Skip to content

CFA Level II · CFA Level II Exam

Machine Learning for CFA Level II: Chapter Guide

Machine Learning in CFA Level II covers how algorithms learn patterns from data. You study supervised methods (penalized regression, SVM, CART, ensembles), unsupervised methods (PCA, clustering) and neural networks. You solve questions by matching the problem in the vignette to the right model, then judging its fit, overfitting risk and limits.

What this chapter covers

This chapter introduces machine learning (ML) as a set of tools for finding patterns in large datasets. It splits into supervised learning, where you have labelled outcomes, and unsupervised learning, where you do not. It also covers neural networks, deep learning and reinforcement learning. The chapter is about concepts, not heavy maths.

You will not be asked to code or to run a model by hand. Item sets give you a short description of a firm, a dataset or a model output, and ask you to pick the right algorithm, spot a flaw, or interpret a result. Expect questions on overfitting, bias-variance trade-off, training versus validation versus test data, and how to read a confusion matrix.

The chapter links to several other areas. Penalized regression and overfitting build on regression in Quantitative Methods. Big data and alternative data link to Equities and Portfolio Construction. ML in credit scoring and fraud detection links to Fixed Income and Financial Statement Analysis. Clustering and PCA connect to risk factor and diversification ideas in portfolio work.

Machine Learning sits within Quantitative Methods, which carries a weight of 5-10% for the whole topic. No separate weight is published for this chapter. You are tested on definitions, model selection and interpretation, so careful study turns into points without long calculations. Because every question sits in an item set and must be answered from the vignette, practising how to match a described situation to an algorithm is what earns marks here. Candidates who skip this chapter as too technical give up easy points.

Machine Learning: topics in the order to study them

  1. 1Overview of Machine Learning and OverfittingIt sets the vocabulary (features, targets, training and test sets, bias, variance) that every later topic uses.
  2. 2Supervised Learning: Penalized Regression and SVMPenalized regression extends ordinary regression you already know, so it is the easiest first model and shows how overfitting is controlled.
  3. 3Supervised Learning: CART and Ensemble LearningTrees and ensembles add classification ideas and show how combining models reduces error, building on the supervised framework.
  4. 4Unsupervised Learning: PCA and ClusteringOnce you know supervised learning, the contrast with unlabelled data is clear, and these methods reduce or group data.
  5. 5Neural Networks, Deep Learning and Reinforcement LearningThese are the most advanced models and are easier to place once you hold the earlier concepts, so they come last.

How to prepare Machine Learning

Aim for clear concepts and fast model recognition. Short, frequent sessions work well, including on your phone during a commute.

  1. Build a one-page map first: supervised versus unsupervised, regression versus classification, and which algorithm belongs where.
  2. For each model, learn four things: what problem it solves, how it works in one sentence, its main strength and its main weakness.
  3. Master overfitting early. Be able to explain bias, variance, and how cross-validation, penalties and pruning address them.
  4. Practise reading short vignettes and underlining the clues: labelled or unlabelled data, number of features, need for interpretability, categories or numbers as the target.
  5. Learn to read a confusion matrix and the measures from it: precision, recall, accuracy and F1. Practise these calculations until they are quick.
  6. Do item sets under time pressure and review each wrong answer by naming the clue you missed.
  7. Revise with a short list of contrasts, such as LASSO versus ridge, bagging versus boosting, and PCA versus clustering.

Common mistakes in Machine Learning

  • Mixing up supervised and unsupervised methods

    Fix: Ask first whether the vignette gives a known target. If yes, it is supervised; if it only asks to group or simplify, it is unsupervised.

  • Confusing LASSO with ridge

    Fix: Remember that LASSO can set coefficients to zero and so selects features, while ridge only shrinks them.

  • Treating overfitting as only a training problem

    Fix: Tie overfitting to out-of-sample performance: strong training results with weak validation or test results signal it.

  • Swapping precision and recall in confusion matrix questions

    Fix: Precision asks how many predicted positives are right; recall asks how many actual positives were found. Write the formula before computing.

  • Treating PCA components as original, interpretable variables

    Fix: Remember that components are combinations of the original features and are often hard to interpret, which is a stated drawback.

  • Choosing the most complex model as the best answer

    Fix: Pick the model that fits the data size, the need for interpretability and the problem type given in the vignette.

Last-day revision: Machine Learning

  • Supervised learning uses labelled data; unsupervised learning finds structure without labels; reinforcement learning learns from rewards.
  • Overfitting means a model fits training data too closely and performs poorly on new data.
  • High bias means underfitting; high variance means overfitting.
  • Data is split into training, validation and test sets; the test set is used only for final evaluation.
  • Penalized regression adds a penalty on coefficient size; LASSO can shrink coefficients to exactly zero, so it selects features.
  • SVM finds the boundary that gives the widest margin between classes.
  • CART builds a tree of splits; a deep tree overfits, and pruning reduces this.
  • Bagging averages many models trained on resampled data; boosting builds models in sequence to fix earlier errors.
  • A random forest is a bagged ensemble of decision trees, each trained on a bootstrap sample and using a random subset of features at each split.
  • PCA reduces many correlated features into fewer uncorrelated components; clustering groups similar observations.
  • K-means needs you to choose the number of clusters; hierarchical clustering does not.
  • Precision = TP ÷ (TP + FP); recall = TP ÷ (TP + FN); accuracy = (TP + TN) ÷ all observations.

Machine Learning in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

Machine Learning: frequently asked questions

Do I need to know coding or maths for Machine Learning at Level II?

No coding is tested. You need concepts, model selection and interpretation of outputs, plus simple confusion matrix calculations. Read the vignette for clues and match them to the right algorithm.

How should I order my study of this chapter?

Start with the overview and overfitting, then supervised methods, then unsupervised methods, and finish with neural networks and reinforcement learning. This order moves from familiar ideas to the most advanced ones.

What is the quickest way to tell which algorithm to use in an item set?

Check whether the data has labels, whether the target is a category or a number, and whether interpretability matters. Those three clues eliminate most wrong options.

How do I avoid confusing bagging and boosting?

Bagging trains many models independently on resampled data and averages them to cut variance. Boosting trains models one after another, each focusing on the errors of the last.