Predicting U.S. business cycles with recurrent neural networks : An extensive multivariate time-series analysis for comparing LSTM and GRU networks
Year: 2020
Type: dissertation
Abstract: This study examines how 22 different long short-term memory (LSTM) and gated recurrent unit (GRU) network architectures suit predicting U.S. business cycles. The networks create 91-day forecasts for the dependent variable by using multivariate time-series data comprising 26 leading indicators’ values for the previous 400 days. The proposed models are evaluated by using a train-test split, where the proposed models are trained with data from 1980 to 2005, and the out-of-sample set consists of data between 2005 and 2015. The performance is evaluated by using mean squared error (MSE) and mean absolute error (MAE), and early warning signs are also considered beneficial. The training algorithm consists of typical deep learning methods. MSE and L1 regularization are used for determining the cost, and minibatches of 32 examples are applied together with Nesterov accelerated momentum (NAG) learning algorithm. Early stopping is introduced to halt the training process when strong signs of overfitting are detected. Each proposed recurrent neural network (RNN) architecture is trained three times, and these three networks’ averaged predictions are examined when comparing the architectures. Performance-wise, a few LSTM networks stand out from the other proposed networks. Although the performance results favor the proposed LSTM networks slightly over their GRU equivalents, the difference is not substantial and, in turn, the proposed GRU networks offer less deviation in MSE and MAE between each architecture. However, these steadier performance results do not generate less volatile forecasts. Instead, the best performing networks and architectures differentiate by offering less volatile predictions that also vary less from the real values. Most of the models generate a considerable amount of early warning signs before the 2007 recession, which indicates their suitability for detecting turning points in business cycles. Moreover, a wide range of the proposed LSTM and GRU network architectures learn the general pattern, also the smaller architectures comprising only one hidden layer and less than 500 optimizable parameters. This suggests that these methods offer noteworthy solutions for business cycle forecasting and, more widely, supports applying nonlinear machine learning methods with multivariate data for macroeconomic forecasting tasks where prevalent methods have been found unable to deliver adequate accuracy.
Source: UTUPub (University of Turku)
Author: Markus M. Mäkelä
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W3093480170