Project

Olist EDA — Brazilian e-commerce

Exploratory analysis of 100k+ Olist orders (2016–2018), answering 5 business questions with pandas and matplotlib — revenue trends, category mix, geography, delivery impact on reviews, and customer retention.

title: Olist EDA — Brazilian e-commerce blurb: Exploratory analysis of 100k+ Olist orders (2016–2018), answering 5 business questions with pandas and matplotlib — revenue trends, category mix, geography, delivery impact on reviews, and customer retention. personas: [data] stack:

• Python
• pandas
• numpy
• matplotlib
• seaborn
• scipy
• JupyterLab links: github: https://github.com/SyrineLarbi/ecommerce-eda nbviewer: https://nbviewer.org/github/SyrineLarbi/ecommerce-eda/blob/main/notebooks/01-olist-eda.ipynb

What it is

Five business questions a real e-commerce exec would ask of the Olist Brazilian E-commerce dataset (9 relational CSVs, ~100k orders, 2016–2018), each answered as one chart + one quotable finding.

Highlights

• Revenue trend — 5.4× growth in 11 months; Black Friday 2017 alone hit BRL 1.2M.
• Categories — PCs and agro/industry are revenue-heavy but volume-light (premium); garden tools dominate volume without leading revenue.
• Geography — Top-5 states drive 75% of orders and 76% of revenue. The market is heavily São Paulo–weighted.
• Delivery → review — Spearman correlation −0.31 between delivery days and review score; deliveries past 26 days drop the average review by ~1.8 stars.
• Retention — 96.9% of customers order once and never return. The growth came from new acquisition, not repeat business.

What it took

• Plumbing — 9 raw CSVs → one cleaned delivered orders frame with delivery_days, order_year_month, and joined order_revenue. Foreign-key sanity checks before any join so dropped rows would have been caught early.
• Analysis — five independent notebook sections, each a tight aggregate → plot → finding. All five plots saved to figures/ so the README can embed them directly.
• Reproducibility — pinned requirements.txt, Kaggle-CLI dataset download, make-like setup in the repo README. Anyone can git clone + pip install -r requirements.txt + run the notebook end-to-end.

Stack at a glance

pandas · numpy · matplotlib · seaborn · scipy · jupyterlab — deliberately the smallest dependency set that does the job. No notebook magic, no hidden helpers.