Forecast foreign exchange with both linear and non-linear models coupled with trading rules for selected currencies

Document Type

Conference Proceeding


Modelling and Simulation Society of Australia and New Zealand

Place of Publication



Weber, T., McPhee, M.J. and Anderssen, R.S.


School of Business and Law




Originally published as: Ling, L., Tsui, Al., & Zhang, Z. (2015). Forecast foreign exchange with both linear and non-linear models coupled with trading rules for selected currencies. In Weber, T., McPhee, M.J. and Anderssen, R.S. (eds) MODSIM2015, 21st International Congress on Modelling and Simulation. Modelling and Simulation Society of Australia and New Zealand, December 2015, (pp. 490–496). http://www.mssanz.org.au/modsim2015/E7/ling.pdf. Available here.


The importance of forecasting exchange rate is evident both academically and practically, but it is not an easy task to perform as the foreign exchange market has long been considered complex, erratic and exhibits apparently random behavior. The challenge is posited in a number of studies that highlight the poor out-of-sample forecasting performance of a variety of structural exchange rate models and conclude that none of these models could significantly outperform a simple random-walk model in both short- and medium-terms. An extensive subsequent literature using non-linear econometric techniques, different currencies, data periodicity and samples also draw similar conclusion that exchange rates, just like other financial time series, can be well modeled using a random walk model.

In this paper we attempt to employ a “hybrid” model to investigate the effectiveness of monetary fundamentals and other macroeconomic variables in predicting the bull and bear market longer-term trends (macro-cycles) and in forecasting the exchange rate movements. In particular, we intend to use a combined model of both parametric Markov Logistic model and a nonparametric multilayer feedforward neural network coupled with technical trading rules to predict the macro-cycles of the selected currencies by using the macroeconomic fundamental variables as inputs. When applying the linear models, most existing studies seem to use the same specification for estimation and forecasting, but the dynamic impact of the concerned variables is ignored. In this study we allow for variations in model specification throughout the forecasting period to address this stylized fact, and, furthermore, we combine the linear model and nonlinear neural network model by adopting both an equal weighted approach and a profit weighted approach to capture both the linear and nonlinear components of the exchange rate mechanism. It is expected that the combined hybrid models will outperform those single models in terms of predicting power and trading advantage in different market condition. We choose three pairs of currencies including the US dollar (USD), the Japanese yen (JPY) and Canadian dollar (CAD) in this study. The USD and JPY are one of the mostly traded currencies in the world, and the Canadian dollar is chosen because of its close economic ties to the United States. The bilateral exchange rate of CAD and JPY is studied as it is interesting to see if our model works for the less traded currencies, and also to complete the “triangle” of the three currencies.

The results confirm that the combination models have a significant predictive and market timing ability and outperform the benchmark models in terms of returns, even although their advantage diminishes in the periods of central bank intervention.