Huseyin Ince, Ali Fehim Cebeci, and Salih Zeki Imamoglu (2019)
An Artificial Neural Network-Based Approach to the Monetary Model of Exchange Rate
This paper contributed by Ince, Cebeci and Zeki (2019) has formed a thorough investigation into how exchange rates can be forecasted based on various strands of model settings. Compared to the most widely applied approaches such as random walk model and linear static framework, the proposed nonparametric and nonlinear model, which combines the vector autoregressive regression and Multilayer Neural Networks, outperformed the previous models in its forecasting power. The VAR model and its transformations are based upon the monetary model on how exchange rates are determined, in which the relative income level, money supply and the interests rates are taken into account. The Neural Network model experimented in this article is simply built upon the multilayer perceptron framework. By varying the forecasting error measurement, like the RMSE, and the objective value, such as the mean return and the Sharp ratio, the researchers found the performance is robust to all these scenarios.
The researchers have collected data “comprising 215 monthly observations from January 1998 through October 2015”, which include “six different exchange rate series, monetary aggregates, interest rates, and industrial production indexes of four countries”. And different instances of interest are dissected, like exchange rated between different countries over various horizons. However, the models in fact were not really combined together and in fact were run in parallel. I am interested in extending the work to more specifications of model settings, and will try to develop more convincing workflows that can chain these models together, which might enable us to make even better forecasts of exchange rates.
First of all, the VAR models are good at capturing long term structural relationships between various variables, and the models utter consistent predictions that are determined by these endogenous variables simultaneously. The process is often characterized by the mean-reverting behavior, and its capability to make dynamic predictions over longer horizons will decay substantially as we attempt to predict further. In contrast, the Multilayer Neural Networks model is outstanding in capturing complex system dynamics, which can probably compensate for what the VAR model might fail to look after. The two models can work hand in hand, for example by feeding the neural network with the residuals from the VAR model. The Neural Networks model can even be chained into a cascade and be learned recursively using the backward propagation. This allows more room for exploring prediction approaches with longer horizons dynamically. It is also worthwhile to try out expansions on feature candidates. Especially kernel tricks can be played with that leads to investigation into higher dimensions of feature space, which might offer some untapped predicting factors.
Except the Neural Network model, the GARCH model and its extensions can also serve to predict the exchange rates. These models are good at predicting conditional volatility, and can adjust for asymmetric impact of changes in exchange rates, which might be beneficial to raise the prediction accuracy.
Huseyin Ince, Ali Fehim Cebeci, and Salih Zeki Imamoglu (2019) An Artificial Neural Network-Based Approach to the Monetary Model of Exchange Rate
This paper contributed by Ince, Cebeci and Zeki (2019) has formed a thorough investigation into how exchange rates can be forecasted based on various strands of model settings. Compared to the most widely applied approaches such as random walk model and linear static framework, the proposed nonparametric and nonlinear model, which combines the vector autoregressive regression and Multilayer Neural Networks, outperformed the previous models in its forecasting power. The VAR model and its transformations are based upon the monetary model on how exchange rates are determined, in which the relative income level, money supply and the interests rates are taken into account. The Neural Network model experimented in this article is simply built upon the multilayer perceptron framework. By varying the forecasting error measurement, like the RMSE, and the objective value, such as the mean return and the Sharp ratio, the researchers found the performance is robust to all these scenarios.
The researchers have collected data “comprising 215 monthly observations from January 1998 through October 2015”, which include “six different exchange rate series, monetary aggregates, interest rates, and industrial production indexes of four countries”. And different instances of interest are dissected, like exchange rated between different countries over various horizons. However, the models in fact were not really combined together and in fact were run in parallel. I am interested in extending the work to more specifications of model settings, and will try to develop more convincing workflows that can chain these models together, which might enable us to make even better forecasts of exchange rates.
First of all, the VAR models are good at capturing long term structural relationships between various variables, and the models utter consistent predictions that are determined by these endogenous variables simultaneously. The process is often characterized by the mean-reverting behavior, and its capability to make dynamic predictions over longer horizons will decay substantially as we attempt to predict further. In contrast, the Multilayer Neural Networks model is outstanding in capturing complex system dynamics, which can probably compensate for what the VAR model might fail to look after. The two models can work hand in hand, for example by feeding the neural network with the residuals from the VAR model. The Neural Networks model can even be chained into a cascade and be learned recursively using the backward propagation. This allows more room for exploring prediction approaches with longer horizons dynamically. It is also worthwhile to try out expansions on feature candidates. Especially kernel tricks can be played with that leads to investigation into higher dimensions of feature space, which might offer some untapped predicting factors.
Except the Neural Network model, the GARCH model and its extensions can also serve to predict the exchange rates. These models are good at predicting conditional volatility, and can adjust for asymmetric impact of changes in exchange rates, which might be beneficial to raise the prediction accuracy.
Reference: Ince, H., Cebeci, A.F. & Imamoglu, S.Z. Comput Econ (2019) 53: 817. https://doi.org/10.1007/s10614-017-9765-6**