Use of Machine Learning Algorithms
Friday, Jan. 3, 2020 10:15 AM - 12:15 PM (PDT)
- Chair: Ying Zhu, University of California-San Diego
Credit Growth, the Yield Curve and Financial Crisis Prediction: Evidence from a Machine Learning Approach
AbstractWe develop early warning models for financial crisis prediction using machine learning techniques on macrofinancial data for 17 countries over 1870-2016. Machine learning models mostly outperform logistic regression in out-of-sample predictions and forecasting. We identify economic drivers of our machine learning models using a novel framework based on Shapley values, uncovering nonlinear relationships between the predictors and crisis risk. Throughout, the most important predictors are credit growth and the slope of the yield curve, both domestically and globally. A flat or inverted yield curve is of most concern when nominal interest rates are low and credit growth is high.
Machine Learning for Predicting ‘Rare-Events’ of Infant and Neo-Natal Mortality and Identifying Leading Indicators for Early Interventions
AbstractPredicting ‘rare events’ with accuracy is a fundamental challenge in econometrics. The importance of identifying leading indicators or reliable predictors of many high-impact though not-so-common events (such as mortgage foreclosure, loan default or child death) cannot be overstated. We use a large household survey data from India which includes social, economic, and health (birth and pregnancy related) information on infants (less than one year), newborns (less than one month) and their mothers. Using this information, we aim at identifying characteristics (leading indicators) that are commonly associated with higher probabilities of neonatal or infant deaths and thereby selecting a high-risk group for early interventions. However, since death of a child is much rarer compared to survival, this particular classification problem essentially involves high imbalance in the binary dependent variable (very low occurrence of ‘events’ compared to ‘non-events’). Standard econometric techniques do not predict such ‘rare events’ well. We employ several Machine Learning (ML) methods including Logistic LASSO, Classification Random Forest, Boosted Classification Trees and compare their prediction performance with traditional logistic regression. We also use ML methods such as RUSBoost and SMOTEBoost that are tailor-made to deal with rare events. Our ML methods not only render up to 42% higher prediction accuracies compared to logistic regression, but also largely agree upon set of predictors associated with higher probabilities of child deaths, thereby enabling us to identify high-risk group for making targeted interventions by policy makers. For neonatal deaths such characteristics include various information related to pregnancy, newborn care, delivery complications, mother’s employment and education, income, and child’s birth weight. For infant mortality, the list adds vaccinations (also indicating health awareness of the family). Our research can be useful for global ‘Every Newborn Action Plan’ at large, and especially for India’s ‘New born Action Plan’, targeted to combat child mortality problem.
Predicting Access to Healthy Food in the United States with Machine Learning
AbstractA balanced, nutritious diet is essential for good health, but around eight percent of all U.S. households still lack access to healthy food. Those households rely on local fast-food restaurants and other sources of inexpensive, processed food with little or no nutritional value, because quality fresh food means paying higher or travelling far. Surveying access to healthy food is costly, and finding the determinants of access remains convoluted for multidimensional nature of socio-economic variables.
The study uses Random Forests—a popular machine-learning model—to predict the retail food environment in the United States at census tract level. The predicted variable, modified Retail Food Environment Index (mRFEI), refers to the percentage of healthy food retailers in a tract. This specification also generates two binary measures of access to healthy food: a census tract is food ‘desert’ if no healthy food retailer exists, and food ‘swamp’ if healthy retailers are disproportionally outnumbered by unhealthy ones. We train the model with socio-economic information of 45 thousand tracts, and test on another independent five thousand tracts. Random Forests searches over 282 variables and ranks them by their prediction efficiency. Findings suggest that access to healthy food significantly differs across demographic features. Only 17 demographic variables—especially population density, property value, and employment—can predict access to healthy food with high accuracy. The model accurately identifies food desert and food swamp 100% of the time in the training data, and respectively 78% and 61% of the time in the independent test data. Our model can be used to get a sensible prediction of access to healthy food for any U.S. tract with a handful demographic information.
Predicting Success Among Female Entrepreneurs: Evidence from Three African Countries
AbstractWe use key business indicators, heuristic economic models, and machine learning algorithms to identify successful female entrepreneurs in Ethiopia, Tanzania, and Togo. When concentrating on the country with the largest dataset (Tanzania), we show that past performances - such as baseline profits, sales, and number of employees - are powerful predictors of future profit levels and growth rates. Machine leaning algorithms achieve higher performance in some, but not all, simulations. Nevertheless, the overall (out-of-sample) goodness-of-fit of all approaches is rather low. We then show that we can substantially raise the accuracy of our models when increasing the size of the training sample by combining together the available data from the three African countries. In particular, we find large improvements in the predicting power of machine learning algorithms and when we focus on identifying firms in the top tail of the profit distribution.
Predictive Power at What Cost? Economic and Racial Justice of Data-Driven Algorithms
AbstractData-driven algorithms are increasingly being used to inform decision making across consequential settings. Conventional data-driven algorithms leverage many input variables to maximize overall predictive power. In doing so, algorithms may include variables that only marginally improve predictive power, without consideration of whether using these variables introduce unequal treatment across groups. Unequal treatment can result if group differences are disproportionately explained by certain group variables that actually contribute little to predictive power. Using a decomposition framework, I assess how algorithms trade off predictive power and equal treatment across groups. I examine a nationally used recidivism risk assessment tool using pretrial defendant case and risk score data from 2013-2016 in Broward County, Florida, combining publicly available data with novel data that I assemble. I find that variables -- such as defendants' neighborhoods -- can marginally improve recidivism prediction and help to explain risk scores received by defendants, while inducing substantial disparities in risk scores across race and economic status. In particular, average differences in defendant's neighborhood characteristics explain less of the overall variation in risk scores, than they explain of the Black-White gap in risk scores and the gap in risk scores between “indigent” defendants who use a public defendant, and “non-indigent” defendants who do not. These dramatic disparities reveal that black and indigent defendants have disproportionately higher risk scores on average. Higher risk scores may lead to longer pretrial incarceration and downstream consequences, by impacting labor market outcomes. These findings underscore that machine learning objectives tuned to maximize predictive power can be in conflict with racial and economic justice.
- C5 - Econometric Modeling