The four kinds of analytics

In depth

Descriptive analytics: getting the facts right

A large share of analytics work is descriptive, and for good reason: every other kind depends on it. Descriptive work turns raw records into summaries: totals (orders in August), averages (the average amount spent per order), rates (the share of visits that end in an order) and comparisons (this month against last month, or against a target). Module 3 of this course defines these measures exactly and teaches you to read them.

Descriptive analytics sounds easy and is easy to get wrong. A count of “orders” depends on what counts as an order: are cancelled orders in or out? A monthly comparison depends on whether the months are comparable: September has 30 days and August 31. A total depends on the data being complete and free of duplicates.

The next lesson deals with definitions and module 2 with data quality. For now, note that a common failure in analytics is not an exotic model but a wrong number in a simple table.

Diagnostic analytics: from pattern to explanation

Diagnostic work starts where description finds something surprising, and its main tools are simple. Drilling down splits a total into its parts: orders by channel, by device, by product, by day. Comparison sets the group that changed beside one that did not, such as the website, whose checkout was redesigned, beside the app, whose checkout was not. Timing asks whether a change began when a suspected cause began. The third lesson of this module, on problem-solving frameworks, gives this search a structure, so that you look everywhere the cause could be, and look once.

A diagnosis is a hypothesis until it has survived a test. Two things that move together need not be cause and effect: visits and orders might both fall in a heatwave, with neither causing the other. The last lesson of this course, and later B05 Correlation Analysis, treat this properly. Until you have ruled out the other explanations, say “the evidence is consistent with” rather than “the cause was”, and where you can, confirm with an experiment that changes one thing and watches what happens, the subject of B12.

Predictive analytics: estimating what you cannot yet see

A forecast is an estimate, made in advance, of a future value such as next month’s orders. Predictive analytics also estimates present values that are not yet known, such as whether a payment just made is fraudulent. Its methods run from the very simple to the very complex: the simplest forecast says next month will be like this month. A seasonal version says December will be like last December, which Lantern cannot use: its shop opened in March 2026, so it has no December to copy. Statistical models and machine learning fit patterns to many past observations at once, and B09 Forecasting and Time Series teaches them.

Every prediction assumes that the pattern it learned from will continue, and every prediction is uncertain. Good predictive work therefore says how uncertain it is, for example by giving a range, or a few scenarios with probabilities, rather than one number. It also checks itself afterwards. The forecast error is the actual value minus the forecast, once the actual is known. Tracking errors over time shows whether a method is biased (always too high, say) and how far it can be trusted.

Prescriptive analytics: choosing what to do

A prediction alone does not make a decision. Knowing that December’s gift-box orders will probably be somewhere between 160 and 320 does not say how many boxes to assemble; that also depends on what a box sold earns, what an unsold box costs and how much risk the business will accept. Prescriptive analytics brings these together. Its methods include optimisation, which searches for the choice that gives the best result within constraints (the cheapest delivery routes, the staff rota that covers every shift at the lowest cost); simulation, which plays out many possible futures to see how each option fares; and decision rules built on expected value, the probability-weighted average of the possible outcomes, which F00 introduced. B11 Optimisation and Decision Analytics develops these methods.

Prescriptive analytics does not take the decision away from people. It makes the trade-offs explicit: given these forecasts, these costs and this goal, this option does best. If Ana disagrees with a recommendation, the useful question is which input she disagrees with, and that is a far more productive argument than one about gut feeling.

Value and difficulty

The staircase makes one point and hides another. The point: moving from description towards prescription generally adds value, because the output gets closer to the decision, and adds difficulty, because each step needs the earlier ones plus more data, more assumptions and more skill. That is why so much work stays descriptive: it is cheaper, quicker and easier to trust.

What the staircase hides is that the four kinds are not grades of quality, and that you do not have to climb it. A correct monthly report that shows a manager something she did not know can be worth more than an elaborate model built on bad data. Some decisions need only description: whether the shop met a target is answered by a comparison. And the boundaries blur. A dashboard that turns a number red when it misses its target is already part-way to a decision rule, and a diagnosis often rests on an implicit prediction of what would have happened without the change.

Where machine learning fits

Machine learning is a family of methods in which a computer fits a model to data, learning a relationship between inputs and an outcome from examples rather than from rules a person writes down. In supervised learning the examples come with the answer attached, such as past customers labelled by whether they ordered again, and the model learns to predict that answer for new cases. In unsupervised learning there is no answer to predict, and the method looks for structure, such as groups of similar customers.

Machine learning is therefore not a fifth kind of analytics but a set of tools used inside the four, most often for prediction: scoring each customer’s chance of ordering again, forecasting demand, flagging a payment that looks fraudulent. Grouping customers into segments is descriptive work done with machine-learning methods, and some prescriptive systems learn which action works best by trying several. Whatever the method, the same discipline applies: a clear question, trustworthy data and an honest check of how well the model performs on cases it has not seen. B08 Machine Learning for Business teaches the methods; this course builds the habits they depend on.

The analytics process

Every analysis in this course, simple or complex, follows the same five steps, which together make up the analytics process.

1. Question. State the decision the analysis supports, and turn a vague worry into a measurable question: what exactly will be measured, for whom, over what period and compared with what. The next lesson teaches this.
2. Data. Find or collect the data the question needs, understand what each value means and check whether it can be trusted. Module 2 covers this, and B02 Data Cleaning and Preparation goes deeper.
3. Analysis. Describe, diagnose, predict or prescribe, as the question requires, using measures that are defined exactly. Module 3 covers the measures and how to read them.
4. Decision. Present the finding, with its uncertainty, to the person who decides, together with what you recommend and why.
5. Measure. After the decision, measure what happened and compare it with what was expected. The result answers the old question and usually raises the next one.

The last step is the one most often skipped, and it is what makes analytics improve over time. A forecast that is never compared with the actual cannot get better, and a change to a website whose effect is never measured teaches the business nothing.

Other sources draw the process with different boxes. OpenStax’s Principles of Data Science describes a five-stage data science cycle (problem definition, data collection, data preparation, data analysis and data reporting) and notes that collecting and preparing the data usually take the most time and effort. CRISP-DM, which you will meet in the third lesson of this module, uses six phases. They agree on the essentials: start from the question, earn trust in the data, and finish with something someone can act on.