19 FAQ

19 · FAQ

Q — Why are the headline R²/AUC numbers ≈ 0? Is the project broken? No — that is the proven result on this specific dataset, which is statistically signal-free. The pipeline reaches cross-validated R² ≈ 0.85 on signal-bearing data (05-Predictive-Capability). The project demonstrates both competence and honesty.

Q — Isn't a "no-result" a weak portfolio piece? The opposite. Detecting, proving (permutation test + positive control), and communicating a signal-free, target-leaked dataset is the clearest demonstration of senior judgement here. Most pipelines would mask it.

Q — How do I know the model pipeline actually works? The positive control runs the identical pipeline on a known-signal target and reaches R² ≈ 0.86; SHAP recovers the true drivers; the learning curve rises with data. See 05-Predictive-Capability.

Q — Are the external APIs real or mocked? Three of four are validated live (World Bank macro, CO₂ emissions, FX); fuel is mock-first because the WB archived its pump-price indicator — reported honestly as mock. See 06-External-Data-Augmentation.

Q — Why generic-looking elasticities in the simulator? They are segment-specific luxury priors (own-price ≈ −0.3 with Veblen effects, income ≈ 2.2), from the automotive/luxury-goods literature, with Monte-Carlo uncertainty. See 09-Scenario-Simulator.

Q — Why DuckDB instead of a database? Portable, reviewable SQL directly over the CSV — no server, no ETL. Matches the project's zero-infrastructure ethos. See 10-SQL-Analytics.

Q — Why is deep learning included if it isn't used? To justify the choice with numbers. The MLP is benchmarked against boosting; both are no-skill, so boosting (cheaper, explainable) is the production choice. The DL code stays as evidence of the comparison (08-Machine-Learning-and-Deep-Learning).

Q — How do I run it? make app locally, docker run … ghcr.io/maxime2476/bmw-sales-analytics:latest, or open the live demo. See 15-Reproducibility-and-Commands.

Q — Where are the design decisions explained? Nine ADRs — 16-Architecture-Decision-Records — and on the docs site.

Q — Can I trust the numbers in the README? Yes — and CI enforces it: guard tests fail the build if the README's coverage gate or ADR count drifts from the actual config (13-Testing-and-Quality).

---

Back to Home.