Data science projects: make the decision and evidence visible
A portfolio project is useful when another person can understand the question, reproduce the analysis and see why you chose the method. A large model does not compensate for a missing evaluation plan.
No ranking, interview or hiring outcome is guaranteed.
Four strong project shapes
- ForecastingPredict a future quantity with a time-aware split and compare against a seasonal or last-value baseline.
- ClassificationChoose a threshold around the real cost of false positives and false negatives.
- Experiment analysisEstimate an effect with uncertainty and state assumptions that could change the conclusion.
- Data productBuild a small reproducible pipeline or dashboard whose user and decision are explicit.
A portfolio-ready write-up
Question
Who needs the answer and what decision will it affect?
Data
Where did it come from, what is missing and what may be biased?
Baseline
What simple approach must a model beat?
Evaluation
Which split and metric match the use case?
Results
What changed, how uncertain is it and where does it fail?
Reproduction
Can another person run the code and obtain the same output?
Show the decisions between the code cells
A reviewer can see code in the repository; the write-up should reveal judgment. Explain why the target represents the problem, why the split matches how the system would be used, why the baseline is credible and why the chosen metric reflects the costly mistakes. Record alternatives you rejected rather than presenting the final pipeline as inevitable.
Include one failure analysis with examples. A classification project might compare errors across important groups; a forecast might show periods where the baseline wins; an experiment analysis might test sensitivity to an assumption. This turns limitations into evidence that you know how to investigate a result instead of marketing a single score.
A repository another person can review
- A short entry pointThe README states the question, data source, setup, main result and limits before asking the reader to open a notebook.
- A reproducible pathDependencies, data instructions and commands are precise enough for another person to obtain the documented output.
- Outputs are intentionalKeep the tables and figures that support the decision; remove exploratory noise that no longer contributes to the explanation.
Questions
How many data science projects do I need?
A few complete, explainable projects are stronger evidence than many copied notebooks. Quality and relevance matter more than a fixed count.
Can I use a common public dataset?
Yes, if your question, evaluation and analysis add something beyond following an existing tutorial.
How many data science projects should a portfolio include?
A few complete, distinct projects are more useful than many unfinished notebooks. Each should make the question, decisions, evaluation and limits easy to inspect.
Do failed experiments belong in a portfolio project?
Selected failures do. Explain what you expected, what happened, how you investigated it and what decision followed. Remove only noise that teaches the reader nothing.