Guided Learning Path
From intuition to working experiments
Use this path when you want a clear route through the site: understand the pattern, build a small model, evaluate it honestly, then test assumptions in a browser lab or sandbox.
Beginner
ML Foundations
~60 min · no prior ML required
Step 1
Know Your Data
Distribution shape, outliers, and why summaries can hide failure modes.
Step 2
How Models Learn
Gradient descent as a practical training loop: loss, updates, stopping, and failure.
Step 3
Classification Metrics
Precision, recall, false positives, and why accuracy is often the wrong target.
Step 4
Honest Evaluation
Cross-validation, leakage checks, and the difference between tuning and testing.
Build
First Real-Data Workflow
~75 min · includes local CSVs
Case 1
California Housing
20,640 real districts: why random cross-validation flatters a model, and a price cap hidden in the target.
Step 2
Bias vs Variance
Diagnose underfitting, overfitting, and whether more data will help.
Step 3
Regularization
Control model complexity without blindly tuning the test set.
Run
JupyterLite Lab
The housing case as a notebook, in Jupyter in the browser — or run each step on the case page.
Intermediate
Time Series and Drift
~90 min · production-minded validation
Step 1
Stationarity
Check whether the structure of the series is changing over time.
Step 2
Walk-Forward Validation
Avoid random splits and evaluate models in time order.
Step 3
Data Drift
Monitor PSI/KS and connect distribution shift to model decay.
Case 2
Household Electricity
Four years of real meter readings: missing minutes, six naive baselines walk-forward, a model against the 7-day mean.
Advanced
Explainability and Market Evidence
~90 min · evidence-aware decisions
Step 1
SHAP Values
Explain predictions while avoiding causal overclaims.
Step 2
Market Evidence Legend
Separate statistical risk tools from heuristic chart and indicator hypotheses.
Step 3
Validate Indicators
Treat indicator signals as hypotheses and test costs, drawdown, and walk-forward stability.
Case 3
The 200-Day Rule
A century of daily US market returns: where the edge came from, and a one-day look-ahead bug.
Guides
Start From the Problem
Decision guides and checklists: which test, which metric, which forecasting model, is my backtest honest.
Cases
Real-Data Case Studies
Housing, credit, electricity and a century of market returns — every step runnable in the browser.
Sandbox
Build Mode
Save, compare, export JSON, copy Python, and reset interactive experiments.
Lab
JupyterLite
The case notebooks in Jupyter, in the browser, with nothing to install.