branch · ★☆☆

Product

A product analyst is the product's eyes. You watch how people actually use the app or the site, turn their behavior into numbers, and turn those numbers into decisions — what to fix, what to build, what to kill.

what you do

Day to day: you compute metrics (how many came back, reached checkout, got stuck in the funnel), build dashboards so the team can see the picture, and test hunches with A/B tests — 'is the new button really better, or did we just want it to be?'.

stack

SQL is your main language (pulling events and orders from the database). Python with pandas — for cohorts, retention, and A/B. A BI tool like Metabase — to surface metrics without code. Plus a pinch of statistics: proportions, confidence intervals, and an honest p-value.

who it suits

This branch is for you if you like understanding people, not just crunching numbers: why a user leaves, what keeps them, what makes them pay. It's the most hireable analytics track — everyone with an app or a website is hunting for product analysts.

how it differs

Unlike a marketing analyst (they're about ads and acquisition) or a BI analyst (more about reporting), you're about what happens INSIDE the product: behavior, retention, experiments. And unlike a data scientist, you care more about a fast, honest answer that drives a decision than about a fancy model.

L0

Read the product's pulse

a couple of evenings

Get a plain-English feel for the core product metrics: active users (DAU/WAU/MAU), conversion, average order value, revenue per user. Not hard math — just learn to read them and explain what each one means.

tool Google Sheets for first back-of-envelope work, or DuckDB (duckdb.org) — SQL straight over a CSV. Charts in seaborn or right inside a BI tool. dataset Any product dataset that hooks you: Olist orders (kaggle.com/datasets/olistbr/brazilian-ecommerce), Online Retail UK transactions (kaggle.com/datasets/carrie1/ecommerce-data), Telco churn (kaggle.com/datasets/blastchar/telco-customer-churn). Pick whatever you find interesting.
what you do A cheat-sheet: 6–8 key metrics, what each one shows and when it lies. Plus your first over-time chart on any product dataset.
Ready for the next level when:
  • You can explain in your own words how DAU differs from MAU, and why their ratio — stickiness — matters.
  • You know the difference between revenue, average order value, and ARPU — and when each one is the right lens.
  • You compute conversion as a proportion and always keep 'from what to what' in mind.
  • You spot when an average misleads because of outliers — a couple of whales drag the average order value up.
  • You plot a metric over time and read the trend, not just a single number.
L1

Funnels and cohorts

1–2 weeks

Learn to see the user journey as a funnel (where people drop off) and to slice users into cohorts — groups by entry time or trait — so you can compare them fairly.

tool SQL (DuckDB) to assemble the funnel, pandas for the cohort tables. dataset Olist (kaggle.com/datasets/olistbr/brazilian-ecommerce) — an order → payment → delivery funnel across 9 linked tables; or a synthetic event stream (github.com/bytewax/awesome-public-real-time-datasets) for DAU and a signup → activation → payment funnel.
what you do A step-by-step funnel with conversion between each step, and a cohort table (say, by month of first purchase).
Ready for the next level when:
  • You build a funnel from events and compute step-by-step conversion, not just the overall rate.
  • You find the funnel's bottleneck and form a hypothesis about why people are lost there.
  • You understand what a cohort is and can split users by entry time or by a trait.
  • You compare cohorts against each other without confusing 'different people' with 'change over time'.
  • You confidently join several tables (events, users, orders) to assemble a funnel.
L2

Retention: who comes back

1–2 weeks

Get to grips with retention — the heart of product analytics. Compute day and week retention, build a retention curve, and read its shape.

tool pandas (cohort pivots), seaborn for the heatmap, Metabase (metabase.com) to show it to the team. dataset Olist by customer_unique_id (do customers return); Online Retail UK for monthly cohorts; synthetic events for weekly retention. Pick by interest.
what you do A retention curve (D1, D7, D30 or weekly) and a cohort retention heatmap. Plus a takeaway: where the product holds people and where it loses them for good.
Ready for the next level when:
  • You compute retention as the share of a starting cohort that returns, and you get why that's not the same as 'active users'.
  • You build a retention curve and explain its shape: the sharp early drop, then the plateau.
  • You tell classic (N-day) retention from rolling/range retention and know when each fits.
  • You read retention as the key product-market-fit signal — a flat plateau above zero means the product genuinely matters to someone.
  • You tie retention to segments: which cohort or user type sticks around better.
L3

Stats basics and the A/B mindset

2–3 weeks

Add the statistical core: randomness, confidence intervals, hypotheses, and — honestly — the p-value. Understand why a difference in metrics can turn out to be just noise.

tool scipy.stats and pingouin (test plus effect size in one call); for intuition — Seeing Theory (seeing-theory.brown.edu) and rpsychologist.com/viz. dataset Palmer Penguins (t-test, seaborn.load_dataset), Students Performance, Tips — small clean sets so the focus stays on the method, not on cleaning.
what you do A two-group comparison with a verdict of 'real difference or noise': a confidence interval for the difference plus a correctly read p-value.
Ready for the next level when:
  • You state a null and an alternative hypothesis for a product question.
  • You explain the p-value HONESTLY: the chance of a result this extreme (or more) IF there were no effect — and you don't mix it up with 'the probability the hypothesis is true'.
  • You compute a confidence interval (including by bootstrap) and always report an estimate with its margin, not a bare number.
  • You pick the right test: a t-test for means, a proportions test or chi-square for conversions.
  • You understand Type I and Type II errors and what statistical power is.
  • You keep in mind that statistical significance is not the same as practical importance.
L4

A full A/B experiment

2–3 weeks

Run an A/B end to end: hypothesis → design and sample size → analyze the difference in proportions → confidence interval for the difference (bootstrap) → effect size → a business decision.

tool Python: pandas, scipy.stats / pingouin, a hand-rolled bootstrap (numpy). For intuition — StatKey (lock5stat.com/StatKey). dataset Cookie Cats — a real mobile-game A/B, gate at level 30 vs 40, D1/D7 retention (kaggle.com/datasets/mursideyarkin/mobile-games-ab-testing-cookie-cats). Another option — Marketing A/B (kaggle.com/datasets/faviovaz/marketing-ab-testing).
what you do A finished write-up of a real A/B: retention proportions compared across groups, a CI for the difference, the effect size, and a plain-words 'ship it or not' verdict.
Ready for the next level when:
  • Before touching the data you fix the hypothesis and the metric, so there's no after-the-fact p-hacking.
  • You get the role of randomization: why it, specifically, lets you claim causation rather than mere correlation.
  • You compare proportions across groups with the right test (proportions z-test or chi-square) and build a bootstrap CI for the difference.
  • You size the effect and answer the real question: is this lift even worth shipping.
  • You write a business takeaway in plain language — what to do with the result, not just 'p < 0.05'.
  • You watch for the traps: peeking before the test ends, multiple comparisons, sample bias.
L5

A product-analytics portfolio

finish · as long as it takes

Pull 2–3 finished cases into one story — from a product question to a decision. This is what you show at interviews and what makes you competitive.

tool Git + GitHub (portfolio), Metabase/Superset (dashboard), the whole Python/SQL stack. An optional AI layer — 'ask your data in plain language'. dataset Everything above, chosen by interest: Olist, Online Retail, Telco, Cookie Cats — or your own product / pet-project if you have access to its data.
what you do A GitHub portfolio: metrics and a dashboard for one product, a retention/cohort deep-dive, and a full A/B case (Cookie Cats, say) — each with a README of 'question → data → what I found → what to do'.

Finish: The finish line: you can take any service's raw product data and drive it to a conclusion the team can act on. You're competitive.

  • Each case lives on GitHub with a clear README — question, data, method, takeaway — readable by a non-analyst too.
  • You have at least one BI dashboard (Metabase/Superset) with product metrics, a funnel, and retention.
  • You have one A/B case taken all the way, with an honest verdict and an effect-size estimate.
  • Every case ends in a decision or a recommendation, not just a chart — you sell the conclusion, not the table.
  • In an interview you can walk through any case in 5 minutes: why you measured it, what you found, what you'd change in the product.