Week 1 — Prediction Problems & Data · ~180 min
Weekly project: a written experiment plan
ProjectFree previewThe brief
Before anyone trains a model at a serious company, someone writes a plan. Write week1/experiment-plan.md framing two prediction problems on an authorised dataset — one regression and one binary classification. You may use the scikit-learn datasets, a public dataset with a clear licence (e.g. Kaggle, data.gov.ng, World Bank open data), or synthetic data you generate. Do not use data containing real people's personal information that you are not authorised to use.
Required sections, for each problem
- Decision — who acts on the prediction and how
- Observation, features, label — one sentence each; label units or classes
- Time of prediction — and a list of at least three features that are available then and one tempting feature that is not
- Cost of errors — what each type of mistake costs, and which is worse
- Baseline — named, and computed on a training split, with its score
- Data profile — rows, missing values, class balance, provenance, one representation concern
- Split strategy — random, stratified, or time-ordered, and why
- Metric — which one you will use and why it matches the cost of errors
- Leakage risks — at least two, with how you will prevent them
- How you will use the held-out test set — when, and how many times (the right answer is: once, at the end)
Deliverables
experiment-plan.md- A notebook that loads the data, prints the profile and computes both baselines
requirements.txt- AI-use note in your submission
Acceptance checklist
- Both problems framed with all 10 sections
- Baselines actually computed, not just named
- At least one "tempting but unavailable" feature identified per problem
- The test-set section says it is used once
- Notebook runs top to bottom on a fresh kernel
Reviewer checks
Can the learner explain, without notes, why their tempting feature leaks? Is the metric linked to the cost of errors rather than chosen by habit? Does the split strategy match the data (time-ordered data split by time)?
Sign in to submit your work.