Intended use
The material is designed for learners with basic Python and pandas experience. A complete pass takes approximately 3–5 hours: 45–60 minutes of guided reading, 90–120 minutes of notebook practice, and 60–90 minutes completing the six project artifacts.
It can be taught as one half-day workshop, three shorter sessions, or a self-paced assignment. It is a compact module rather than a replacement for a full-semester data-science curriculum.
Session plan
| Segment | Time | Instructor action | Learner output |
|---|---|---|---|
| Course map | 10 min | Explain why modelling starts after decision framing. | Workflow sketch |
| Business and data | 35 min | Review target timing, baseline and leakage. | Decision brief and data profile |
| Preparation and modelling | 40 min | Demonstrate a pipeline and a leaky alternative. | Prepared-data workflow and comparison table |
| Evaluation and communication | 35 min | Discuss threshold, error slices and uncertainty. | Evaluation report and decision memo |
| Notebook labs | 60–90 min | Let learners alter one assumption and compare results. | Modified notebook or short result note |
Common misconceptions
- “More advanced model means stronger project.” The house-price example shows the opposite: the simple baseline wins.
- “Accuracy is the default target.” Ask what error costs and select a metric that reflects the action.
- “Preprocessing is separate from modelling.” Transformations must be fitted inside the training boundary.
- “A notebook that runs is deployable.” Delivery requires documented inputs, ownership, monitoring and review triggers.
- “A good chart proves the conclusion.” A chart explains evidence; it does not replace the research protocol.
Assessment rubric
| Criterion | Developing | Proficient | Strong |
|---|---|---|---|
| Decision framing | Names a model and target only | Identifies user, action, baseline and metric | Includes capacity, error cost and rejection conditions |
| Data validity | Describes columns after cleaning | Profiles coverage, missingness and target balance | Documents availability time and tests leakage risk |
| Reproducibility | Notebook runs once on one machine | Uses a deterministic split and saved parameters | Rebuilds from a clean environment and preserves evidence |
| Evaluation | Reports one headline score | Compares with a baseline and explains the threshold | Shows stability, error slices and decision consequences |
| Communication | Presents notebook cells sequentially | Provides a clear result and limitation | Gives a decision-ready recommendation and review trigger |
The project has automated evidence that the code, links and responsive layouts work. It has not yet been validated through a controlled study of learner outcomes. A useful next test is to ask 3–5 target learners to complete one case without assistance and observe where they become blocked.
Review questions
- Can the learner state what decision the model changes?
- Can the learner identify a field that would leak the future?
- Can the learner explain why a baseline is necessary?
- Does the evaluation recommend an action rather than just report a score?
- Could a second analyst rebuild the result from the repository?