Overview

Topic 1 built the ERM recipe around a regression target — a real number. Most quant classification problems — will this bond default, is this a fraudulent trade, will this options position expire in the money — have a discrete target instead, and squared-error loss is the wrong tool for it. This topic swaps the hypothesis class and loss for classification: logistic regression as the discriminative workhorse, its Newton's-method optimiser, the perceptron and softmax for multi-class problems, the generalized linear model (GLM) framework that shows linear and logistic regression are the same algorithm applied to different output distributions, generative classifiers (Gaussian discriminant analysis, naive Bayes) as the alternative modelling philosophy, and the evaluation metrics — precision, recall, ROC/AUC — that a squared-error test-set score cannot supply for a classifier.