◆ BMW Luxury Sales Analytics

Production-grade analytics, econometrics & decision intelligence for the BMW luxury-car market — honest ML/DL, external-API augmentation, SHAP, a Bayesian-flavoured Scenario Simulator, a DuckDB SQL layer, and a fully containerised, CI/CD-tested codebase.

Open the live dashboard Source on GitHub

Data source

Base dataset: BMW Sales Dataset on Kaggle by eshummalik. All macro/fuel/CO₂/FX context is added by this project's augmentation layer.

The defining principle — intellectual honesty

The dataset is structurally pristine but signal-free, and Sales_Classification is a leaked deterministic threshold on Sales_Volume. Rather than inflate metrics, this project detects, proves and communicates the issue — then delivers business value through a clearly labelled Scenario Simulator.

Headline, reproducible results

Check	Result
Permutation (label-shuffle) test	p ≈ 0.90 — indistinguishable from chance
Positive control (same pipeline, synthetic target)	R² ≈ 0.86 — pipeline is sound
Regression R² / leakage-free ROC-AUC	≈ 0.00 / ≈ 0.51
Leaked classification ROC-AUC	1.00 🚩

How the docs are organised

• Architecture Decisions — the nine ADRs that record every significant choice (architecture, the data-integrity finding, augmentation, DL justification, DevOps, the signal audit, SQL & quality gates, decision-under-uncertainty, and observability & docs).
• API Reference — auto-generated from the bmw_sales package docstrings.

Quickstart

make install-dev     # dependencies
make audit           # statistical signal audit (permutation + positive control)
make sql             # DuckDB business-insights report
make test            # full suite + coverage gate
make app             # launch the dashboard

---

Built by Maxime GOURGUECHON.