Structured data

Regression

price, lifetime value, delivery time, energy load

A continuous target per row.

Maximum likelihood

Linear algebra

Bias & variance

Hypothesis testing

Foundations

Least squares is maximum likelihood under Gaussian noise. Regularisation is the bias-variance tradeoff made explicit.

Frame the problem

Frame

Decide whether you need the mean, a quantile, or an interval. Errors are rarely symmetric; say which direction hurts.

Sourcing & signal

Exploring the data

Hunting for leakage

Cleaning the data

Data & labels

Look at the target distribution; long tails usually want a log transform. Leakage hides in fields filled in after the outcome: final invoice, actual delivery time.

Encode & scale

Create features

Reduce & select

Represent

Split the data

Cross-validation

Split

Group by the entity if it repeats. Temporal if prices drift.

Dumb baseline

Linear models & GLMs

Regularised linear

K-nearest neighbors

Gradient boosting

Kernel methods

Model

Predict the mean, then a linear model, and read its coefficients. Ridge when features are many and correlated, lasso when you want fewer. Then gradient boosting. Kernel methods or GPs when data is small and you need uncertainty.

Loss functions

Hyperparameter tuning

Over- & underfitting

Ensembling

Train

Pick the loss to match the framing: MAE for medians, quantile loss for intervals, Huber for outliers.

Regression metrics

Interpretability

Error analysis & significance

Evaluate offline

RMSE and MAE together. Residuals against every feature. Slice by price band.

Evaluate online

Nothing unusual here.

Features at serving time

Serving & release

Monitor & retrain

Ship & monitor

React Flow