Learn
Six short explainers with a two-question check. Progress and your daily streak are saved in this browser only.
Pick one customer who churned and one who stayed, at random. AUC is the probability that your model gave the churner the higher score. 0.5 is a coin toss; 1.0 is perfect ranking.
It measures ranking, not calibration: a model can rank well and still be wrong about the actual probabilities. Gini is just 2 × AUC − 1.
A model has AUC 0.80. What does that mean?
AUC is about ranking pairs, not accuracy.
Gini for AUC 0.75 is…
2 × 0.75 − 1 = 0.50.
Sort customers by model score and cut them into ten equal deciles. Lift is the churn rate in a decile divided by the overall churn rate.
A top-decile lift of 3.5× means that slice contains three and a half times the average concentration of churners. The useful question is where the curve flattens: that is where extra contact stops paying.
Top-decile lift is 3×. Overall churn is 10%. Churn in the top decile is…
Lift × base rate = 3 × 10%.
Why contact only the top deciles?
Diminishing returns, not zero returns.
Any single target invites gaming: by people, and even more relentlessly by optimising algorithms. That is Goodhart's law at work.
Pair each primary metric with one that would reveal gaming: speed with repeat contacts, closures with complaints, conversion with long-term retention. Then look at samples of actual behaviour, not just the dashboard.
A call centre is targeted on average handle time. Best counter-metric?
Rushed calls come back.
Goodhart's law says…
The classic paraphrase.
Common dimensions: completeness (is it there?), validity (does it fit the rules?), accuracy (is it true?), consistency (do systems agree?), timeliness (is it fresh enough?) and uniqueness (no duplicates).
Cheap defences: a metric dictionary, automated checks on volumes, nulls and ranges at each pipeline step, a named owner per critical dataset, and known-answer tests on every refresh.
Two systems report different balances for the same account. Which dimension?
Systems disagreeing is a consistency failure.
Best first step before building a dashboard?
Agree definitions first.
India's Digital Personal Data Protection Act, 2023 and the DPDP Rules, 2025 apply to organisations that decide why and how digital personal data is processed (Data Fiduciaries). Core obligations apply in phases, with the main set 18 months after the Rules were notified.
Key themes: clear notice, valid consent or a legitimate use, reasonable security safeguards, breach intimation, rights for individuals, retention limits and extra care with children's data. Check the official text and take legal advice.
Who carries the main obligations?
The fiduciary decides purpose and means.
Withdrawing consent should be…
The Act requires withdrawal to be comparably easy.
Automation decisions are design decisions. Three questions help: Is the action reversible? Who bears the cost if it is wrong? Could a customer or regulator ask us to explain it?
Low-stakes, reversible, explainable actions can be automated aggressively. Anything touching money, access or dignity gets a human checkpoint, a log and an owner.
Which is the best candidate for full automation?
Reversible and low-stakes.
A human checkpoint should come with…
Accountability needs a record and a person.
General explanations for learning. Not financial, legal or tax advice.