LettersNeural network protocols and model performance
Introduction
Ever since the seminal paper of Meese and Rogoff [14] it has been generally accepted that it is hard to beat a random walk in predicting exchange rates out-of-sample, possibly because there is no coherent underlying economic theory of short-to-medium-run movements in exchange rates [13]. Subsequent studies including [19] provide support for some of the long-run relationships implied by economic theory, but suggest further attempts to offer explanations of short-term exchange rate movements based only on macroeconomic fundamentals may not prove successful.
However, motivated by the growing use of technical analysis in currency trading [2] and recent nonlinear modeling of exchange rates using neural networks [6], [12], we build univariate neural network prediction models using time-delay embedding technique to forecast short-run exchange rate movements. In doing so, common procedures such as normalisation and out-of-sample validation were omitted to reduce problems of data snooping and only model architectures approximating accepted rules of thumb were considered. Even so, the results showed significant outperformance of accepted benchmarks and interestingly of previous comparable studies for all four exchange series.
Section snippets
Data and neural network models
Daily dollar rates for the Japanese Yen (JPY), British Pound (GBP), Deutsche Mark (DEM) and Swiss Franc (CHF) from January 2, 1986 to November 11, 1999 (3616 observations for each currency) were obtained from the Datastream/ICV International Database and converted to returns by taking the first difference of log rates. The third moment of return series distribution, skewness, is statistically significant at the 1% level for $/JPY and $/GBP rates indicating deviation from normal distribution.
Results
Normalized mean squared error (NMSE), RMSE and mean absolute error (MAE) along with out-of-sample sign and direction change statistic for network predictions and error ratios relative to the mean value predictor are presented in Table 1. In financial markets, profits may depend more on accuracy in predicting direction of movements than error in predicting and this is captured by the sign of return statisticwhere for N pairs of predictions yt and outcomes xt, at=1 if xtyt>0 or xt=
Concluding remarks
Contrary to earlier findings, our one-step-ahead forecasts of daily exchange rate returns significantly outperform the random walk and mean value predictor benchmark forecasts and the results of comparable previous research. This performance is achieved using robust models which avoid some of the problems of overfitting and information losses through data preprocessing inherent in conventional approaches. In any event the use of parsimonious procedures appears to improve transparency without
Acknowledgements
The authors would like to thank the anonymous reviewers for their helpful comments.
References (21)
- et al.
Currency traders and exchange rate dynamicsa survey of the US market
J. Internat. Money Finance
(2001) On the approximate realization of continuous mappings by neural networks
Neural Networks
(1989)Linear, nonlinear and essential foreign exchange rate prediction with single technical trading rules
J. Internat. Econom.
(1999)- et al.
Testing the equality of prediction mean squared errors
Internat. J. Forecasting
(1997) - et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989) - et al.
Empirical exchange rate models of the seventiesdo they fit out of sample?
J. Internat. Econom.
(1983) - et al.
Forecasting with artificial neural networksthe state of the art
Internat. J. Forecasting
(1998) Neural Networks for Pattern Recognition
(1995)- et al.
Comparing predictive accuracy
J. Business Econom. Statist.
(1995) - et al.
Time series forecasting with neural networksa comparative study using the airline data
Appl. Statist.
(1988)
Cited by (0)
- 1
The results and interpretations in this paper are the author's alone and do not represent the position of Accenture.