Forecasting exchange rates with the iMLP: New empirical insight on one multi-layer perceptron for interval time series (ITS)

Abstract

The foreign exchange market (FX) is a market for converting the currency of one country into that of another country. Spot exchange rates movements are carefully observed every minute around the world. But governments, banks and multinational companies generally develop decision-making under other frequencies of time like days, weeks, months, quarters, or semesters. Interval time series (ITS) assign an interval of values at every period of time. For example, daily or monthly lows and highs values are key examples of ITS. Several forecasting methods have been developed for ITS. Neural networks have attracted research focused on FX forecasting. The Multi-Layer Perceptron (MLP) with one hidden layer is one of the best networks for forecasting crisp time series. For ITS, the iMLP (interval MLP) was proposed as an extension of the MLP. The number and type of inputs and the number of neurons in the hidden layer (15 is the usual number) are key parameters to rank different architectures of the network. We analyze these hyperparameters in the forecasting performance of the iMLP through the EUR/USD on a low–high daily basis, on different behaviors such as uptrend, downtrend, or sideways; on different accuracy measures, including coverage and efficiency rates, and incorporating other rates such as GBP/USD or AUD/USD. The election of 15 neurons is discarded. Moreover, we compare these iMLP networks with the interval random walk and results are quite promising. Finally, we conclude that in any context of FX, several iMLP networks should be considered which opens new research avenues.