Time-Series Prediction by Answer-in-Weights Neural Network Consisting of Three Fluctuation Estimators

Abstract

In time-series data prediction, periodic and long-term components are often separately handled in a pre-divided form. We have investigated a neural network model that can identify the two parts, simultaneously. The model consists of two parts of the network arranged in the answer-in-weights structure. In this report, we propose to use an answer-in-weights network consisting of three component estimators. To show the effectiveness of the network, we perform computer simulation on economic data. As the result, we confirm the effectiveness of using the answer-in-weights neural network for time-series prediction.