Statistical Learning — see every algorithm move

Eighteen interactive modules covering the modern statistical learning canon — the topics of the classic graduate curriculum (Hastie, Tibshirani & Friedman's The Elements of Statistical Learning), rebuilt here as plain-English lessons with a live visual lab in every module. Drag a bandwidth and watch bias appear; add boosting rounds one at a time; pile features onto a fixed dataset until least squares breaks. One thread ties all eighteen together: every method is a different position on the same bias–variance dial — and the honest way to set that dial is always held-out data.

🔎 Brand new? Take the intro track first

10 gentler modules — the same story with no math assumed.

🚀 Practitioner? Jump to boosting

Watch the XGBoost idea assemble a curve one stump at a time.

Foundations

The vocabulary and the honest scoreboard: overfitting, bias-variance, cross-validation, the bootstrap.

  1. 🎯
    Module 1Available

    What is statistical learning?

    Learning a rule from data, and the central tension: flexibility captures signal — and then starts memorizing noise.

    🎛 Overfitting playground

  2. ⚖️
    Module 2Available

    Two poles of prediction

    Rigid linear rules vs memorize-your-neighbors, the bias-variance decomposition, and the curse of dimensionality.

    🎛 kNN vs linear lab

  3. 🧪
    Module 7Available

    Model assessment & selection

    Training error lies. K-fold cross-validation, the one-standard-error rule, and the leakage traps that fake accuracy.

    🎛 Cross-validation lab

  4. 🎲
    Module 8Available

    The bootstrap & averaging

    Resample your own data to measure any estimator's uncertainty — and the short road from bootstrap to bagging.

    🎛 Bootstrap lab

Linear models & smooth functions

The linear machine and three ways to make it flexible: shrinkage, basis expansions, local weighting.

  1. 📏
    Module 3Available

    Regression, ridge & lasso

    Why unbiased least squares fails with correlated features, and how L2 shrinks while L1 shrinks and selects.

    🎛 Coefficient-path lab

  2. 🪓
    Module 4Available

    LDA vs logistic regression

    Model the class clouds or model the boundary — same straight line, different philosophies, different failure modes.

    🎛 Boundary lab

  3. 🧵
    Module 5Available

    Basis expansions & splines

    Curves from linear machinery: why global polynomials thrash at the edges and natural splines don't.

    🎛 Spline lab

  4. 🔍
    Module 6Available

    Kernel smoothing

    Prediction as a locally weighted average: the bandwidth dial, and the boundary bias local regression fixes.

    🎛 Smoother lab

Trees, boosting & modern learners

From one interpretable tree to the ensembles and networks that dominate applied prediction.

  1. 🌳
    Module 9Available

    Decision trees

    Greedy recursive partitioning: interpretable, interaction-friendly, piecewise-constant — and unstable on purpose to fix later.

    🎛 Regression-tree lab

  2. 🚀
    Module 10Available

    Gradient boosting

    Weak stumps fitted to residuals, added a sliver at a time — the engine inside XGBoost, and why slow learning wins.

    🎛 Boosting lab

  3. 🧠
    Module 11Available

    Neural networks

    Hidden units as basis functions that position themselves, weight decay, and the bridge to deep learning.

    🎛 Hidden-layer lab

  4. 🛣️
    Module 12Available

    Support vector machines

    The widest street through the data, the few points that pave it, and the kernel trick that bends it.

    🎛 Margin lab

  5. 🌲
    Module 15Available

    Random forests

    Bagged, feature-decorrelated deep trees: variance averaged away, OOB error for free, and why more trees never overfit.

    🎛 Forest lab

  6. 🤝
    Module 16Available

    Ensembles & stacking

    Blend models that fail differently, and let stacking learn the weights — on held-out predictions, or else.

    🎛 Blending lab

Neighbors, structure & scale

Memory-based methods, learning without labels, dependency graphs, and the p ≫ N endgame.

  1. 📍
    Module 13Available

    Nearest neighbors & prototypes

    The training set is the model: live decision maps, choosing k, and compressing the memory into prototypes.

    🎛 Decision-region lab

  2. 🧭
    Module 14Available

    Unsupervised learning

    k-means stepped iteration by iteration, PCA's variance-hunting axis, and why validating clusters is the hard part.

    🎛 k-means + PCA lab

  3. 🕸️
    Module 17Available

    Undirected graphical models

    Correlation vs partial correlation: zeros in the precision matrix are missing edges, and the graphical lasso finds them.

    🎛 Dependency-graph lab

  4. 🌌
    Module 18Available

    High-dimensional problems

    p ≫ N: least squares interpolates noise with certainty, regularization becomes mandatory, and lucky backtests multiply.

    🎛 p ≫ N lab

Why this matters for traders

Every quantitative trading question is a statistical learning question wearing a costume: signal research is feature selection under p ≫ N (Module 18), backtest evaluation is Module 7's honest-scoring problem, portfolio dependence is Module 17's graph, and regime detection is Module 14 without labels. The  fixed incometrack's credit-ML module applies exactly these tools to a live trading book.

Original interactive lessons and examples; all labs run synthetic data in your browser. Educational analysis, not investment advice.