FRM Part I · FRM Exam Part I

Machine Learning and Prediction for FRM Part I

Machine learning uses data to find patterns and make predictions without hand-written rules. For FRM Part I, you must know supervised, unsupervised and reinforcement learning, overfitting, regularization, PCA, clustering, trees, neural networks and classification metrics. Solve questions by identifying the method, then applying its formula or logic step by step.

What this chapter covers

This chapter covers how machine learning methods build models from data and how you judge whether those models predict well. It starts with the types of learning: supervised (labelled outcomes), unsupervised (no labels) and reinforcement (learning from rewards). It then moves to the core problem of any model: fitting the data you have without failing on data you have not seen.

The middle of the chapter covers tools. Regularization (Ridge, LASSO, Elastic Net) controls model complexity. PCA reduces many correlated variables to a few components. Clustering groups similar observations. Decision trees, random forests and neural networks handle non-linear patterns. The chapter ends with evaluation: confusion matrices, precision, recall and related measures.

This chapter builds on Quantitative Analysis. Regression, estimation, hypothesis testing and variance ideas all return here in new form. It also links to the Valuation and Risk Models topic, where you judge model quality and model risk, and to Foundations of Risk Management, where model governance and data quality matter. Questions are often conceptual, with some short calculations.

Machine learning is a newer part of the FRM curriculum, and many candidates under-prepare it because it feels less familiar than VaR or bond maths. That is your opening. The concepts are finite and the calculations are short, so careful study can turn this chapter into reliable marks on a 100-question exam. Many questions test whether you can pick the right method for a stated problem, and that skill is learnable. It also reinforces regression and bias-variance ideas that you need elsewhere in Quantitative Analysis. Always check the current GARP Study Guide and Learning Objectives, since the curriculum is revised every year.

Machine Learning and Prediction: topics in the order to study them

1Machine Learning Basics and Types of LearningIt gives you the vocabulary (features, labels, supervised, unsupervised, reinforcement) that every later topic uses.
2Overfitting, Bias-Variance Tradeoff and Data SplittingThis is the central problem of the chapter, and it explains why the next methods exist.
3Regularization: Ridge, LASSO and Elastic NetIt is the first direct fix for overfitting and builds on regression you already know.
4Dimension Reduction and Principal Components AnalysisIt moves you to unsupervised learning and uses variance and covariance ideas from Quantitative Analysis.
5Clustering: K-Means and Hierarchical MethodsIt completes the unsupervised methods once you are comfortable with distance and similarity.
6Decision Trees, Ensembles and Random ForestsIt returns to supervised learning with non-linear models, and ensembles reduce variance, linking back to the bias-variance tradeoff.
7Neural Networks and Deep LearningIt is the most flexible and complex model family, so it is easier after trees and regularization.
8Model Evaluation and Classification MetricsIt ties everything together: you now score the models you have learned, using confusion-matrix measures.

How to prepare Machine Learning and Prediction

Treat this as a concepts-plus-short-calculations chapter. Aim to explain each method in two sentences and compute each metric by hand.

Read the GARP learning objectives for this chapter first, and turn each one into a question you must be able to answer.
Build a one-page table: for each method, note whether it is supervised or unsupervised, what it does, its main strength and its main weakness.
Learn the bias-variance logic until you can predict what happens to training error and test error as complexity rises.
Practise the calculations: confusion-matrix metrics, share of variance explained by principal components, and simple distance steps in K-Means. Write each formula and the numbers before computing.
Compare methods in pairs, such as Ridge versus LASSO, K-Means versus hierarchical, single tree versus random forest. Exams often test the difference.
Do timed mixed questions, about 2 to 3 minutes each, and log every miss by cause: concept, formula or misreading.
In the final days, rework only your error log and the quick revision list.

Common mistakes in Machine Learning and Prediction

Mixing up precision and recall.
Fix: Remember the denominators: precision divides by everything predicted positive (TP + FP); recall divides by everything actually positive (TP + FN).
Saying LASSO and Ridge both drive coefficients to exactly zero.
Fix: Link LASSO to selection (exact zeros) and Ridge to smooth shrinkage. Elastic Net combines both effects.
Using the test set to tune the model.
Fix: Tune on the validation set or by cross-validation. Touch the test set once, at the end, for an honest estimate.
Treating PCA as a supervised method or as variable selection.
Fix: PCA ignores outcomes and builds new combinations of the original variables. It does not choose a subset of them.
Trusting accuracy when classes are imbalanced.
Fix: Check the base rate and look at precision, recall and the confusion matrix. A model that always predicts the majority class can score high accuracy and still be useless.
Assuming a more complex model is always better.
Fix: Judge models by out-of-sample performance. More flexibility can raise variance, hurt interpretability and add model risk.

Last-day revision: Machine Learning and Prediction

Supervised learning uses labelled outcomes; unsupervised finds structure without labels; reinforcement learns from rewards.
Overfitting means low training error but high error on new data; the model has learned noise.
Higher model complexity usually lowers bias and raises variance.
Use separate training, validation and test sets; the test set is for final evaluation only.
Ridge shrinks coefficients toward zero but does not usually set them exactly to zero.
LASSO can set coefficients exactly to zero, so it performs variable selection.
Elastic Net blends the Ridge and LASSO penalties.
PCA components are uncorrelated; the first explains the most variance, and the shares of variance across all components sum to 100%.
K-Means needs you to choose the number of clusters K in advance; hierarchical clustering does not.
Random forests average many trees built on random samples and features to reduce variance.
Precision = TP ÷ (TP + FP); recall = TP ÷ (TP + FN); accuracy = (TP + TN) ÷ total.
Accuracy can mislead when classes are imbalanced.

Machine Learning and Prediction practice questions

Machine Learning and Prediction 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 and Prediction: frequently asked questions

Do I need to know programming for the machine learning chapter?

No. The FRM Part I exam is multiple choice and tests concepts, interpretation and short calculations. You do not write code. You should understand what each method does and when to use it.

How much maths is in this chapter?

Mostly light. Expect confusion-matrix metrics, variance explained by components and simple reasoning about penalties and errors. Heavy derivations are unlikely, but always check the current GARP learning objectives.

Which topics in this chapter should I study first?

Start with the types of learning and the overfitting and bias-variance tradeoff. Almost every other topic is a response to the overfitting problem or a way to evaluate it, so these two give you the framework.

Is this chapter worth the effort given the whole exam has 100 questions?

Yes. GARP does not publish a pass mark, so you should aim to be solid on every topic. The machine learning material is finite and learnable, which makes it a good place to secure marks.