How to become a Data Scientist

A complete, honest roadmap for the skills data scientists actually use, from programming and statistics through data wrangling, exploratory analysis, inference, machine learning, experimentation, communication, productionising, and LLMs. It runs top to bottom, foundational to advanced, so you always know what comes next. Free to read, no signup required.

1. 01

   Step 01Programming foundations

   Data science runs on code. Before models or maths, get comfortable manipulating data programmatically, which is the everyday work under everything else.

   • Python: The default language for data science. Learn data structures, control flow, and working with data in code.
   • The data science libraries: The numerical and dataframe tools (NumPy and pandas style) that do the heavy lifting of real analysis.
   • Notebooks & reproducibility: Working in notebooks while keeping analysis organised and repeatable, so results can be trusted and rerun.
   • Good coding habits: Version control and clean, readable code, because a data scientist’s work is increasingly real software.

   PythonNumPy & pandasNotebooksGitClean code
2. 02

   Step 02Statistics & mathematics

   Statistics is what makes data science a science rather than guessing with data. This is the part that most separates a data scientist from a data analyst.

   • Probability & distributions: How likelihood and randomness work, the language behind models and uncertainty.
   • Descriptive & inferential stats: Summarising data honestly, and reasoning from a sample to a broader conclusion.
   • Linear algebra & calculus: Enough to understand how models are built and trained, at the level of intuition rather than proofs.
   • Correlation vs causation: The idea that runs through everything: two things moving together doesn’t mean one causes the other.

   The statistics is what keeps you honest. Plenty of impressive-looking analysis is wrong because someone skipped the reasoning that checks whether a result is real.

   ProbabilityDistributionsInferenceLinear algebraCorrelation vs causation
3. 03

   Step 03Data wrangling & cleaning

   Most of a data scientist’s time goes to getting data into a usable state. The quality of this work sets a ceiling on every result that follows.

   • Messy real-world data: Handling inconsistent formats, duplicates, and errors, which is what real data always looks like first.
   • Missing values: Deciding how to handle gaps without distorting the analysis, a judgement call more than a rule.
   • Reshaping & joining: Combining and restructuring data from several sources into one clean table to work from.
   • Getting data: Pulling data with SQL and from APIs and files, since it rarely arrives in one convenient place.

   Data cleaningMissing valuesReshapingSQLAPIs
4. 04

   Step 04Exploratory data analysis & visualization

   Before modelling, you look. Exploratory analysis is how you understand a dataset, find patterns, and spot problems worth knowing about.

   • Exploration: Summarising, slicing, and plotting data to build a real feel for what’s in it.
   • Visualization: Making clear charts that reveal patterns and outliers, both to understand data and to communicate it.
   • Spotting problems: Finding the outliers, biases, and data-quality issues that would quietly wreck a model later.
   • Forming hypotheses: Turning what you see into questions worth testing, rather than jumping straight to a model.

   EDAVisualizationOutliersPatternsHypotheses
5. 05

   Step 05Statistical inference & experiment design

   Drawing sound conclusions from data is a core data-science skill. This is where careful reasoning separates a real finding from a lucky-looking coincidence.

   • Hypothesis testing: Framing a testable question and judging whether the evidence really supports it.
   • Confidence & significance: Understanding uncertainty and significance so you don’t read noise as a real effect.
   • Experiment design: Setting up experiments and A/B tests that can actually answer the question you care about.
   • Causal thinking: Reasoning about cause and effect, and the traps that make correlation look like causation.

   Hypothesis testingSignificanceExperiment designA/B testingCausal inference
6. 06

   Step 06Machine learning fundamentals

   Machine learning is central to modern data science. Learn the fundamentals well, because they apply to every model and keep advanced methods honest.

   • Supervised & unsupervised learning: Learning from labelled data versus finding structure in unlabelled data, the two core paradigms.
   • Core algorithms: Regression, trees, and clustering, which solve a large share of real problems on their own.
   • Overfitting & generalisation: Building models that work on new data, not only the data they were trained on, the central ML challenge.
   • Evaluation: Choosing metrics and validation that reflect the real goal, since the wrong metric hides a broken model.

   Don’t rush past the fundamentals to chase the newest model. Overfitting and evaluation are exactly what separate a data scientist who ships reliable work from one who ships impressive demos.

   Supervised / unsupervisedCore algorithmsOverfittingValidationMetrics
7. 07

   Step 07Feature engineering & model building

   Turning raw data into signal and choosing the right model is where a lot of real value is created. It’s often more decisive than the algorithm itself.

   • Feature engineering: Crafting features that carry signal, frequently the single biggest lever on how well a model performs.
   • Model selection: Choosing a model that fits the problem and the data, and preferring the simplest one that works.
   • Tuning & validation: Improving a model with proper cross-validation, without fooling yourself by tuning to the test set.
   • Baselines: Comparing against a simple baseline so you know a complex model is actually earning its complexity.

   Feature engineeringModel selectionCross-validationTuningBaselines
8. 08

   Step 08Deep learning & modern methods

   Deep learning powers a growing share of data-science work, especially with text and images. You should understand when it helps and how it works, at least conceptually.

   • Neural networks: How networks learn, and the intuition for why they excel at unstructured data like text and images.
   • When to use it: Recognising that classic methods still win on much tabular data, and reaching for deep learning deliberately.
   • Frameworks: Enough familiarity with a deep-learning framework to build and train a model when the problem calls for it.

   Neural networksText & imagesWhen to useFrameworks
9. 09

   Step 09Communication & storytelling

   A model or finding nobody acts on is wasted. Communicating results to non-technical people is what turns analysis into decisions, and it separates good data scientists from great ones.

   • Explaining results: Translating a model or analysis into plain language and a clear recommendation.
   • Data storytelling: Building a narrative that leads to the point, supported by honest visuals, rather than dumping numbers.
   • Stakeholder work: Understanding what decision-makers need and answering the question they actually have.
   • Honesty about uncertainty: Being clear about confidence and limits, so decisions rest on an honest picture of what you know.

   The gap between a good data scientist and a great one is usually communication, not modelling. The insight only matters if it changes a decision.

   Explaining resultsStorytellingStakeholdersUncertainty
10. 10

    Step 10Productionising & working with engineers

    A model in a notebook rarely delivers value. Getting work into production, and collaborating with engineers to do it, is an increasingly expected part of the role.

    • From notebook to product: Turning exploratory work into something reliable enough to run in a real system.
    • MLOps awareness: Understanding deployment, monitoring, and drift enough to work well with ML engineers.
    • Data pipelines: Knowing how data reaches your model in production, and where it can quietly go wrong.
    • Reproducibility: Making results and models reproducible, which matters far more once real decisions depend on them.

    ProductionisingMLOps awarenessPipelinesReproducibility
11. 11

    Step 11LLMs & the modern data scientist

    Large language models have shifted what data science looks like. Knowing where they fit, and using AI tools well, is now part of the job.

    • Using LLMs: Applying language models to text problems, and knowing when a classic model is the better tool.
    • AI in the workflow: Using AI assistants to write code, explore data, and draft analysis, with your judgement checking the output.
    • Evaluating AI output: Judging quality when there’s no single correct answer, which needs care and honest evaluation.
    • Keeping the fundamentals: Anchoring new tools in statistics and evaluation, so you can tell real progress from hype.

    AI speeds up the mechanics of analysis, which raises the value of the judgement it can’t do: framing the question, checking the data, and standing behind the conclusion.

    LLMsAI in the workflowEvaluationFundamentals first