Data Scientist Roadmap for Complete Beginners

Data Scientist Roadmap for Complete Beginners

A data scientist spends the week pulling data with SQL, testing a specific hypothesis with statistics, and building a model in Python to answer a business question — then explaining the honest uncertainty behind the answer to people who are not statisticians. Realistic time to employable from zero is twelve to eighteen months part time, or six to nine months full time.

Updated for 2026. Salary figures are indicative and move quarterly.

What This Role Actually Does

The job ad says "leverage machine learning to unlock insights and drive innovation." The real week is narrower, slower, and far more about judgement than algorithms.

What people wrongly imagine this role is: training cutting-edge neural networks daily, discovering dramatic hidden patterns, and being treated as an oracle by executives who instantly act on every finding.

What it actually is: discovering that a "surprising" trend is actually a data collection artifact, and having to deliver that unglamorous but important finding to someone who was hoping for a more exciting answer.

The first surprise for beginners is how little of the job resembles a Kaggle competition. Most business problems do not need a deep neural network — they need a correctly framed question, clean data, and a model simple enough that you can explain why it works. The second surprise is how much the job depends on skepticism: checking your own result for a confound before presenting it, because a confidently wrong data scientist causes far more damage than an analyst ever could.

A third, quieter surprise is how much time goes into scoping the question itself. Stakeholders rarely arrive with a well-formed statistical question; they arrive with "can you look into why churn went up," and a large part of the actual skill is turning that into something testable — a specific cohort, a specific time window, a specific comparison group — before writing a single query. Beginners who skip this step and jump straight to modelling routinely produce technically correct answers to the wrong question, which is worse than no answer at all because it looks credible.

The title itself is one of the least standardised in tech. "Data scientist" job postings genuinely range from a data analyst role with a nicer title to a near-research role requiring published statistical work — read each posting's actual task list, not just the title.

The Roadmap

Six stages. Each assumes the one before it.

Stage 1 — Statistics Foundation

What to learn. Descriptive statistics, distributions, sampling and sampling error, the central limit theorem, confidence intervals, hypothesis testing and p-values with genuine conceptual understanding rather than a memorised threshold, correlation versus causation, and common bias traps such as Simpson's paradox and survivorship bias. Then experimental design basics: what makes an A/B test valid, and the difference between correlational and causal claims.

Realistic time estimate. Eight to ten weeks. This is the stage beginners most often rush, and it shows up as bad judgement later.

Free resources by name. Khan Academy's Statistics and Probability course, complete and free. StatQuest with Josh Starmer on YouTube for genuinely clear intuition on almost every topic in this stage. Seeing Theory (seeing-theory.brown.edu) for interactive visual explanations of probability concepts. OpenIntro Statistics, a complete free textbook available as a PDF.

Portfolio project. Design and run — or simulate — an A/B test end to end: state the hypothesis, calculate the required sample size, run the test, and write a one-page memo stating the result, the confidence interval, and what evidence would be needed before recommending a rollout.

How you know you are done. You can explain to a non-technical person why a small sample "significant" result is not trustworthy, without using the word significance.

Stage 2 — SQL for Data Scientists

What to learn. SELECT, filtering, sorting, every join type, GROUP BY and HAVING, subqueries, common table expressions, CASE logic, date and string functions, and window functions — ROW_NUMBER, RANK, LAG, running aggregates. Learn one dialect properly; PostgreSQL is the standard default because it is free and its syntax transfers cleanly to warehouse tools like BigQuery and Snowflake.

Realistic time estimate. Six to eight weeks.

Free resources by name. SQLBolt for interactive basics. Mode's SQL Tutorial, written specifically for data roles. Kaggle Learn's free Intro to SQL and Advanced SQL courses. pgexercises.com for graded practice against a real database.

Portfolio project. Load a real multi-table dataset into PostgreSQL and answer ten analytical questions that each require at least one join and one window function, publishing the queries alongside the questions.

How you know you are done. You write a query with a CTE and a window function without a reference, and you instinctively verify row counts after every join.

Keep the SQL cheat sheet open while you practise.

Stage 3 — Python for Data Science

What to learn. Python fundamentals if you do not already have them, then pandas for data manipulation, NumPy for numerical operations, matplotlib and seaborn for visualisation, and Jupyter Notebook as your working environment. Add requests for pulling data from public APIs.

Realistic time estimate. Six to eight weeks.

Free resources by name. Kaggle Learn's Python and Pandas micro-courses. freeCodeCamp's Data Analysis with Python certification. The official pandas documentation's "10 minutes to pandas" and user guide. Python for Data Analysis by Wes McKinney, free to read in its open online edition.

Portfolio project. Pull data from a public API on a schedule, clean it with pandas, and produce a notebook that identifies and visualises one genuinely non-obvious pattern, with the data collection step documented.

How you know you are done. You reach for pandas instead of a spreadsheet the moment a file exceeds about fifty thousand rows, without dread.

The pandas and NumPy cheat sheet covers the daily syntax reference.

Stage 4 — One Machine Learning Library, Properly

What to learn. Scikit-learn to real depth: linear and logistic regression, decision trees, random forests, gradient boosting, train/validation/test splitting, cross-validation, regularisation, feature engineering, and choosing the right evaluation metric — precision, recall, and the confusion matrix matter more here than raw accuracy in most real business problems.

Go deep on one library before touching deep learning frameworks. This is not the machine learning engineer path — production deployment and neural network depth belong there instead.

Realistic time estimate. Eight to ten weeks.

Free resources by name. Andrew Ng's Machine Learning Specialisation on Coursera, audit-free. Kaggle Learn's Intro to Machine Learning and Intermediate Machine Learning courses. scikit-learn's official documentation and user guide. Kaggle competitions for practice against messy real-world data.

Portfolio project. A complete classification or regression project on a real dataset with proper splitting, at least three model types compared against a sensible baseline, and a written justification for the chosen evaluation metric.

How you know you are done. Given a new dataset, you can correctly frame the problem type, choose an appropriate baseline, and explain why accuracy alone would mislead a stakeholder.

The machine learning cheat sheet is a useful reference during this stage.

Stage 5 — Communicating to Non-Technical Stakeholders

What to learn. The half of the job most courses skip entirely. Structuring a finding answer-first, choosing a chart type for the message rather than for variety, writing an executive summary that a busy stakeholder reads once, and defending a methodology calmly when someone senior disagrees with your number. Alongside this, enough business context — revenue, margin, retention, cohort thinking — to frame findings in terms the business actually cares about.

Realistic time estimate. Three to four weeks, then permanently ongoing.

Free resources by name. Storytelling with Data's blog and free chart guides. Google's Analytics Academy for digital metric fluency. Lenny's Newsletter's free posts for how product and growth teams reason about metrics.

Portfolio project. Take one earlier project's findings and rewrite the summary twice more: once for an engineering audience and once for an executive audience, with the same underlying data producing two genuinely different documents.

How you know you are done. A non-technical reviewer can state your project's conclusion correctly after reading only your summary paragraph, without opening the notebook.

Stage 6 — Job Search and Interview Practice

What to learn. Packaging everything above: a resume with quantified outcomes, three to four written-up portfolio projects with real messy data, and a GitHub or portfolio site that a reviewer can evaluate in under a minute. Then interview-specific practice: live SQL exercises, statistics conceptual questions, take-home case studies, and product-sense or experiment-design questions that test judgement rather than syntax.

Realistic time estimate. Eight to twelve weeks of active searching, running alongside continued practice.

Free resources by name. StrataScratch and DataLemur free tiers for SQL and statistics interview questions. Kaggle for continued practice datasets. Government open data portals — data.gov, data.gov.uk, Statistics Canada — for realistically messy source data for new projects.

Portfolio project. The search itself: track applications, response rates, and interview stages in a small dataset you analyse. It is a legitimate portfolio piece and it makes the search feel measurable rather than formless.

How you know you are done. You have a signed offer. Not before.

One practical note on this stage: many candidates over-prepare for the machine learning portion of interviews and under-prepare for the SQL and statistics portions, which is backwards, since most data scientist interview loops weight the fundamentals more heavily than an obscure modelling question. Drill SQL and statistical reasoning specifically in the final weeks, not just modelling.

Salary and Job Titles

Figures below are indicative ranges compiled from the pattern of public aggregates — Levels.fyi, Glassdoor and Indeed self-reported data, the US Bureau of Labor Statistics occupational data, and the Stack Overflow Developer Survey. These sources disagree with each other, they lag the market, and they move quarterly. Treat them as orientation, not as a quote.

United States (USD base salary, excluding bonus and equity):

Large tech companies and finance firms sit above these ranges once bonus and equity are included; smaller companies and non-tech employers typically sit below them.

Canada (CAD base salary — note this is CAD, not USD):

Toronto, Vancouver, and Montreal sit at the top of the Canadian range, with a persistent gap versus US compensation that remote roles for US employers partly close.

Who Should Not Take This Path

You want to build products, not answer questions about them. Data scientists rarely ship user-facing features themselves. If shipping something users directly touch is the satisfying part of tech for you, look at backend or frontend development instead.

You dislike deep engagement with uncertainty. Statistics is fundamentally the study of how confident you can be, not what is true. If ambiguity about "how sure are we" feels uncomfortable rather than interesting, this role will frustrate you constantly.

You want to skip straight to deep learning and production models. That combination of interests points toward the machine learning engineer path, which is a genuinely different job with more software engineering and less experimental design.

You dislike explaining the same finding multiple ways to different audiences. A large share of the job is translation — the same result explained differently to an engineer, a product manager, and an executive — and doing it well is not optional.

You want fast execution over careful reasoning. Good data science is often deliberately slow: checking a result for confounds before presenting it. If your instinct is to ship fast and iterate, data analytics rewards speed more directly than this role does.

The Five Mistakes

1. Jumping to machine learning before statistics and SQL are solid. Machine learning courses are the most exciting content available and the easiest to start, which is exactly why beginners over-invest there. Without statistics judgement and SQL fluency, the modelling has nothing solid underneath it.

2. Reporting a model's accuracy with no discussion of baseline or class balance. A stated ninety-four percent accuracy on a dataset that is ninety-three percent one class is a red flag to any experienced reviewer, not an achievement. Learn to report the metric that actually matters for the problem.

3. Never checking a "surprising" result for a mundane explanation. The most common cause of a dramatic finding is a data collection change, a tracking bug, or a duplicate join — not a genuine business insight. Rule those out before presenting anything.

4. Treating "data scientist" as one fixed job description. Job postings under this title vary enormously in what they actually require. Read the actual task list in a posting, not just the title, before assuming your skills match or your interview prep is correct.

5. Skipping the communication stage entirely. A technically excellent analysis that a stakeholder cannot understand or act on produces zero business value, and is indistinguishable from no analysis at all from the business's perspective.

Print This Section

DATA SCIENTIST ROADMAP FOR BEGINNERS — STAGE SUMMARY

1. STATISTICS FOUNDATION       8-10 weeks
   Distributions, CIs, hypothesis testing, bias traps
   Done when: you explain a false-significant result plainly

2. SQL FOR DATA SCIENTISTS     6-8 weeks
   Joins, GROUP BY, CTEs, window functions
   Done when: CTE + window function without reference

3. PYTHON FOR DATA SCIENCE     6-8 weeks
   pandas, NumPy, matplotlib/seaborn, Jupyter, APIs
   Done when: pandas beats spreadsheets at 50k+ rows

4. ONE ML LIBRARY (scikit-learn)  8-10 weeks
   Splits, cross-validation, metrics, regularisation
   Done when: correct metric choice on a new dataset

5. STAKEHOLDER COMMUNICATION   3-4 weeks, then ongoing
   Answer-first writing, chart choice, defending methods
   Done when: a stranger states your conclusion correctly

6. JOB SEARCH                  8-12 weeks
   Resume, 3-4 written-up projects, SQL/stats interview prep
   Done when: offer signed

TOTAL: ~12-18 months part time, ~6-9 months full time
BIGGEST RISK: chasing machine learning before stats and SQL

🔗 Next: compare the faster entry point in the data analyst roadmap, or start from the pillar, Tech Career Roadmaps Compared.