Data Science Roadmap 2026: A Complete Step-by-Step Guide

Many data science roadmaps are lists of topics. The trouble with a list is that it cannot tell you when to stop studying one thing and start the next, so people either rush into machine learning before they can clean a dataset, or stay in tutorials for months because there is always one more topic. This roadmap puts the topics in dependency order and gives every stage an exit test: something you should be able to do, unaided, before moving on.

One idea runs through all of it. Pick one real dataset early and keep working on it at every stage. Clean it in stage 3, question it with statistics in stage 4, model it in stage 5, and explain it in stage 6. By the end you have a project with a history instead of a folder of unrelated notebooks.

First, check that it is data science you want

Data science, data analytics and data engineering share tools but answer different questions. Analysts explain what happened and why. Data scientists build models that predict or estimate something they can't observe directly. Data engineers build the systems that move and store data. If you mainly enjoy dashboards and business questions, the analytics path gets you working sooner. Our data science vs data analytics comparison sets the two side by side. The rest of this roadmap assumes you want to build and evaluate models.

Stage 1: Python for data work

You don't need all of Python. You need enough to load data, reshape it and write small functions without looking up syntax every line: variables and types, lists and dictionaries, loops and comprehensions, functions, reading files, and handling errors. The official Python tutorial covers this ground well. If you are starting from zero, follow our Python roadmap up to its data branch and come back.

Then learn pandas and NumPy, the two libraries nearly every data science workflow is built on. The pandas guide 10 minutes to pandas is a good map of what to learn: selection, missing data, grouping, merging, reshaping and time series.

Exit test: given a CSV you have never seen, load it, fix the column types, handle missing values, and produce a grouped summary. Do it in under an hour without a tutorial open.

Stage 2: SQL

In most companies the data you need sits in a database, and you will be expected to query it yourself. Learn joins, aggregation with GROUP BY and HAVING, subqueries and CTEs, and window functions. Those four cover much of the SQL a data scientist writes. The PostgreSQL tutorial is a solid free reference, and SQL for data science lists the query patterns that come up most.

Exit test: write a query that returns each customer's first and most recent order, their total spend, and their rank within their city. Then explain why you used a window function instead of GROUP BY for the rank.

Stage 3: Cleaning and exploratory analysis

This is where your chosen dataset comes in. Exploratory data analysis (EDA) means looking at distributions, missing-value patterns, outliers and relationships between variables before you model anything. It is not a formality. Many of the decisions that make or break a model are made here: which rows are unusable, which columns leak the answer, and what a sensible baseline looks like.

Exit test: a short written EDA of your dataset. List three findings, each with a chart, and state one thing about the data you are still unsure of.

Stage 4: The statistics you will actually use

Statistics is what lets you say how confident you are, and it is an easy stage to skimp on when you learn alone. Focus on:

• Distributions and sampling. Why a sample mean varies, and what a standard error measures.
• Confidence intervals and hypothesis tests, including what a p-value does not tell you.
• Correlation versus causation. Why an A/B test can support a causal claim and an observational dataset usually cannot.
• Linear and logistic regression, as statistical models before they are machine learning models: coefficients, assumptions and residuals.

An Introduction to Statistical Learning is free to download in R and Python editions (the Python edition, ISLP, was published in 2023), and it connects this stage to the next. Our statistics for data science article narrows the list further.

Exit test: on your dataset, estimate one difference between two groups with a confidence interval. Then write one paragraph on whether the difference is causal, and why or why not.

Stage 5: Machine learning, with honest evaluation

Start with supervised learning: regression and classification with linear models, decision trees and one ensemble method such as random forests or gradient boosting. The algorithms matter less than the evaluation discipline, and three habits separate a real result from an accidental one:

1. Always compare against a baseline. A model is only as good as its improvement over the simplest possible guess.
2. Estimate performance with cross-validation, not one lucky split. Our cross-validation guide explains how to choose the splitter your data needs.
3. Keep preprocessing inside a pipeline. scikit-learn's guide to common pitfalls is blunt about it: never call fit on test data. A pipeline guarantees that scaling and other steps are fitted only on each training fold.

The whole habit fits in a dozen lines:

from sklearn.datasets import load_breast_cancer
from sklearn.dummy import DummyClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

X, y = load_breast_cancer(return_X_y=True)

baseline = DummyClassifier(strategy="most_frequent")
model = make_pipeline(StandardScaler(), LogisticRegression(max_iter=1000))

for name, est in [("baseline", baseline), ("logistic regression", model)]:
    scores = cross_val_score(est, X, y, cv=5, scoring="balanced_accuracy")
    print(f"{name:>20}: {scores.mean():.3f} ± {scores.std():.3f}")

Run with scikit-learn 1.8, it prints:

            baseline: 0.500 ± 0.000
 logistic regression: 0.977 ± 0.010

The baseline scores 0.5 on balanced accuracy because it always predicts the majority class. That is the number the model has to beat, not zero. The spread across folds tells you how much the estimate would move with different data. And because the scaler sits inside the pipeline, it is refitted on each training fold and never sees the rows it is scored on.

For the order in which to learn the algorithms and the maths behind them, the machine learning roadmap goes deeper than this stage needs to.

Exit test: a model on your dataset that beats a baseline under cross-validation, with a pipeline, and a written note of one way your estimate could still be too optimistic.

Stage 6: Communicating results

A model nobody understands does not get used. Practise turning a notebook into a one-page summary: the question, the answer, how confident you are, and what should happen next. Use a chart only where it carries the argument, and state limitations plainly. Interviews often test it directly, for example by asking you to explain a project to a non-technical manager.

Exit test: someone outside data science reads your summary and can repeat the main finding and its main caveat back to you.

Stage 7: Specialise, but only one direction at a time

Only now does it make sense to go deep. Choose one direction based on the jobs you want:

Trying to learn all four at once usually leaves you shallow in each.

What to leave out at the start

Big-data tools such as Spark, cloud certifications, and the latest model releases all feel urgent, and none of them is usually a prerequisite for a first data science job. Spark matters when data no longer fits on one machine, and your learning datasets will. Cloud skills matter once you deploy something. Add them when a project or job description demands them, not before.

Turning the roadmap into a portfolio

By stage 6 your running project is a portfolio piece, because it shows cleaning, statistics, modelling and communication on one problem. Add one or two projects in your chosen specialisation, with a README that states the question, the data, the evaluation and the result. Our list of data science projects for freshers suggests datasets and problems at this level.

If you want to follow this sequence in a structured course, the GloryTecks Data Science course covers Python, SQL, statistics and machine learning, with projects.