Fact-check note: Reviewed September 4, 2026. Unsupported market and career forecasts and common statistical misconceptions were removed.

Data science combines domain knowledge, data stewardship, statistical reasoning, computation, and communication to answer defined questions with evidence. It is not synonymous with machine learning, and a complex model is not automatically more useful than a clear descriptive analysis or experiment.

1. Frame the problem

Define users, decision, target population, outcome, time horizon, baseline, alternatives, error costs, constraints, and accountable owner. Decide whether the task is description, prediction, causal inference, optimization, or exploration; each needs different evidence.

2. Understand the data

Document source, collection process, sampling, units, definitions, lineage, rights, consent, access, missingness, measurement error, selection, and changes over time. Preserve raw inputs and create versioned, reproducible transformations. Cleaning cannot prove that data represent reality.

3. Use statistical reasoning

Explore distributions and relationships while distinguishing planned analysis from discovery. A p-value does not give the probability that a hypothesis is true. Report effect estimates, uncertainty, assumptions, missing-data handling, multiplicity, and practical importance. Correlation and predictive accuracy do not establish intervention effects.

4. Build reproducible code

Python and R are both capable ecosystems; choose tools from team skills, libraries, deployment, governance, and maintainability rather than universal rankings. Pin environments, version code and data references, test transformations, validate schemas, log parameters, and protect secrets and sensitive outputs.

5. Validate against deployment

  1. Split by time, person, organization, location, or another real generalization boundary before preprocessing.
  2. Fit preprocessing only on training data and compare with simple baselines.
  3. Use cross-validation to estimate performance under its sampling assumptions; it does not confirm that a model is correct.
  4. Report uncertainty, subgroup results, calibration where relevant, error severity, latency, cost, and human-review burden.
  5. Keep a final test set and document every selection made using validation results.

6. Communicate limitations

Match tables and visualizations to the question, label units and denominators, show uncertainty, use accessible colors and text, and distinguish observed results from interpretation. State exclusions, residual risk, transportability limits, and what evidence would change the conclusion.

7. Monitor decisions, not just models

After deployment, monitor inputs, outcomes, drift, calibration, subgroup effects, overrides, incidents, cost, and user impact. Revalidate material changes and retain fallback, rollback, and retirement paths. A launched dashboard or model is not evidence of value.

Practice data preparation with data cleaning best practices, improve presentation through data visualization best practices, and manage experiments using MLOps best practices.