The Supervised Learning Problem and the ML Pipeline

A quant who fits a model without first fixing what "fitting" means — what counts as data, what counts as an error, what counts as success on unseen data — will produce a number that looks like a Sharpe ratio and behaves like noise. This note fixes the vocabulary before any algorithm is introduced: a hypothesis class, a loss function, an optimiser, and a train/validation/test protocol, assembled into the single objective every supervised-learning algorithm in this unit minimises — empirical risk. It also places supervised learning inside the broader taxonomy of learning paradigms, since later topics (Topic 7, Topic 10) drop the labels or the losses entirely.