Machine Learning, Autumn 2021

CSE 5523, The Ohio State University, Autumn 2021

Economics has a long tradition in processing data that holds a different view compared with Computer science. Applied Economists dig into data to find the correlation and causality between different factors, while Computer scientists have developed Machine Learning to predict out-of-sample data points. In other words, Economists try to interpret data in the past to understand the causality, while Computer scientists try to make future predictions based on the pattern existing in the data. This report aims to investigate the effectiveness of the recurrent neural network to replicate the interpretation power that linear regression can provide, and also the preciseness of the linear regression in out-of-sample prediction. My result shows that (1) linear interpretation is good within-sample interpretation but performs devastatingly in out-of-sample prediction, (2) trained with long enough epochs, the recurrent neural network can reach the interpretation power of linear regression, but maintaining good enough out-of-sample predictions power requires a high number of layers, and (3) Both GRU and LSTM performs generally better than SimpleRNN, and the design of reset gate in GRU can prevent noise from outliers, while the outcome generated by LSTM exacerbates with outliers.

Article tags: Miscellaneous