Jordan neural network for inflation forecasting

Abstract

In times of pronounced nonlinearity in macroeconomic variables and in situations when variables are not normally distributed, i.e. when the assumption of i.i.d. is not fulfilled, neural networks (NNs) should be used for forecasting. In this paper Jordan neural network (JNN), a special type of NNs is examined, which is because of its advantages in time series forecasting suitable for inflation forecasting. The variables used as inputs include labour market variable, financial variable, external factor and lagged inflation, i.e. the most commonly used variables in previous research. The research is conducted at the aggregate level of euro area countries in the period from January 1999 to January 2017. Based on 250 estimated JNNs, which differ in selected variables, sample breaking point and varying parameters (number of hidden neurons, weight value of the context unit), the model adequacy indicators for each JNN are calculated for two periods: “in-the-sample” and “out-of-sample”. Finally, the optimal JNN for inflation forecasting is obtained as the best compromise solution between low mean squared error “in-the-sample” and “out-of-sample” and low number of parameters to estimate. This paper contributes to existing literature in using JNN for inflation forecasting since it is rarely used for macroeconomic time series prediction in general. Moreover, this paper defines which set of variables contributes to the best inflation forecast. Additionally, JNN is examined thoroughly by fixing certain parameters of the model and alternating the other parameters to contribute to the JNN literature, i.e. finding the optimal JNN.