The ask rarely arrives with a definition attached. Somebody wants the numbers for Thursday, three people in the room mean three different things by "active customer", and whoever is holding the query editor has to settle it before the meeting starts. Learning this job is mostly learning to close that gap in public, without inventing certainty and without hiding behind a dashboard.
The work is deciding what counts
A data analyst turns a badly framed question into a number somebody can act on, then explains how the number was made. The software is a small part of that. Most of the job is definition, checking, and translation: deciding what counts as a customer, an order, a churn, a month, then defending the choice to people who will spend money or change a process because of it.
Take one ordinary phrase from a request and look at it as an object rather than a label. "Monthly active users" is a stack of decisions wearing a single name. Which month, calendar or billing. Active means what, one login, one meaningful action, one action that a bot cannot fake. A user is one account or one person, and those differ at every company that has ever offered a family plan. Each decision is defensible. The number changes by thirty percent depending on which ones you pick, and nobody in the meeting will know which set you used unless you tell them.
That is the hidden labor of the role. A junior analyst answers the question as written. A strong one notices that the question was a proxy for something else, usually a decision about budget, staffing, or whether a project continues. The useful move is to ask what will change depending on the answer. If nothing will change, the analysis is decoration and should be labeled as such.
Three audiences read the work. The requester wants an answer fast and will not read a methodology note. The decision meeting wants a sentence it can repeat afterward without embarrassing itself. The person who inherits the dashboard in eight months wants to know why the filter on transaction type exists. A deliverable that serves only the first audience becomes a number nobody can defend. A deliverable that serves only the third becomes a document nobody reads.
Trust accumulates in a specific way in this trade. An analyst is believed in proportion to how often their numbers survive a second look from someone who is trying to find the error. That is a slower currency than speed. It is also the only one that keeps paying after the first year, because the requests that matter go to whoever has already been checked and passed.
There is a boundary worth naming early, because beginners cross it by accident. Deciding what the company should do belongs to someone else. The analyst's job is to make the options legible, including the option of not acting. Analysts who start recommending before they can measure get promoted quickly and then stall, because the recommendation is only as good as the measurement underneath it.
What each part of the catalog is for
Four parts of the catalog carry this role, and each one fails in its own way when it is thin. A learner with SQL and no statistics produces precise nonsense. A learner with statistics and no SQL produces opinions about data they cannot retrieve. The four below are the ones this roadmap rests on.
Data Science & Analytics supplies the descriptive vocabulary: distributions, variation, sampling, correlation, the difference between a proportion and a count. The catalog holds 31 live resources in Data Science & Analytics, 26 of them free and 5 paid. This area is where you learn to look at a column before you summarize it, which prevents most of the errors that later look like technical bugs.
Databases & Backend is where the rows actually live and where the joins go wrong. The catalog holds 31 live resources in Databases & Backend, 27 of them free and 4 paid. The point is not to become a database administrator. The point is to know why a join multiplies rows, why a timestamp arrives in a different timezone than the one on the report, and why a table with no primary key will eventually embarrass you.
Business Strategy & Operations gives the numbers a reason to exist. Margins, unit economics, churn cost, the difference between a leading and a lagging indicator, the way a strategy becomes a set of measures that a team is judged on. In Business Strategy & Operations the catalog holds 30 live resources, 27 of them free and 3 paid. Analysts who skip this area can answer a question correctly and still be useless in the room, because they cannot say which number the business is actually steering by.
Machine Learning matters less than beginners expect and more than mid-career analysts assume. Machine Learning is the largest mapped area at 32 live resources, 25 of them free and 7 paid. You need enough of it to recognize when a prediction task is worth doing, when a regression answers the question better than a dashboard, and when the data cannot support the model someone has already promised a client.
The areas overlap in practice. A churn question needs the database area to build the cohort, the statistics area to compare cohorts fairly, the business area to know what retention is worth, and occasionally the machine learning area to rank who is likely to leave. The order below moves through them in the sequence that keeps a beginner honest.
An order that follows the request
Most curricula teach technique in the order textbooks present it. The order below follows the shape of an actual request instead: read what you were given, get the data yourself, describe it correctly, put it in front of a decision, then build a check that catches you next time. Each step has a done test you can apply alone.
1. Read the raw table before you trust a query
Open the extract and count things. How many rows, how many distinct customer identifiers, how many null dates, how many categories that look like the same category typed twice. Look at the smallest and largest values in every numeric column, because a negative order total or a date in 1970 explains more errors than any tutorial. This is the step beginners skip, and it is the step that separates an analyst who is trusted from one who is fast.
Kaggle Datasets (FREE) is a good place to practice on material that has real mess in it. DuckDB Documentation (FREE) is worth an hour even if you never use DuckDB at work, because a local analytical database lets you poke at a file without asking anyone for access. Our World in Data (FREE) is a useful counterexample to study: clean, documented, and explicit about what a series does and does not cover.
Done means you can describe a new table in five sentences, including the grain of the table, which is the thing each row represents.
2. Get SQL to the point where nobody has to write it for you
SQL is the one skill in this role that is not negotiable. You need selects, joins, aggregates, window functions, and enough of a mental model to know when a query is doing more work than it should. Learn it by writing queries against real tables rather than by reading syntax pages.
SQLBolt (FREE) is the shortest honest introduction, and its interactive exercises cover the clauses you will use daily. Mode SQL Tutorial (FREE) goes further into aggregation and window functions with a business framing that matches the job. SQLZoo (FREE) is a drill site and good for repetition. Select Star SQL (FREE) teaches the same material through a single messy dataset, which is closer to work than a clean tutorial table. DB Fiddle (FREE) lets you test a query in a browser without installing anything, and PostgreSQL Official Tutorial (FREE) is the reference to read once you have opinions about how a query should run.
If you need a structured course with a deadline, Dataquest (PAID) is a subscription built around writing code in a browser. Buy one month when you have a specific gap, not a year because the landing page was persuasive.
Done means you can answer ten business questions against an unfamiliar schema without help, including one that needs a self-join and one that needs a window function, and you can explain why a join changed your row count.
3. Learn the statistics that stop a wrong sentence
The statistics you need is narrower than a degree and more useful: variation, sample size, selection, the difference between a rate and a total, and the reason a percentage change on a small base is not news.
Khan Academy Statistics and Probability (FREE) covers the ground from descriptive statistics through hypothesis testing at a pace that fits evenings. OpenIntro Statistics (FREE) is a full open textbook, and its exercises are the part that makes the material stick. Statology (FREE) is a lookup site for the moment when you need to remember which test applies. Python Data Science Handbook (FREE) connects the same ideas to working code, which matters once you move past spreadsheets.
Done means that given a sentence like "conversion rose twelve percent", you can list four ways the sentence could be true and still useless, including a change in the denominator and a change in who was measured.
4. Learn to move data without breaking it
Most analyst work is reshaping: filtering, grouping, joining, pivoting, and producing a tidy table that someone else can use. Do this in code rather than by hand, even when the file is small, because the second request will always be a variation of the first.
pandas Getting Started Guide (FREE) is the official entry point and covers loading, selecting, and grouping. Kaggle Learn: Pandas (FREE) is a short hands-on course that pairs well with it. NumPy Learn (FREE) is worth reading for the array model underneath. freeCodeCamp Data Analysis with Python (FREE) is a full video course if you prefer a person walking through the workflow. R for Data Science (FREE) covers the same ground in R and is the better choice if your team already works in R.
Done means you can take a raw export, clean it reproducibly, and hand someone a script plus a table that a second person can regenerate without asking you a question.
5. Put the number in front of a decision
A chart is an argument with axes. Learn to make one that survives a projector, a phone screen, and a skeptical manager who has seen a misleading graph before. Learn too when a table is better than a chart, which is more often than tool vendors suggest.
Tableau Public (FREE) is a full visualization tool you can learn on public data, and publishing there is a portfolio by itself. Power BI Learning Catalog (FREE) covers the Microsoft stack that a large share of employers run. Matplotlib Tutorials (FREE) and Seaborn Tutorial (FREE) cover the code route, which is what you need when the chart has to be regenerated every Monday.
For the business half of this step, Strategyzer Free Knowledge Library (FREE) is useful for connecting a measure to a business model, and MIT Sloan Management Review (FREE) gives you the vocabulary of people who make the decisions you are supporting.
Done means one page that states the question, the answer, the definition used, the main caveat, and the decision the page enables.
6. Build the check that catches you later
Every analyst eventually ships a number that is wrong. The difference between a career and an incident is whether a check catches it before the meeting. Build the habit early: a row count comparison, a total that should reconcile, a duplicate check on the join key, a date range assertion.
Microsoft Data Science for Beginners (FREE) is a broad course that includes the workflow discipline around a result, and Awesome Data Science (FREE) is a curated list worth browsing when you need a technique you have not used before. Google Analytics Academy (FREE) is a free way to learn how a measurement system can be configured wrongly, which is a useful education in what your source data may already have lost.
Done means another person can reproduce your number from your notes and arrive at the same figure without calling you.
The six steps are not a calendar. They overlap, and the overlap is where the learning happens. But the sequence matters, because a learner who starts at visualization builds attractive charts on top of undefined metrics, and a learner who starts at machine learning builds models on data they have never counted.
Picks that earn a bookmark
Some entries do not belong to a single step but repay a place in your bookmarks.
For SQL fluency: PostgreSQL Exercises (FREE) is a set of graded problems against a real database, and PostgreSQL Documentation (FREE) is the reference you will eventually need for window frames and indexes. Khan Academy: Intro to SQL (FREE) is slower and gentler than the drill sites, useful if the first attempt at joins did not land.
For the shape of the job: Google Machine Learning Crash Course (FREE) is the fastest honest introduction to what a model does, and Elements of AI (FREE) covers the concepts without code. StatQuest with Josh Starmer (FREE) explains statistical methods visually, which is often what you need when a stakeholder asks what a p-value is.
For the business frame: FourWeekMBA (FREE) is a compact reference on business models, and MIT Sloan MBA First-Semester Core (MIT OpenCourseWare) (FREE) is a serious free course if you want the finance and strategy vocabulary that analysts are often assumed to already have. Harvard Business Review (PAID) is a subscription, worth it only if you are moving into a role where you are expected to argue strategy rather than measure it.
For practice material with real mess: State of AI Report (FREE) is a useful example of a document that presents numbers with sources and caveats, and Awesome Data Science (FREE) is the index to browse when you want a dataset for a specific technique.
Money, months, and the first number someone acts on
The free base here is unusually strong. Data Science & Analytics includes 26 free resources against 5 paid, and the entire six-step order can be completed without spending anything. That fact is why an analyst can be self-taught in a way a surgeon cannot.
Paid material is worth considering in three situations. The first is structure: a subscription such as DataCamp (PAID) or Dataquest (PAID) supplies a sequence and deadlines, which is exactly what a learner with a full-time job lacks. Buy one month, finish it, then decide. The second is a specific tool your employer uses and no free course covers. The third is a credential that a hiring process explicitly requires, which for this role is rarer than bootcamp advertising suggests.
Time is the honest constraint. At a steady part-time pace, expect three to six months before you can take a request from a stranger and return something defensible. The first month feels like syntax. The second month feels like you are getting worse, because you start noticing how many questions were ambiguous. The third month is when a colleague asks you a question instead of the tool.
What should exist by month four: three analyses you can walk someone through end to end, each with the question, the definition, the query, the chart, and one thing you got wrong and corrected. That last item does more for your credibility in an interview than any certificate, because it shows you know where errors come from.
What not to expect: that the first job arrives because of a course completion. Most first analyst roles come from a department you already work in, a small company that needs someone to make sense of its spreadsheet, or a referral from someone who has seen your work. The training exists so that when one of those openings appears, you can be handed the data without supervision.
Where an analysis quietly stops being true
Six failures account for most of the wrong numbers that reach meetings.
The first is joining before counting. A one-to-many join duplicates the left side, so revenue doubles and nobody notices because the chart still looks like a chart. Check row counts before and after every join, and know the grain of both tables.
The second is averaging an average. A conversion rate computed per store and then averaged across stores is not the company conversion rate unless every store had the same traffic. Weight by the thing that made the groups different.
The third is a filter that silently removes the population you were asked about. Dropping empty fields looks tidy and can delete exactly the customers who never activated. Count what you dropped and say so.
The fourth is a chart that flatters the finding. A truncated axis exaggerates a small change, and a dual axis can imply a relationship that does not exist in the data. The test is whether a reader who dislikes your conclusion would call the chart dishonest.
The fifth is answering the easy version of the question. "How many users churned" is easier than "which users churned and were worth keeping", and the second one is what the meeting needed. When you substitute, say that you substituted.
The sixth is shipping a number with no definition attached. The next person recomputes it with their own definition, gets a different figure, and now two teams are arguing about whose number is right when the real disagreement is about what the word meant. Write the definition in the same file as the number.
Each failure has a cheap check, and the checks compound. An analyst who counts rows, states dropped records, labels axes honestly, and writes definitions down will still make mistakes. They will just be mistakes that a colleague can find before a customer does.
What to open on Monday morning
Go deeper in the areas this page rests on: Data Science & Analytics, Databases & Backend, Business Strategy & Operations, and Machine Learning. Where one step above felt weak, the matching area page is the next stop rather than a new course.
The neighboring roadmaps are worth reading for contrast. Business Intelligence Analyst sits closer to reporting and governance, Data Engineer owns the pipelines your queries depend on, and Data Scientist carries the modeling work further. Compare the wider set at Browse by Profession.
If you want one concrete action rather than another tab, take the last report you built or received and write down the definition of every number on it. Where a definition is missing, that is your next week of study.
FAQ
I have a spreadsheet job and no SQL. Where do I start?
Start with SQLBolt (FREE) for two weeks, then move to Mode SQL Tutorial (FREE) for aggregation and window functions. Practice on an export from your current job if you are allowed to, because your own messy columns teach faster than a clean tutorial table. Keep the spreadsheet for one more month while you rebuild one recurring report in SQL, then compare the two outputs until they reconcile.
How do I answer a request that has no defined metric?
Ask what decision the number will support, then propose one definition and one alternative. Show how much the answer moves between them. If the requester cannot choose, pick the more conservative definition and label it on the page. Guessing what they meant is the wrong move. Make the choice visible so that a later disagreement is about the definition rather than about your competence.
Is a paid bootcamp ever the right buy for this role?
It can be, in one case: you need a structured sequence, a deadline, and feedback from a person, and you have already tried free material and stopped. A paid subscription such as DataCamp (PAID) is cheaper than a bootcamp and covers the same fundamentals. Bootcamps earn their price mainly through career placement, so ask for the placement numbers before paying, and compare them with the actual job listings in your city.
What should a portfolio contain when my work is confidential?
Rebuild the shape of the work on public data. Take a question you answered at work, replace the dataset with Kaggle Datasets (FREE) or a public source, and publish the question, definition, query, chart, and a paragraph on what you would check next. Three of these demonstrate more than a certificate, because a reviewer can see how you think when the numbers do not behave.
Can an analyst move across from another department?
Yes, and it is the most common route. Pick one recurring report in your current team, rebuild it in SQL and a reproducible script, and show the difference in time and accuracy. That artifact is your application. Analysts are hired on evidence of judgment with data, and judgment is visible in a rebuilt report that someone else can run.
How much statistics is enough before applying?
Enough to describe a distribution, explain sampling and selection, read a confidence interval, and refuse a causal claim from observational data. OpenIntro Statistics (FREE) covers that ground in its first six chapters. You do not need to derive estimators. You do need to notice when a comparison is between groups that were never comparable.
SkillCache Editors · Updated November 13, 2026
