Statistical Learning

Statistical learning is the framework for learning from data: choose a model class, define loss, and fit by minimising empirical risk (e.g. squared error, cross-entropy). Core ideas include bias–variance tradeoff (simple models underfit, complex overfit), regularisation (penalise complexity: Ridge, Lasso), and cross-validation (hold out data to estimate generalisation). The additive rule and linear models are building blocks; modern ML extends this with flexible function classes and optimisation. The discipline is in controlling overfitting, choosing the right complexity, and interpreting what was learned—not just fitting curves.

Define the target and the loss; choose model complexity appropriate to sample size. Use cross-validation or holdout to estimate generalisation; regularise to reduce overfitting. In business: treat "learning from data" as statistical—expect noise, guard against overfitting, and validate on out-of-sample or out-of-time data.

Statistical learning fits models from data with controlled complexity (bias–variance, regularisation) and validation (cross-validation, holdout). Use it whenever you learn from data: define target and loss, avoid overfitting, and validate on unseen data.