Screenshot 2023-05-29 130622

Machine learning functions well when financial theory does not dictate the choice of variables or when incorporating linear specifications is undetermined. Machine learning models use flexible functional forms to capture complex (and potentially non- linear) variable interactions, which differs from what is required with linear regression models. Explanatory variable independence, normal distributions, statistical significance, error-term testing, and quality of regression fit are critical for conventional models, but they are not applicable to “supervised” machine learning, which is focused on prediction accuracy. The differences in key terminology between conventional econometrics and machine learning are shown in Figure 25.1. Figure 25.1: Terminology Differences Conventional Econometrics Machine Learning Independent Variables Inputs or Features (values) Dependent Variables Outputs or Targets (forecasted values, known as labels) Supervised, Unsupervised, and Reinforcement Learning Models LO 25.g: Differentiate among unsupervised, supervised, and reinforcement learning models. Machine learning methodologies land in the following three categories: » Supervised learning: used to predict either the value of a variable (e.g, the value of a car) or the classification of an observation (e.g, the outcome of a sports team’s next game—win or loss). The algorithm uses “labeled data” to learn. For the value of a car, this data would include the make, model, year, size, engine type, etc. For the outcome of a sporting event, this data would include current record, location of the event, player years of experience, etc. » Unsupervised learning: pattern recognition in data with no specific target. Data clustering and the identification of a small group of explanatory factors may be involved. » Reinforcement learning: incorporates a trial-and-error approach to make decisions in a changing environment. Supervised learning can be used to predict a variable’s value in a time-series (predicting Treasury note yields one year from now) or a cross-sectional scenario for data not in a sample (the list price for a house based on comparable prices in the neighborhood). Unsupervised learning is not used for predictions; however, it is very useful for characterizing and learning the structure of a dataset. An internal auditor may use unsupervised learning to evaluate various company transactions on specific ledger accounts to determine if further review and investigation is needed. 191