Using Artificial Neural Networks to forecast Exchange Rate, including VAR‐VECM residual analysis and prediction linear combination

• Author: Alejandro Parot,Werner D. Kristjanpoller,Kevin Michell
• Published date: 01 January 2019
• DOI: http://doi.org/10.1002/isaf.1440
• Date: 01 January 2019

RESEARCH ARTICLE

Using Artificial Neural Networks to forecast Exchange Rate,

including VAR‐VECM residual analysis and prediction linear

combination

Alejandro Parot |Kevin Michell |Werner D. Kristjanpoller

Departamento de Industrias, Universidad

Tecnica Federico Santa Maria, Valparaiso,

Chile

Correspondence

Werner D. Kristjanpoller, Universidad Tecnica

Federico Santa Maria, Departamento de

Industrias, Avenida España 1680, Valparaiso,

Chile.

Email: werner.kristjanpoller@usm.cl

Summary

The Euro US Dollar rate is one of the most important exchange rates in the world, mak-

ing the analysis of its behavior fundamental for the global economy and for different

decision‐makers at both the public and private level. Furthermore, given the ma rket effi-

ciency of the EUR/USD exchange rate, being able to predict the rate's future short‐term

variation represents a great challenge. This study proposes a new framework to improve

the forecasting accuracy of EUR/USD exchange rate returns through the use of an Arti-

ficial Neural Network (ANN) together with a Vector Auto Regressive (VAR) model, Vec-

tor Error Corrective model (VECM), and post‐processing. The motivation lies in the

integration of different approaches, which should improve the ability to forecast regard-

ing each separate model. Thisis especially true given thatArtificial Neural Networks are

capable of capturing the short and long‐term non‐linear components of a time series,

which VECM and VAR models are unable to do. Post‐processing seeks to combine

the best forecasts to make one that is better than its components. Model predictive

capacityis compared accordingto the Root Mean Square Error(RMSE) as a loss function

and its significance is analyzed using the Model Confidence Set. The results obtained

show that the proposed framework outperforms the benchmark models, decreasing

the RMSE of the best econometric model by 32.5% and by 19.3% the best hybrid. Thus,

it is determined that forecast post‐processing increases forecasting accuracy.

KEYWORDS

Embedded Models, Artificial Neural Network, Exchange rate return, VAR, VECM, Cointegration

1|INTRODUCTION

Parities or values betweencurrencies are fundamental to the exchange

of goods and services. This is even more true when there is strong

international trade as the exchange rate can generate incentives for

imports or exports. In this context, one of the main global exchange

rates is the Euro US Dollar (EUR/USD). This rate is essential for the

decision‐making of governments, businesses, and people. For this rea-

son, having more accurate predictions has become a challenge; thus,

accurate exchange rate forecasting,both in terms of return and volatil-

ity, is a problem of global importance, especially in the financialsector,

where a poor forecast can cause big losses. Due to financial market

efficiency, it is difficult to make good short‐and long‐term forecasts

(Kilian & Taylor, 2003). In the case of the EUR/USDexchange rate, this

rate has long been the mosttraded exchange rate, making it a very effi-

cient market where abnormalreturns cannot be frequently obtained. In

this study, we aim to forecast next day returns, which is more compli-

cated as basic techniquesare often ineffective, in order to achieve bet-

ter results that consider changes in interestrates. In addition, exchange

rate series are generally non‐linear, chaotic, non‐parametric, and

dynamic (Zhang & Wu, 2009). Therefore, different techniques and

methodologies have been developed to predict the exchange rate in

a better way than the random walk model. Econometric models have

been widely applied for this task; however, they have alsobeen heavily

Received: 9 January 2018 Revised: 22 November 2018 Accepted: 23 November 2018

DOI: 10.1002/isaf.1440

Intell Sys Acc Fin Mgmt. 2019;26:3–15. © 2019 John Wiley & Sons, Ltd.wileyonlinelibrary.com/journal/isaf 3