Data Science Roadmap
Data Scientist
A data scientist turns messy data into a decision someone acts on. The sequence is Python and SQL first, then statistics, then the unglamorous majority of the job — cleaning, exploring and visualising — then machine learning, then experimentation and the communication skills that decide whether any of it gets used.
By the end: Take a business question from raw data to a recommendation people act on.
The Data Scientist Roadmap
Python and SQL~2 months
The two languages the job is actually done in.
Python fundamentals
Enough to manipulate data confidently — you are not building applications, but notebooks still rot without structure.
Pandas
Load, filter, join, group and reshape. The single most-used skill in the role.
NumPy
Arrays and vectorised operations, which is what pandas is standing on.
SQL
Most real data lives in a warehouse. Joins and window functions are the way in.
Notebooks and reproducibility
Jupyter for exploration, with enough version control that a result survives being questioned.
Statistics~2 months
What stops you presenting noise as a finding.
Descriptive statistics
Central tendency, spread and shape — and why the mean is often the wrong summary.
Probability and distributions
The distributions that keep appearing, and what each implies about a process.
Inference and hypothesis testing
Confidence intervals, p-values and the traps that generate confident nonsense.
Experimentation and A/B testing
Design, power and stopping rules. Frequently the highest-value thing a data scientist does.
Bayesian thinking
OptionalPriors and updating — useful whenever data is scarce and expensive.
Data Wrangling and EDA~2 months
The unglamorous majority of the job.
Cleaning and preprocessing
Missing values, duplicates, types and encodings. Expect this to take most of your time.
Exploratory data analysis
Look before you model. Most of the insight and all of the surprises arrive here.
Outliers and anomalies
Decide deliberately what is an error and what is a genuine extreme — that choice changes conclusions.
Visualisation
Charts that answer a question rather than decorate a slide.
Feature engineering
Turn raw columns into signal. Usually worth more than swapping the model.
Time series
Trend, seasonality and why random train/test splits are invalid on temporal data.
Machine Learning~2 months
Model when a description is not enough and a prediction is needed.
Supervised learning
Regression and classification, starting with the linear models you can explain.
Ensembles
Random forests and gradient boosting — the usual winners on tabular business data.
Unsupervised learning
Clustering and reduction for segmentation and exploration.
Evaluation and validation
The right metric for the business question, and a test set you have not contaminated.
Model explainability
SHAP and friends. A model nobody can interrogate rarely gets deployed.
Deep learning, awareness level
OptionalEnough to know when a neural network is the answer — on tabular data it usually is not.
Communication and Delivery~1 month
Analysis nobody acts on has no value, however correct it is.
Storytelling with data
Lead with the recommendation, then the evidence. Stakeholders are not reading your notebook.
Dashboards and BI
Put recurring questions somewhere self-serve so you stop answering them by hand.
Working with stakeholders
Translate a vague business question into something data can answer, then push back when it cannot.
Getting models into production
Enough of the deployment path to hand over cleanly, or ship it yourself.
Portfolio and Interviews~1 month
Prove it on data that was not cleaned for you.
End-to-end projects
Two or three analyses on real messy data, from question to recommendation. Avoid tidy competition datasets.
SQL interview practice
Often the largest single component, and the one candidates most underestimate.
Statistics and ML questions
Expect to justify a metric choice and to explain a model to a non-technical interviewer.
Resume and applications
Quantify outcomes — "cut churn 4%" beats "built a churn model" every time.
Frequently Asked Questions
Python or R for data science?
Python, in almost every commercial setting — it carries the whole ML ecosystem and the engineering path out of notebooks. R remains excellent for statistics and is still standard in academia, pharma and some research teams. Learn Python first unless your target industry says otherwise.
How much statistics do I need?
More than most bootcamps teach. Distributions, sampling, confidence intervals, hypothesis testing and the traps around p-values are load-bearing — they are what stops you shipping a confident conclusion from noise, which is the failure mode that damages trust in the whole function.
Do I need a masters degree?
It helps in research-heavy and regulated roles and is sometimes a hard filter, but plenty of working data scientists do not have one. The substitute is a portfolio of end-to-end analyses on real, messy data — not clean Kaggle sets, where the hardest part has already been done for you.
How much SQL does a data scientist need?
A lot, and it is underrated. Most working data lives in a warehouse, and joins, window functions and aggregation are how you get at it. Many data science interviews are more SQL than machine learning, because that is what the day job looks like.