Stop 0 · Foundation
Pack your bags
Pick your question, load a spreadsheet, and set a baseline to beat. Everything starts here.
Machine learning is simpler than everyone tells you — simple enough to guess the very day a tycoon duck cracks open his vault to buy. Here's how it works… and the Monday it actually called it, five days early and blind.
Strip away the jargon and every model on Earth does one humble thing: it studies past examples, spots a pattern, and guesses the next answer. That's it. You already do this every day — you just don't call it a "model."
Guess whether a café will be busy. You picture past visits — rainy Tuesday, sunny Saturday — spot the pattern, and predict. A machine does the same, only it can weigh thousands of past examples at once and never gets tired.
1. Gather rows of past examples. 2. Point at the answer you wish you could predict. 3. Let the model find the pattern. 4. Test it honestly on examples it never saw. No calculus required.
The computer does the arithmetic. You choose the question and read the answer.
It all runs free in your browser through Google Colab. Nothing to break.
Seven models, each a single plain-English trick. Learn one, add it, repeat.
Each lesson is one simple idea for turning clues into a guess. Two of them — the nearest-neighbour trick and patient correction — are the very models that made the real prediction below.
Pick your question, load a spreadsheet, and set a baseline to beat. Everything starts here.
To guess a new duck's move, find the handful most similar and copy the crowd.
Score every clue, add them up, out pops a number. Simple as a recipe.
Sorts each duck into a category and says how sure it is. "80%" is a lean, not a promise.
A chain of simple questions leads to the answer. Stop it before it starts memorising.
A hundred little trees voting together beat any single clever one.
Small trees, each fixing the last one's mistakes. This is the model that made the prediction.
No answer column? Find the natural groups hiding among your rows.
There's no fixed order. Do the foundation once, then wander the trail at your own pace — nothing you can break.
In the real market, thousands of moneyed "houses" trade India's biggest blocks of shares. We set the model a bold, testable dare — no hindsight, no do-overs:
The market closed. Every call got graded. The verdict, in six numbers:
Illustrative names. The real study ranked 584 actual NSE trading houses; the true top five all printed a buy.
Summed across the shortlist, the model's forecast of the day's total buying landed within a whisker of reality.
A prediction is only honest if you say where it breaks. The model dazzled at who will trade — but that's a narrower magic than it looks.
The houses it nailed are money-changers: they buy and sell the same stock in the same breath. It predicts who prints — not who's quietly getting richer. Net position ≈ zero.
Three high-probability regulars simply sat Monday out. A house that trades most days but skips this one is the hardest call of all.
Roughly one in three of any day's buyers are first-timers the model has never met. You can't rank a duck that isn't in your book.
A brilliant "who opens their vault today" engine — not a treasure map to riches. It came true exactly where it claimed skill, and stayed humble about the rest.
The exact same models that called Monday are waiting for you in a free, gentle, no-nonsense course. Bring any spreadsheet you're curious about — sales, students, sports, your own budget — and predict something.
Do the foundation once, then add each new model underneath it. If Colab forgets its place, just Run all — nothing is ever lost.
Your numbers depend on your data, so they won't match anyone else's. The only fair rival is your own baseline.
Tune on practice data; peek at the set-aside test set a single time, at the very end. That keeps the score honest.