Present an unsuccessful model project as honest learning

A project does not become successful because its final slide is positive. If the candidate loses to a valid baseline, the technically credible conclusion is to reject it under the tested conditions. The portfolio value comes from the quality of the decision and evidence.

State the losing result directly

The career evidence lab uses an invented forecast.

from career_cases import honest_failure_case

result = honest_failure_case()
assert result["failure_not_relabelled_as_success"] is True
assert result["narrative"]["baseline_mae"] == 8.2
assert result["narrative"]["candidate_mae"] == 9.7
assert result["artifact_count"] == 3

The candidate MAE is 9.7 versus 8.2 for seasonal naive. The project rejects the candidate, notes a structural break and weak post-break sample, and retains the backtest, residuals and decision memo. These values belong to an authored fixture, not a real forecasting system.

Show why the comparison is fair

Describe the forecast origins, horizons, known-at-prediction-time fields and metric policy. Confirm that both candidates faced the same rolling windows. Report performance by horizon and before/after the break. If the sample is too small for a stable conclusion, say so; do not promote an inconclusive slice as a win.

Then distinguish findings from hypotheses. Residual evidence may show errors increased after a break. It does not by itself prove the causal reason. A next test could collect more post-break periods, revise a known-future covariate or evaluate an intervention-specific forecast.

Build the presentation around the decision

Use five slides or sections: decision and baseline, evaluation design, primary result, error evidence, recommendation and next evidence. Include the rejected configuration and reproduction command. Explain what would make you revisit the decision.

Avoid hiding the baseline in an appendix, changing metrics after seeing results or calling failure a “90% improvement opportunity.” Honest negative results help organizations avoid needless complexity and show reviewers that you can protect a decision from your attachment to a model.

The Data Science course preserves adverse results in its draft labs so learners can practise this defence.

Exercise

Take a project where your preferred candidate lost or became inconclusive. Rebuild the five-section presentation, label every causal interpretation as evidence or hypothesis, and ask a reviewer whether the rejection rule was defined early enough.

Continue learning

This article is part of the Data science careers and portfolio decisions sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Plan a domain-expert transition into data science.
• Continue with Read a machine learning job description critically.

Reference: NIST engineering statistics handbook on model validation.