§ 01 Problem
A six-figure-row transactional sales dataset where the interesting patterns are buried under inconsistent records — the kind of table where any model trained on the raw data would mostly be learning the dirt. Before any modelling is honest, the shape of the data has to be understood: what sells, where, and how steadily.
§ 02 Approach
- Profiled the full transactional table before touching a single feature — 100K+ rows spanning 7 product lines and 19 country markets, roughly $354M in total sales, with deals bucketed into three size tiers (Small / Medium / Large).
- EDA surfaced the structure the raw table hides: Classic Cars dominate revenue at ~$122M — well ahead of Vintage Cars — the USA is by far the largest single market (ahead of Spain and France), and monthly sales read as a noisy, non-trending series across 72 months (2018–2024). The dark-theme charts render straight from the cleaned frame, so every figure on this page is inspectable, not asserted.
- Systematic cleaning and feature engineering in featureng.ipynb (21 code cells, plus a companion Colab notebook): normalizing categories, deriving sale-level features, and documenting each transformation so the lineage from raw column to engineered feature stays auditable.
§ 03 Outcome
A cleaned, charted, feature-rich dataset ready for downstream modeling — deliberately presented as what it is: the exploratory and preparation work done carefully, not a modeling result inflated into one.
100K+
transaction rows profiled
7 / 19
product lines / country markets
$354M
total revenue in the sample set
72
months of revenue trend (2018–2024)
§ 04 For the record
§ 05 Evidence
evidencefeatureng.ipynb — 21 code cells, cleaning + feature engineering
evidenceUpdated_Expanded_Auto_Sales_Data.csv — 100K+ rows, 7 product lines, 19 markets
