WSU Lexicon: Quantitative Methods
Machine Learning
Machine learning (ML) is a set of computational methods that let algorithms identify patterns in data and improve their predictions through experience, without the analyst specifying the model's rules in advance. In finance it supports classification, forecasting and pattern discovery in large datasets.
Generative AI and large language models are specialized branches of deep learning, itself a subset of machine learning.
The entry
Supervised, Unsupervised and Deep Learning Approaches
Supervised learning uses labeled data: each observation pairs a set of input features with a known target value, and the algorithm learns the mapping from one to the other. When the target is continuous, such as a future return, the task is regression; when it is categorical, such as default or no default, the task is classification. Unsupervised learning works without a target and instead searches for structure, either grouping similar observations through clustering or compressing many correlated features into fewer factors through dimension reduction.
Deep learning stacks many layers of interconnected nodes in a neural network, allowing highly nonlinear relationships to be captured from images, text or audio. Reinforcement learning trains an agent to choose actions that maximize cumulative reward through trial and error, an approach explored in trade execution.
The entry
How Models Are Trained, Tuned and Tested
Practitioners divide data into three parts. The training sample is used to fit the model, the validation sample is used to tune hyperparameters such as the number of network layers or the depth of a decision tree, and the test sample is kept untouched until the end to measure out-of-sample performance. Where data are scarce, k-fold cross-validation rotates each slice of the data through the validation role.
Total out-of-sample error has three sources: bias error, which reflects a model too simple to capture the true relationship (underfitting); variance error, which reflects a model so complex that its fit changes with each sample (overfitting); and base error from irreducible randomness. Adding complexity lowers bias but raises variance, so the goal is the level of complexity that minimizes total error, not training error.
The entry
Generative AI and Large Language Models in Research
Large language models are deep neural networks trained on vast text corpora to predict the next word, or token, in a sequence. That simple objective produces systems able to summarize annual reports, extract figures from earnings call transcripts, draft research notes and classify news sentiment at scale. Generative models of this kind create new content rather than only scoring existing data.
Their weaknesses matter for investment use. Models can produce fluent but fabricated statements, known as hallucinations, may reflect biases in their training data, and rarely explain how an output was reached. Feeding client or confidential data into external tools also raises privacy and professional conduct concerns, so outputs require human verification before they inform recommendations.
The entry
Applying Machine Learning Concepts on the Exam
Candidates should be able to match a problem to a method. Predicting a bond's credit spread from issuer ratios is supervised regression; flagging whether a borrower will default is supervised classification; grouping funds by return behavior with no predefined categories is unsupervised clustering. They should also recognize the symptoms of poor generalization, such as a large gap between training and test accuracy.
A further distinction is purpose. Traditional econometrics emphasizes inference, testing whether a specific coefficient is significant under stated assumptions. Machine learning emphasizes prediction accuracy and tolerates complex, less interpretable structures, which is why explainability, data quality and the risk of data mining remain central concerns when ML outputs drive portfolio decisions.
Common trap
“A model that fits its training data almost perfectly will forecast well on new data.”
Near-perfect in-sample fit usually signals overfitting: the model has learned noise specific to the training sample. Performance must be judged on held-out validation and test data, where an overfit model's errors rise sharply.
In practice
Supervised Learning in Card Fraud Detection
Card networks and issuing banks train classifiers on historical transactions labeled fraudulent or legitimate. Each new authorization is scored in milliseconds using features such as amount, merchant type and location, and high-risk payments are declined or flagged for review.
Candidates ask
Questions on this term
- Is machine learning just a more complex form of regression?Linear regression is one of the simplest supervised learning models. Machine learning extends it with flexible, nonlinear methods and prioritizes out-of-sample prediction over testing hypotheses about individual coefficients.
- What is overfitting, and how do analysts reduce it?Overfitting occurs when a model is complex enough to fit noise in the training sample, so it performs well in-sample but poorly on new data. Analysts reduce it by holding out validation and test samples, using k-fold cross-validation, limiting complexity through techniques such as penalizing large coefficients or pruning decision trees, and preferring the simplest model with acceptable out-of-sample accuracy.
- How do generative AI and large language models relate to machine learning?They sit inside it. Deep learning is a subset of machine learning built on multilayer neural networks, and large language models are deep learning systems trained on text. Generative AI refers to models that produce new content such as text or images. For analysts they are productivity tools whose outputs must be checked, because they can state errors with apparent confidence.
See also
Related entries
- R-SquaredR-squared, or the coefficient of determination, is the fraction of the variation in the dependent variable explained by the regression model.
- ResidualA residual is the difference between an observed value of the dependent variable and the value predicted by the estimated regression line.
- Robo-AdviserA robo-adviser is an automated digital platform that builds and manages a client portfolio using algorithms driven by an online questionnaire covering goals, time horizon and risk tolerance.
- Simple Linear RegressionSimple linear regression estimates a straight-line relationship between one independent variable and one dependent variable.
One entry from the WSU Lexicon every day. Browse the full archive.