How Data Science Works: From Questions to Decisions

What data science is—and what it is not

Data science is a way to learn from data using methods from statistics, computer science, and a field such as transportation, retail, biology, or public policy. A data scientist may describe what has happened, investigate why a pattern appears, predict what could happen next, or help decide what action to take. The right approach depends on the question and the information available.

It is not simply making charts, writing code, or applying artificial intelligence. Those can be parts of the work, but useful analysis also requires clear definitions, reliable data, appropriate methods, and careful interpretation. A sophisticated model cannot fix a poorly framed question or information that does not represent the situation being studied.

How a data science project works

Projects often move through these stages, though teams may revisit earlier steps when they find a problem or learn something new:

Define the question. Turn a broad concern into something answerable. “How can deliveries be more reliable?” might become “Which factors are associated with late deliveries in this service area?” Agree on what counts as “late” and what decision the analysis should inform.
Find and understand the data. Identify relevant sources, how the information was gathered, what each field means, and whether access is appropriate. Data may come from transaction records, sensors, surveys, public datasets, or operational systems.
Prepare the data. Inspect for missing values, duplicates, inconsistent units, incorrect entries, and mismatched records. Decide how to handle each issue and document choices. This step can take substantial effort because real-world data is rarely analysis-ready.
Explore and analyze. Summaries and visualizations help reveal distributions, unusual cases, and possible relationships. Statistical methods can test whether an observed pattern is consistent with the question, while domain knowledge helps distinguish a meaningful signal from a misleading coincidence.
Build and evaluate a model, if needed. A model is a simplified representation used for tasks such as forecasting or classification. Set a suitable baseline, train using one portion of the data, and evaluate on separate data that was not used to fit the model. Choose measures that reflect the real consequences of errors.
Explain and use the result. Share findings, limitations, and uncertainty in terms the intended audience can act on. If a model is put into routine use, monitor its performance and the data it receives; conditions can change and results can degrade.

Not every project needs a predictive model. If a careful summary answers the question, adding a complicated algorithm may make the result harder to explain without making it more useful.

Data collection and preparation

Data can be structured, like rows in a spreadsheet, or less structured, like text, images, and audio. Before analysis, a team needs to know how records were produced and what they leave out. For example, a delivery dataset may include completed trips but omit canceled orders. An analysis of only completed trips cannot directly describe delays among all attempted deliveries.

Cleaning does not mean making data look neat at any cost. Missing information may itself be meaningful, and removing unusual values can erase important events. Analysts should check whether records are comparable, choose transparent ways to handle problems, and preserve a record of changes. They should also consider whether the data can be used for the intended purpose and whether sensitive information needs safeguards.

Statistics, machine learning, and interpretation

Statistics helps summarize uncertainty and assess patterns. Machine learning uses algorithms to find patterns that can support predictions or classifications. These approaches overlap, but neither automatically explains why a pattern occurs. A model that predicts late deliveries from distance, for instance, does not prove that changing distance would prevent lateness; other factors may be involved.

Evaluation should resemble the conditions in which a result will be used. Randomly splitting records can be inappropriate when time order matters, because future information might leak into training. For a forecasting task, testing on later observations is often more realistic. For a classifier, accuracy alone can hide poor performance on an uncommon but important category; false positives and false negatives may have different consequences.

How to judge whether the result is useful

A useful result is not necessarily the most technically elaborate one. Ask whether the original question was answered, whether the data reflects the people or events of interest, and whether the method was evaluated fairly. A clear baseline—such as predicting that tomorrow will resemble a simple historical average—helps show whether a more complex model adds value.

Check the target: Is the outcome defined in a way that matches the decision?
Check the evidence: Are important groups, periods, or cases missing?
Check the errors: Which mistakes matter most, and who could be affected?
Check the explanation: Can the intended audience understand the limits and uncertainty?
Check after deployment: Does performance remain acceptable as data and conditions change?

Data science can support a decision, but it does not make the decision automatically. People still have to weigh goals, constraints, and consequences that may not be captured in a dataset.

Common problems and misconceptions

A frequent misconception is that more data always produces better answers. More records can help, but irrelevant, biased, outdated, or poorly measured data can reinforce the wrong conclusion. Another is that a strong association proves cause and effect. A relationship may arise because of a third factor, selection effects, or the way data was collected.

Overfitting is another risk: a model can learn quirks of its training data instead of patterns that hold in new cases. Testing on data kept separate from training helps reveal this, but it does not guarantee that future conditions will match the test set. Teams should also guard against data leakage, where information unavailable at prediction time accidentally enters model development.

Finally, a technically sound result can still fail in practice if it is hard to use, poorly explained, or disconnected from a real decision. Good communication is part of the method, not an optional finishing touch.

Frequently Asked Questions

What does a data scientist do?

A data scientist frames questions, examines and prepares data, applies statistical or computational methods, evaluates results, and explains what they mean. The balance of coding, analysis, communication, and subject knowledge varies by project.

Is data science the same as artificial intelligence?

No. Data science is a broader practice of learning from data. Artificial intelligence and machine learning can be tools within it, but many data science projects use descriptive statistics or simpler analyses instead.

Does data science always use machine learning?

No. A well-designed summary, comparison, or statistical analysis may answer the question more clearly. Machine learning is useful when its predictive or pattern-finding abilities suit the task and can be evaluated responsibly.

Why does cleaning data take so much work?

Information is often collected for operational purposes rather than analysis, so fields may be missing, inconsistent, duplicated, or defined differently across sources. Resolving these issues carefully helps prevent misleading conclusions.

Can data science prove that one thing caused another?

Observational patterns alone generally do not establish causation. Stronger causal conclusions require an appropriate design and assumptions, such as a well-conducted experiment or a suitable causal analysis.

Further reading: National Institute of Standards and Technology. Details can change, so check the official page for anything that matters to you. Spotted a mistake? Tell us. How we write an answer.