Date Thesis Awarded

5-2019

Document Type

Honors Thesis

Degree Name

Bachelors of Science (BS)

Department

Economics

Advisor

Matthew Klepacz

Committee Members

John Lopresti

Martin White

Abstract

This research asks whether satellite imagery of metropolitan statistical areas (MSAs) can aid in the prediction of unemployment over different time horizons. I make these predictions using Convolutional Neural Networks (CNNs). The data consist of 1735 Landsat Analysis Ready Data (ARD) images from 1990 to 2011, each averaged quarterly with a subject that is one of 27 MSAs. I match each quarterly image with a known value of unemployment collected by the Bureau of Labor Statistics. Ten training configurations use different combinations of data augmentation and image preprocessing. I find CNN training accuracy around 10 to 20 times the accuracy of random selection, suggesting that satellite imagery contains information about unemployment amenable to machine learning analysis. Test accuracy at best, by contrast, is around double the accuracy of random selection, raising concerns about out-of-sample generalization. Two training/test set splits then explore the use of CNN predictions as features in linear regression forecasts of unemployment. In the random split, in two out of ten data configurations, mean-squared error is reduced. In the time-based split (all training data before 2001, all testing data afterwards), in eight out of ten data configurations, mean-squared error is reduced. Multiple models are estimated with other features such as national macroeconomic data, quarter/year fixed effects, MSA fixed effects, and unemployment lags. The MSE reductions are corroborated by findings that the CNN prediction regressor is widely statistically significant among both training/test splits, data configurations, and model types.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License.

Share

COinS