The team defines the decision the app must make before moving money: given the customer's situation today, how likely are they to stay covered until payday if the app transfers a small amount into savings now?
By the end of this path you will build machine learning models yourself and understand every piece, because you will have built each piece first.
You hand-build a tiny prediction engine for a real app, with nothing but counting.
You watch a model learn the same kind of engine automatically, on a brand-new problem.
You came here to build machine learning models, and the first courses in this path are probability and statistics. It is natural to wonder how the three connect. So instead of explaining that connection in the abstract, we will show it through one small team building one real app, stage by stage.
The app is simple to understand. It automatically moves small amounts from a customer’s account into savings, but only when the customer can spare it. The goal is to help savings grow without leaving the customer short before payday.
That sounds like a product decision, but underneath it is an ML problem. Before every transfer, the app has to ask: given this customer’s balance, bills, spending pattern, and days to payday, how likely are they to stay covered if we move money now?
Getting that answer right is not a single trick but a whole lifecycle: understanding the business decision, studying account history, preparing the data, building the model, testing it, and finally using it in the live app.
Step through the six phases and see the probability and statistics under each one.
The team defines the decision the app must make before moving money: given the customer's situation today, how likely are they to stay covered until payday if the app transfers a small amount into savings now?
This ML decision is a conditional probability problem underneath. The app is not asking whether customers stay covered in general. It is asking how likely this customer is to stay covered given today's balance, bills still due, spending pattern, and days to payday.
This ML decision is a conditional probability problem underneath. The app is not asking whether customers stay covered in general. It is asking how likely this customer is to stay covered given today's balance, bills still due, spending pattern, and days to payday.
Statistics does not drive this stage yet. At this stage, the team is defining what a correct decision means, so the later data, model, and evaluation all measure the same outcome: did the customer stay covered or not?
Three parts, nine modules, each taught through its own case study. Tap any module to see what it gives you.
Measure chance and update it with evidence: the language every model speaks.
Prerequisite: School math only, nothing else
Measure chance from counts, and update it as new evidence arrives.
Why it’s on the path: Every classifier you will ever build answers with a probability; this is where that idea gets built by hand.
Case study: a job-posting watcher app
Meet the shapes data takes, and learn what is typical versus rare.
Why it’s on the path: Real data arrives as distributions; you cannot read a dataset, or a model’s assumptions, without knowing their shapes.
Case study: a social media content team
The modules build on each other, 1 through 9. Already know a topic? Skip ahead. Your call.
No deadlines, no live sessions. Pick up exactly where you left off.
The first lessons of every module are free. No account needed to read them.
Read data honestly: summarize it, trust it, and estimate from it.
Prerequisite: Part I
Summarize any dataset honestly in a handful of numbers.
Why it’s on the path: Before any model, a team summarizes what the data says; these are the numbers every dataset is read through.
Tell a real improvement from plain luck.
Why it’s on the path: Every “is it actually better?” question in ML, about a model or a change, is answered with these tools.
Case study: an e-commerce A/B test
Find the values that best explain the data you actually saw.
Why it’s on the path: This is what “training” mathematically means; every fit in Course 3 runs on this idea.
Case study: predicting ad clicks
Put it all together: explore real data and build your first models.
Prerequisite: Parts I and II
Get the lay of the land: what machine learning is, its vocabulary, and how real teams work.
Why it’s on the path: The map of the field; without it, the models that follow are recipes without context.
Take a real, messy dataset and make it ready for modeling.
Why it’s on the path: Models trained on unexamined data learn the mess; this is where you learn to find it first.
Build your first model: predicting a number.
Why it’s on the path: Your first complete build: the simplest model that predicts a number, end to end.
Case study: predicting insurance premiums
Build a classifier whose output is a probability, and see Module 1’s engine learned automatically.
Why it’s on the path: Your first classifier, where probability (Module 1), statistics, and training (Module 5) converge.
Case study: predicting diabetes