Charta is on Google Play: practise on the go. Get the app
Charta

Explain deep, supervised and unsupervised learning and how investment managers use them

CFA® Level I · Quantitative Methods · Introduction to Financial Data Science · QM.20.a

The idea

Machine learning finds patterns in data without the analyst specifying the model's form in advance. Supervised learning trains on inputs labelled with the correct output: predicting a category, such as default or no default, is classification, and predicting a number is regression. Unsupervised learning has no labels and looks for structure, as in clustering similar stocks or principal components analysis, which condenses many correlated variables into a few factors. Deep learning uses neural networks with many layers that learn their own features, suiting large volumes of text, images and audio. The recurring danger is overfitting: a model that fits its training data closely but performs poorly on new data.

A practice question

From the question bank, with the reason each wrong option is wrong.

Supervised learning is best described as training a model on:

  1. inputs paired with known target outputs — correct
  2. inputs with no target outputs at allWhy not: Learning structure from unlabelled inputs is unsupervised learning.
  3. rewards gained from trial-and-error playWhy not: Learning from rewards through trial and error is reinforcement learning.

In supervised learning each training example is labelled with the correct answer, and the model learns a mapping from inputs to that target so it can predict it for new data.

A harder one

A model is trained on past bond issues labelled as defaulted or not defaulted, and then predicts default for new issuers. This task is:

  1. unsupervised clusteringWhy not: The training data carry known default labels, so the learning is supervised.
  2. supervised regressionWhy not: The target is a category (default or not), so the task is classification, not regression.
  3. supervised classification — correct

Labelled data make it supervised; a categorical target makes it classification. Predicting a continuous value such as a credit spread would be regression.

The hardest tier

A stock-selection model is 98% accurate on its training data but only 61% accurate on new data. The most likely problem is:

  1. underfittingWhy not: An underfit model performs poorly in and out of sample alike, not well on the training data.
  2. overfitting — correct
  3. a lack of data labelsWhy not: The model reports training accuracy, so the data must have been labelled.

A large gap between in-sample and out-of-sample performance means the model learned noise specific to the training set. Remedies include simpler models, regularisation and cross-validation.

In the app: 6 practice questions on this outcome, a narrated walkthrough with diagrams, spaced review that brings it back before you forget it.

Take the free diagnostic →

More in Introduction to Financial Data Science