An Artificial Neural Network-Based Approach to the Monetary Model of Exchange Rate

This paper aims to investigate the predictive accuracy of the flexible price monetary model of the exchange rate, estimated by an approach based on combining the vector autoregressive model and multilayer feedforward neural networks. The forecasting performance of this nonlinear, nonparametric model is analyzed comparatively with a monetary model estimated in a linear static framework; the monetary model estimated in a linear dynamic vector autoregressive framework; the monetary model estimated in a parametric nonlinear dynamic threshold vector autoregressive framework; and the naïve random walk model applied to six different exchange rates over three forecasting periods. The models are compared in terms of both the magnitude of their forecast errors and the economic value of their forecasts. The proposed model yielded promising outcomes by performing better than the random walk model in 16 out of 18 instances in terms of the root mean square error and 15 out of 18 instances in terms of mean return and Sharpe ratio. The model also performed better than linear models in 17 out of 18 instances for root mean square error and 14 out of 18 instances for mean returns and Sharpe ratio. The distinguishing feature of the proposed model versus the present models in the literature is its robustness to outperform the random walk model, regardless of whether the magnitude of forecast errors or the economic value of the forecasts is chosen as a performance measure.