour. Product partners join this clinic so the chart is readable outside data science.
-
Feature leakage audit
A deliberately severe session. Students present joins; the room hunts fields that could only exist after the label timestamp.
-
Thailand prepaid and postpaid quirks
Drought states, SIM rotation, tourist lines, family plans. Borrowed US cable lore is not accepted as evidence.
-
Alert design for account managers
Scores are rewritten as queue tickets with a four-minute action test. Handle time and campaign weeks are first-class constraints.
-
Model monitoring after launch
Drift, silent dbt breakage, Songkran volume, year-end campaigns. You leave with a one-page sheet, not a BI graveyard.
Instructor
Arunwadee Lim
Head of Applied Retention at Systemagentgrid. Previously led CRM science at a Bangkok MNO, where she spent more time arguing about involuntary non-pay definitions than about model class. She still reviews every leakage screenshot in the flagship clinic.
FAQ
Do I need production-scale data?
No. A desensitised extract with a stable account key, a bill or top-up date, and a few usage fields is enough. If you cannot identify an account across billing and app logs, wait — the lab will not invent that join for you.
Is this a vendor implementation?
No. We do not install software in your stack. You work in the notebook environment you already use.
What is a real limitation of this lab?
We do not teach deep learning recommenders, real-time feature platforms, or call-centre workforce management. If your bottleneck is agent rostering or a streaming feature store, this syllabus will feel incomplete — by design. We also cannot fix a warehouse that has no reliable subscriber key; several applicants are turned away for that reason each intake.
How large is a clinic?
Fourteen named seats. Recordings exist for missed sessions, but critique happens live.
Reviews of this lab
Module 3 was uncomfortable in the right way. We had been celebrating an AUC that depended on a save-desk field populated after contact.
★★★★☆
Reviewed after Module 3 — Feature leakage audit. Clear, a bit severe, and slower than our internal bootcamp.
The monitoring sheet is still on our wall. The live clinics ending at 16:30 Bangkok time remain awkward for Chiang Mai, which I mentioned on the Reviews page as well.