Regression neural network for error correction in foreign exchange forecasting and trading

Abstract

Predicting exchange rates has long been a concern in international finance as most standard econometric methods are unable to produce significantly better forecasts than the random walk model. Recent studies provide some evidence for the ability of using multivariate time series models to generate better forecasts. At the same time, artificial neural networks have been emerging as alternatives to predict exchange rates. In this paper, we propose an adaptive forecasting approach which combines the strengths of neural networks and multivariate econometric models. This hybrid approach contains two forecasting stages. In the first stage, a time series model generates estimates of the exchange rates. In the second stage, General Regression Neural Network is used to correct the errors of the estimates. A number of tests and statistical measures are then applied to compare the performances of the two-stage models (with error-correction by neural network) with those of the single-stage models (without error-correction by neural network). Both empirical and trading simulation experiments suggest that the proposed hybrid approach not only produces better exchange rate forecasts but also results in higher investment returns than the single-stage models. The effect of risk aversion in currency trading is also considered.

Introduction

The difficulty in predicting exchange rates has been a long-standing problem in international finance as many standard econometric methods are unable to produce significantly better forecasts than the random walk model. Of the various methods of forecasting exchange rates, multivariate time series models have received much attention in the literature and tend to produce the best out-of-sample forecasts. The literature on multivariate time-series models is quite extensive. Specifically in the area of exchange rate forecasting, Canarella and Pollard [1], Wolff [2], Hoque and Latif [3], Liu et al. [4], and Sarantis and Stewart [5] show that multivariate time-series models have some forecasting strength in predicting exchange rates and are able to produce forecasts that are superior to the random walk model. Additional studies using time series techniques on exchange rates include Lothian [6], Joseph [7], and Trapletti et al. [8]. Interested readers can also refer to the exchange rate forecasting book written by Moosa [9].

In the last decade, with the rapid advancement of computer technologies and growing popularity of artificial intelligence, researchers and practitioners have become more inclined to adopt artificial neural network (ANN) as an alternative method in financial forecasting. Although most ANN models share a common goal of performing functional mapping, different network architectures vary greatly in their ability to handle different types of problems. In this study, we select the General Regression Neural Network (GRNN) to conduct currency exchange rate forecasting. It is because GRNN, due to its rather rigid structure, requires relatively less amount of time and effort on training and on the design of architectural construct (see Masters [10]). Our choice is also partly motivated by the promising results reported in Leung et al. [11]. In their paper, performances of GRNN models are compared with those of the more widely used multilayered feedforward network (MLFN), random walk model, and multivariate transfer function.

While GRNN possesses some strength in functional mapping, its intrinsic mathematical operations render it weaker than most econometric models in predicting values in the state space which extends beyond the training data. On the other hand, most econometric models are capable of extrapolating values outside of the existing/known data range during forecasting process. Given this notion, we propose a two-stage model which combines an econometric model with GRNN to overcome this shortcoming and to create synergies in the overall forecasting process. In this two-stage model, an econometric model is used to generate a forecast in the first stage. In the second stage, a GRNN is applied to correct the estimation error in the first stage forecast. It is hoped that this error correction by the neural network can lead to an improvement in the final forecast.

The fundamental concept of the proposed hybrid approach originates from adaptive forecasting which attempts to uncover possible data patterns not captured by a forecasting model in the first pass. During the second pass in the forecasting process, another forecasting model is applied to pick up potential data patterns hidden in the residual term. A vast majority of the models utilized in this adaptive forecasting paradigm have a rather traditional econometric background. Most of these econometric models are linear in their functional forms and hold under specific parametric assumptions. Our proposed hybrid approach deviates from this tradition as it combines the strengths of a neural network and a multivariate econometric model. Furthermore, the approach takes an unconventional step of applying a nonparametric (neural network) model to learn and thus reduce the errors made by various parametric (econometric) models. This approach can also be roughly viewed as nonweighted forecast combination in a sequential manner.

In the current study, adaptive error correction models are used to forecast one-quarter-ahead exchange rates. Their forecasting performances are then compared with several competing models including the ones solely based on single forecasting techniques. Our investigation of GRNN has been motivated by the neural network's attractive properties discussed earlier. In the next section, we will provide the economic rationale associated with currency exchange rate forecasting. In Section 3, we present a brief review and mathematical foundation of GRNN and those econometric techniques used in this research. Then, the data set and our estimation procedures are explained in Section 4. Statistical performances of the two-stage error correction models are reported in Section 5. This section also includes comparisons with the single-stage econometric and GRNN models (i.e., models used to directly forecast the exchange rates). In Section 6, we describe a simulation experiment designed to measure the profit derived from the trading of various currencies. A background of the trading decision rules and a discussion of the trading experiment results are also given. The paper is then concluded in Section 7.