Adaptive Logic Network, Arima, and Linear Regression Forecasts of International Equity Markets During the Crash of October 1987

Abstract

The purpose of this study is to investigate how Artificial Neural Network forecasts, specifically Adaptive Logic Network (ALN) forecasts, compare to those of linear regression (LR) and ARIMA models when analyzing international stock market movements during the crash of October 1987. The results of this methodological study show: 1) ALN forecasts give, on average, 16.5% mean square error (MSE) improvement over the LR forecasts; 2) the MSE of the ALN models compared to the MSE of the ARIMA models is comparable or smaller, depending on the country; and 3) in four of five cases, ALN forecasts have a lower bias component than do ARIMA forecasts. A Theil decomposition reveals a sizeable systematic error component of MSE in all forecast profiles, suggesting the potential for improving forecast performance with an expanded ALN model, which is not possible with univariate ARIMA models.

JEL Classification: C22, C51, C52, F47