[The answer, in four sentences]
[What was predicted, for whom, how well, and what the business should do differently. Lead with the number.]
The question
Unit: [one customer]. Target: [churn next month, yes/no]. Horizon: [30 days]. Decision: [who receives a retention offer], owned by [role]. Today this decision is made by [gut feel / a spreadsheet rule].
Baseline to beat: [predict the majority class → 73.5% accuracy; or same-week-last-year].
Honest data
[Source, size, time span, what had to be cleaned.] Leakage check: [columns dropped and why — e.g., recorded after the outcome].
A validated model
[Pipeline in plain English: imputation, encoding, scaling inside a scikit-learn Pipeline; models compared; how hyperparameters were chosen; the validation scheme and why it matches production (stratified / grouped / time-ordered).]
| Model | Validation [AUC] | Test [AUC] | Notes |
|---|---|---|---|
| [Baseline] | [0.50] | [0.50] | [majority class] |
| [Logistic regression] | [0.83] | [0.82] | [interpretable] |
| [Gradient boosting] | [0.88] | [0.87] | [early-stopped at n trees] |
The decision, with a number
[Cost of a false alarm: $__. Value of a catch: $__. Break-even threshold: __. At that threshold the model recommends [action] for [N] units per [period], worth an estimated [$] versus the current practice. Assumptions stated honestly.]
The trust story
- Why it says that: [SHAP: top drivers globally, and one explained case].
- How sure it is: [calibration check; conformal 90% sets or intervals; share of ambiguous cases routed to a human].
- Where it fails: [error rates by group; the segment with the highest false-negative rate].
- After launch: [what is monitored, how often, and what triggers retraining].
Model card: intended use · training data · metrics by slice · known failure modes · owner · review date.
Tools, data, and AI assistance
[Data source and license. Libraries. Which AI assistants were used, for what, and how the output was verified — the split, the leakage, the metric, the baseline.]