Career roadmap
Data Scientist
Turn messy data into a decision someone acts on, and be able to say how confident you are.
Before you start Data Scientist
- Python and pandas at working level
- SQL: joins, aggregation, window functions
- Secondary-school statistics you are willing to rebuild properly
Statistics that survive scrutiny
The part candidates skip and interviewers test hardest.
The vocabulary underneath every model and every test.
- Random variables and common distributions
- Expectation, variance and covariance
- Conditional probability and Bayes
- Central limit theorem and why it matters
Hypothesis testing is asked in almost every data science interview.
- Sampling distributions and standard error
- Confidence intervals and their interpretation
- p-values, and what they do not mean
- Type I and II errors, power analysis
Linear models remain the most-used and most-interviewed technique.
- Linear and logistic regression
- Assumptions and diagnostics
- Multicollinearity and regularisation
- Interpreting coefficients honestly
The fastest-growing expectation for the role, and a strong differentiator.
- Correlation versus causation in practice
- Confounders and Simpson's paradox
- Difference-in-differences and matching
- Instrumental variables and natural experiments
Most wrong conclusions come from the data, not the model.
- Selection and survivorship bias
- Missing data mechanisms and imputation
- Outliers and robust statistics
- Measurement error
BuildAnalyse a public dataset and write up the findings with confidence intervals and stated assumptions.
Working with data
Most of the job is acquiring, cleaning and shaping data before anything interesting happens.
The most-used tool in the role and a guaranteed interview round.
- Joins, aggregation and grouping sets
- Window functions for cohorts and retention
- CTEs for readable analysis
- Query performance on large tables
pandas fluency, and knowing when to reach for something faster.
- pandas: reshaping, joins, groupby
- NumPy vectorisation
- Polars or DuckDB for larger data
- Reproducible environments
Structured exploration rather than plotting everything and hoping.
- Univariate and bivariate exploration
- Distribution checks and transformations
- Segment analysis
- Documenting what you found and rejected
A chart that misleads is worse than no chart. This is assessed directly.
- Choosing the right chart for the question
- Uncertainty in visualisations
- Avoiding misleading axes and scales
- Narrative structure for findings
An analysis nobody can rerun is an opinion, not a result.
- Notebook discipline and refactoring to scripts
- Version control for analysis
- Data versioning and lineage
- Parameterised, rerunnable pipelines
BuildAn end-to-end analysis from raw source to reproducible notebook with a documented cleaning log.
Machine learning for decisions
Enough modelling to build something useful, with rigorous evaluation.
Gradient boosting still wins most tabular problems, and interviews know it.
- Tree ensembles and gradient boosting
- Feature engineering for tabular data
- Hyperparameter tuning discipline
- When linear models are the better answer
Choosing the metric is a business decision disguised as a technical one.
- Precision, recall, ROC and PR curves
- Class imbalance handling
- Cross-validation and leakage
- Calibration and threshold selection
Segmentation and dimensionality reduction, used carefully.
- Clustering and choosing k
- PCA and embeddings
- Anomaly detection
- Validating unsupervised results
Forecasting is common in business contexts and has its own traps.
- Seasonality and trend decomposition
- Backtesting without leakage
- Classical models versus gradient boosting
- Forecast intervals
Stakeholders act on explanations, not predictions.
- Feature importance and its pitfalls
- SHAP values and partial dependence
- Model cards and documentation
- Fairness assessment
BuildA predictive model with a baseline, honest validation, and a written analysis of where it fails.
Experimentation
At product companies, A/B testing is the single most-used data science skill.
The most reliably asked topic in product data science interviews.
- Randomisation and assignment units
- Power analysis and sample size
- Primary, secondary and guardrail metrics
- Reading results without peeking
Knowing the failure modes is what separates senior candidates.
- Peeking and sequential testing
- Multiple comparisons
- Network effects and interference
- Novelty and primacy effects
Choosing what to measure is more consequential than how you analyse it.
- North star and input metrics
- Proxy metrics and gaming
- Sensitivity versus meaningfulness
- Long-term versus short-term effects
Half of real questions cannot be randomised. Quasi-experimental methods fill the gap.
- Difference-in-differences
- Synthetic control
- Regression discontinuity
- Stating assumptions explicitly
The output of the role is a decision, not a notebook.
- Decision memos and recommendations
- Communicating uncertainty to executives
- Cost of being wrong in each direction
- Following up on the decision made
BuildDesign, run and analyse an experiment, including power analysis and a written decision memo.
Interview preparation
Data science loops are wide: SQL, statistics, modelling, a case, and communication.
Almost always present, and usually the first filter.
- Window functions under time pressure
- Cohort and retention queries
- Funnel analysis in SQL
- Debugging a wrong result
Conceptual questions where precision of language matters.
- Explain a p-value to a product manager
- When would you not run an A/B test
- Identify the confounder in a scenario
- Sample size reasoning aloud
An open business question. Structure is what is being graded.
- Clarifying the actual question
- Choosing metrics and data sources
- Stating assumptions
- Recommending an action with caveats
Fewer algorithm derivations now, more judgement about approach.
Ch — LLM Fundamentals- Framing a business problem as an ML problem
- Choosing and defending a metric
- Handling leakage and imbalance
- Explaining a model to non-experts
One deep analysis of a real question beats five tutorial notebooks.
- A question that mattered, answered end to end
- Documented assumptions and limitations
- Clear visualisations
- Written recommendation
BuildTwo portfolio analyses that answer a real question, with clear write-ups and honest limitations.
Data Scientist tools on your CV
- Python
- pandas / Polars
- SQL
- scikit-learn
- statsmodels
- Jupyter
- dbt
- Tableau / Looker
What Data Scientist employers ask to see
- An end-to-end analysis with a stated recommendation
- An experiment design document with power analysis
- A model with honest evaluation and failure analysis
- A causal analysis with assumptions declared
AI, ML and data science postings grew 163% year on year. The role has split: pure modelling work moved toward ML engineering, while data science increasingly means experimentation, causal inference and decision support.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.