Model Performance and Algorithmic Fairness in Finance

the bias-variance trade-off

The balance between model bias and variance, where an optimal model complexity minimizes both sources of error.

Out-of-Sample R^2

R^2 value calculated on data not used in the model training process

SLIGHTLY HIGH R^2 AND A SLIGHTLY LOWER TRAINING R^2

Having a better training R^2 does not necessarily mean a better test set R^2; more data may be needed for conclusive results. Out-of-sample R^2 can even be negative.

Model Performance and Assessment

Process of evaluating the effectiveness and accuracy of a model in predicting outcomes.

FOR QUIZ: how to pick number of variables

A question regarding the method for determining the optimal number of variables in a model.

Ordinary Least Squares Regression

A regression technique that estimates the relationship between independent and dependent variables by minimizing the sum of the squared differences between observed and predicted values.

linear relationship = linear model

A linear model describes a situation where one quantity changes at a fixed rate, leading to another quantity changing dependently at a different fixed rate.

Lasso and Elastic Net penalized regression

Regression techniques that apply penalties to the coefficient estimates to prevent overfitting, particularly used in high-dimensional datasets.

alleged discrimination

Accusation of discrimination against women with equal financial capabilities.

regulatory violation by Apple or Goldman Sachs

Discussion on whether Apple or Goldman Sachs violated regulations, particularly related to credit limits and discrimination.

pattern in people reporting discrimination

Observation that discrimination reports are often from white men who are extremely rich.

consequences of not getting a needed credit card

Limits financial capabilities and credit growth, impacting access to essentials like food and rental services.

consequences of receiving a lower credit line than deserved

Restricts financial decisions, potentially hindering major purchases or business ventures.

impact of financial system discrimination on underserved groups

Results in lower access to opportunities and perpetuates wealth disparities across racial and minority groups.

NY Regulator instigating Apple Card for possible gender bias

Investigation by the NY regulator into potential gender bias in credit limit setting by Goldman Sachs related to the Apple Card.

How the law got it wrong with Apple Card

Discussion on the legal assessment of the alleged gender bias in the Apple Card credit limit setting.

DFS under current law assess the fairness of the Apple card

Explanation of how the NY Department of Financial Services evaluates the fairness of the Apple Card credit limit algorithm under existing laws.

new vocab for algorithmic fairness

Introduction of terminology related to assessing algorithmic fairness and potential discrimination in financial systems.

Financial services' Catch-22

Dilemma where underrepresented groups may not benefit from system audits due to data collection restrictions, potentially perpetuating algorithmic discrimination.

business implications of AI fairness concerns

Impact on businesses using AI, including policy changes like sharing credit files and the slow response of regulatory bodies to algorithmic discrimination concerns.

Correlation in data analysis

Measure of the linear relationship between variables, ranging from -1 to +1, with 0 indicating no linear relationship.

supervised learning vs. unsupervised learning

Supervised learning involves having a target variable (Y), while unsupervised learning does not; regression is a form of supervised learning.

classification in machine learning

Type of supervised learning with a categorical target variable, such as binary outcomes like yes/no.

regression in machine learning

Supervised learning method that predicts a continuous valued output based on a set of independent variables, often involving least squares methods.

Predicting quality of Wine case study

Case study involving predicting wine quality using linear regression based on variables like age, weather conditions, and historical data.

One-variable linear regression

Regression model with a single independent variable, where prediction is solely based on the average of the dependent variable.

error measures in regression analysis

Different metrics like SSE, RMSE, and R^2 used to evaluate the accuracy and performance of regression models.

Who tested the quality of wine using predition?

Orley Ashenfelter