Click here to start the web book.
I built Machine Learning A to Z because most ML courses teach the calling of libraries, not the craft of machine learning. This is the course I wanted when I was learning: every algorithm taught the way a professional actually needs it — definition, intuition, mathematics, assumptions, a worked example computed by hand, a real Python implementation, and an honest evaluation of the result. It is a complete, self-contained web book: 10 parts, 33 chapters, 29 algorithms, 1,000 quiz questions, and a growing track of applied equity-research projects — all running in the browser from a static site, dark-mode first.
📖 Read the full book
Click here to start the web book.
No sign-up, no install. Start at Chapter 1 and read in order, or jump to any algorithm.What makes it different
- Theory before code, always. Each chapter follows the same anatomy: what the method is, the intuition, the full mathematics with every symbol annotated, the assumptions and what breaks without them — and only then the implementation.
- Every example is worked by hand first. OLS by normal equations, a logistic posterior from a 9.2:1 odds ratio, entropy and information gain on a real split, backpropagation on a 2-2-1 network with exact gradients, Fisher’s discriminant on eight points, a convolution slid across a 5×5 grid, boosting rounds on four data points. If you can’t compute a small case by hand, you don’t understand it.
- Every number is real. Nothing on any page is invented. Each chapter’s dataset is actually loaded and run (scikit-learn, TensorFlow/Keras, statsmodels, XGBoost, NLTK), and the printed confusion matrices, learning curves and benchmark tables are the genuine outputs — including the unflattering ones.
- Honest evaluation is a theme, not a footnote. A dedicated chapter proves that the same model scores anywhere from 88% to 95% across ten random splits — then teaches cross-validation, grid search and the discipline of touching the test set exactly once. Simple baselines are allowed to win when they win.
- It ends like the job ends. The capstone runs one full professional pipeline — raw data to deployed model — and prices the business decision ($200 retention offers, $2,000 customers, a threshold sweep worth $610,800). Applied project lessons then carry the methods into equity research.
The course map
Part I — Foundations (Ch 1–3)
Machine-learning literacy and the train/validation/test contract; a statistics primer (distributions, sampling, correlation, the inference mindset); and the complete preprocessing toolbox — missing data, encoding, scaling, splitting, and the data-leakage rules that everything later depends on.
Part II — Regression (Ch 4–10)
Simple and multiple linear regression with the full assumption suite (linearity, multicollinearity via VIF, residual diagnostics); polynomial regression and the bias–variance trade-off; support vector, decision-tree and random-forest regression; closed by a dedicated evaluation chapter where six regressors compete on one benchmark — and the simple model wins.
Part III — Classification (Ch 11–18)
Logistic regression from log-odds to thresholds; K-nearest neighbours and the curse of dimensionality; SVMs and the kernel trick; Naive Bayes with a hand-computed posterior; tree and forest classifiers — then a metrics chapter that owns the hard questions: precision vs recall, ROC-AUC, class imbalance, cost-based thresholds and calibration, with an eight-model scoreboard on shared data.
Part IV — Clustering (Ch 19–20)
K-Means with elbow and silhouette on real mall-customer segmentation (five business-ready segments); hierarchical clustering and the dendrogram that reproduces them.
Part V — Association Rules (Ch 21–22)
Apriori and Eclat: support, confidence and lift derived numerically from real market baskets — including why lift, not confidence, finds the actionable rules.
Part VI — Reinforcement Learning (Ch 23–24)
The multi-armed bandit: exploration vs exploitation formalised, Upper Confidence Bound, and Thompson Sampling with Beta posteriors — the mathematics of “when should I stop gathering information?”
Part VII — Natural Language Processing (Ch 25)
Text as data: cleaning and stemming by hand, the bag-of-words matrix, TF-IDF with verified arithmetic, and Naive Bayes sentiment classification on 1,000 real reviews — plus an honest look at what transformers changed.
Part VIII — Deep Learning (Ch 26–27)
Neural networks from the single neuron (it is logistic regression) through backpropagation computed entirely by hand, to a real churn network on 10,000 customers; then CNNs — convolution and pooling by exact arithmetic, a real 813,217-parameter cat-dog classifier, and its genuinely overfit learning curve as the teaching moment.
Part IX — Dimensionality Reduction (Ch 28–30)
PCA with the eigendecomposition computed by hand (discarded eigenvalues are the compression error — proven); LDA as the supervised twin (Fisher’s criterion by hand, a perfect 36/36 on the wines); Kernel PCA bending the axes where the rings no linear method can separate — with the gamma dial turned until they snap apart.
Part X — Model Selection, Boosting & the Capstone (Ch 31–33)
k-fold cross-validation fold by fold (the same SVM swings 88–95% across splits; the fix is averaging 271/300 honest predictions); grid search and its six-way-tie lesson; XGBoost from residuals up (hand-boosted MSE 2.5 → 0.29, early stopping at round 107 of 1,000); and the capstone — one end-to-end pipeline from raw bank data to a deployed model with a money-denominated business case.
Applied Projects — ML in Equity Research (new track)
Full project lessons that weld the chapters into professional workflows on real market data. Project 1 (live): a decision tree that classifies stock momentum — every technical indicator taught as finance (RSI computed by hand, MACD, golden crosses, ATR), three genuine bugs found and fixed in the original template, and an honest 46%-vs-34% scoreboard that becomes a lesson in market efficiency. Random Forest and SVR projects are in production.
By the numbers
| Parts / Chapters | 10 / 33 (+ applied projects) |
| Algorithms implemented | 29 |
| Quiz questions | 1,000 (30 per chapter, answer keys verified) |
| Hand-worked examples | in every single chapter |
| Real datasets | salaries, startups, cancers, wines, social ads, mall customers, market baskets, restaurant reviews, bank churn, cat–dog photos, 8-stock market snapshot… |
| Stack | 100% static HTML/CSS/JS — dark-mode first, works offline, no tracking |
Who it’s for
Students who want the why behind the API call; analysts and engineers moving from tutorials to practice; and instructors who want a reference where every claim can be checked against a run. Prerequisites: comfort with basic Python and high-school algebra — the statistics you need is Chapter 2.
The full book — all 33 chapters, the capstone and the applied projects — is free to read here: Click here to start the web book. . Start with Chapter 1, or skip straight to your algorithm.