Technology
Finds patterns in complex information to guide decisions
Every organisation collects data: what customers buy, where deliveries are delayed or how people use a product. A data scientist turns that raw information into a decision. They use statistics, code and knowledge of the organisation. Their work explains what happened, predicts what might happen next and tests whether an idea actually works.
The fascination - and frequent annoyance - is that data never arrives with a trustworthy explanation of itself. A column called “delivery date” may mean the promised date in one system and the actual date in another. A mathematical model can be sound while learning from a process that has just changed. Data scientists must question how the numbers were created as seriously as they adjust the method. They then need an answer useful enough for somebody to act on.
Most graduates enter through a junior data scientist or data analyst role. A degree involving numbers helps. Employers also care whether you can analyse data with code, work with databases and explain a project clearly. A master’s can help graduates changing subject or aiming at research-heavy machine learning. It is not required for every entry-level job.
| When | Lowest 10% earn under | A typical earner | Top 10% earn over |
|---|---|---|---|
| On graduating | £25,000 | £33,000 | £46,000 |
| 1 year in | £27,000 | £36,000 | £52,000 |
| 2 years in | £30,000 | £42,000 | £60,000 |
| 3 years in | £33,000 | £48,000 | £68,000 |
| 4 years in | £36,000 | £54,000 | £78,000 |
| 5 years in | £38,000 | £60,000 | £90,000 |
| 6 years in | £41,000 | £65,000 | £100,000 |
| 7 years in | £44,000 | £70,000 | £108,000 |
| 8 years in | £46,000 | £74,000 | £114,000 |
| 9 years in | £48,000 | £78,000 | £120,000 |
| 10 years in | £50,000 | £82,000 | £125,000 |
Most workplaces are full of persuasive stories about why sales rose, customers left or a new policy succeeded. Data science gives you tools for testing those stories. You learn to separate coincidence from evidence, design fair comparisons and say how sure you actually are. That way of thinking is useful in almost any important decision, not just at work.
The same basic skills can detect fraud at a bank, improve hospital waiting times or forecast energy demand. They can also help explain why people stop using an app. You still need to learn each industry properly, but you are not locked into one sector. If the subject stops interesting you, the craft can move with you.
Statistics, programming, experiments and machine learning all reward depth. A problem that takes two days as a graduate may take an hour after several years. You can feel yourself becoming more precise. Data science offers a satisfying middle ground for people who like variety but still want a hard skill to practise.
You may imagine spending the week training intelligent models. In reality, data is often missing, duplicated, badly labelled or spread across systems that were never designed to work together. A large part of the job is finding the right table, cleaning it and checking that the result has not quietly become nonsense. If you dislike painstaking detail, the interesting ten per cent will not compensate for the rest.
Evidence does not make a decision by itself. A recommendation may be inconvenient or arrive after somebody senior has chosen a direction. You must understand what colleagues care about and explain the result without hiding behind technical language. Sometimes the organisation will still ignore you. Being right is only half the job. Making the answer usable is the other half.
Companies use data scientist, analyst, machine-learning engineer and applied scientist to mean different things. Some supposedly junior roles ask for a master’s, years of experience and a long list of tools. Others use an impressive title for routine reporting. Graduates need to judge the actual work carefully. Starting as a data analyst may be the quickest route into the profession, not a consolation prize.
Usually 3–4 years
Computer science, statistics, mathematics, physics and quantitative economics all provide useful foundations. A data-science master’s can help when changing subject, but is not a routine entry requirement. Build one project around untidy, real-world data while you study. An employer learns more from the choices you made while cleaning and testing it than from a polished chart or the precise name of your degree.
Usually 0–2 years’ experience
You’ll normally work on well-defined questions with help from a more experienced colleague. Much of your time goes into learning where the organisation’s data lives, writing reliable queries and checking your work. The aim is to prove that someone can make a decision from your analysis without being misled by it.
Usually 2–5 years’ experience
You’re expected to take a problem from a vague question to a useful answer yourself. You might design an experiment, build a forecast or prediction model and explain what the organisation should do next. You also start choosing the method rather than merely carrying out an approach somebody else selected.
Usually 5–8 years’ experience
The questions become more ambiguous and more important. You’ll guide other data scientists, challenge weak assumptions and decide how much analytical effort a problem deserves. Depending on the company, you may specialise in experimentation, forecasting, machine learning or a particular industry.
Usually 8–12 years’ experience
This is often the senior individual-contributor route. You set technical standards, shape large projects and become the person others consult when the obvious analysis is not good enough. Some principal data scientists stay deeply technical; leads may also coordinate a small team while still doing substantial hands-on work.
Usually 12+ years’ experience
You’re now responsible for a team’s direction as much as its analysis. You decide which problems deserve data-science investment, hire and develop people and make sure the work is trusted and used across the organisation. The job shifts towards leadership, budgets and company strategy, although technical credibility still matters.
Combines programming, databases and machine-learning foundations, which makes it the largest single subject group. Its technical breadth may leave less room for experimental design and for explaining uncertainty without specialist shorthand.
Teaches you to reason from evidence and understand what a model is actually assuming. Graduates often have the strongest theoretical base, then need to turn classroom coding into software that other people can reliably use.
Trains you to model complicated systems and solve problems within real constraints. The combination of quantitative confidence and practical judgement transfers well, particularly in manufacturing, energy and technology.
Physics, chemistry and biology all involve noisy measurements, experiments and conclusions that must survive scrutiny. Scientists usually arrive comfortable with evidence; the main gap is often production-quality programming or commercial context.
Economics brings statistics, causal reasoning and an instinct for incentives. Psychology and other quantitative social sciences bring experimental design and an understanding of human behaviour, particularly useful in product and policy work.
Business, geography, the arts and many other subjects form a small but useful remainder. Geography can sharpen spatial analysis; linguistics can help with language data; business subjects clarify the decision behind a model. These graduates generally need much stronger statistics and programming, but domain knowledge can reveal when technically tidy data misrepresents the thing being measured.
Hover a subject to see why it helpsTap a subject to see why it helps
08:45 – 09:30Check the overnight numbers
You look at the company’s main measures and a model that predicts which customers may leave. Its predictions suddenly look less accurate. Nothing is obviously broken, so you write down three possible explanations before opening the code. Data science is easier when you resist falling in love with your first theory.
09:30 – 10:00Team stand-up
The data team quickly shares what it is doing and where it is stuck. An analyst needs help defining a customer as “active,” while an engineer warns that a useful data source will change next week. You agree who will investigate each issue. Even apparently mathematical work depends on people using the same definitions.
10:00 – 11:30Follow the trail
You examine the raw data and discover that a recent app update changed when one event is recorded. The model has not suddenly become stupid; part of its input has changed meaning. You compare old and new records and speak to the engineer who released the update. Together, you create a clean version that can be used safely.
11:30 – 12:30Experiment review
A product manager wants to know whether a new sign-up page performed better than the old one. The headline conversion rate increased, but more of the new users came from an unusually successful marketing campaign. You work through the comparison and explain what can be concluded now, what cannot and how to design a cleaner follow-up test.
12:30 – 13:15Lunch
Last night’s leftover pasta.
13:15 – 14:45Build the model
You return to a project forecasting weekly demand. You try a more complex method, test it against unseen data and find that it performs barely better than the simple baseline. You keep the simpler model. Good data science is not about using the cleverest technique; it is about using the least complex one that solves the problem reliably.
14:45 – 15:30Explain the result
You show the forecast to a product manager. They do not want a tour of the mathematics. They want to know how many staff to schedule and what could make the estimate wrong. Their questions reveal a practical limit you had missed, so you adjust the recommendation. Knowledge of the work often improves a model more than another afternoon of adjustment.
15:30 – 16:45Make it repeatable
The analysis currently works only in your notebook. You tidy the code, add checks for missing data and ask a colleague to review it. This is the difference between an interesting answer and something the organisation can trust every week.
16:45 – 17:45Let operations challenge it
You show the corrected results to the delivery manager. They notice that one depot looks excellent only because failed orders are reassigned before the daily count. Together, you change the comparison. You agree to let the model make predictions without controlling decisions for another week. Nothing launches today, but the new test can reveal whether the model helps in real work.
+3 hours compared with the average UK job