A Hybrid Forecasting Model for Foreign Exchange Rate Based on a Multi-neural Network

In this work, a multi-neural network model consisting of three sub-networks and one master network is proposed to combine the fundamental theorem and technical analysis in TWD/USD exchange rate forecasting. The long-term, mid-term, and short-term tendencies of exchange rate are forecasted separately by different sub-networks. Five macro economics factors of price level, interest rates, money supply, imports/exports, and productivity, and seven practical technical indicators of fifteen-day and one-day intervals are selected as the input variables of the three sub-networks. The master network then provides the integrated forecasting according to the three sub-networks. To increase forecasting accuracy, a threshold filtering mechanism was applied in this work. The experiment result shows that the multi-neural network is more effective than the random walk model and single-neural network model, and with the threshold filtering, can achieve high accuracy.