Deep networks for predicting direction of change in foreign exchange rates

• Author: Svitlana Galeshchuk,Sumitra Mukherjee
• Date: 01 October 2017
• DOI: http://doi.org/10.1002/isaf.1404
• Published date: 01 October 2017

RESEARCH ARTICLE

Deep networks for predicting direction of change in foreign

exchange rates

Svitlana Galeshchuk |Sumitra Mukherjee

College of Engineering and Computing, Nova

Southeastern University, Fort Lauderdale,

Florida, USA

Correspondence

Sumitra Mukherjee, College of Engineering

and Computing, Nova Southeastern

University, Fort Lauderdale, FL, USA.

Email: sumitra@nova.edu

Funding information

Fulbright Faculty Development Program

Summary

Trillions of dollars are traded daily on the foreign exchange (forex) market, making it the largest

financial market in the world. Accurate forecasting of forex rates is a necessary element in any

effective hedging or speculation strategy in the forex market. Time series models and shallow

neural networks provide acceptable point estimates for future rates but are poor at predicting

the direction of change and, hence, are not very useful for supporting profitable trading strate-

gies. Machine learning classifiers trained on input features crafted based on domain knowledge

produce marginally better results. The recent success of deep networks is partially attributable

to their ability to learn abstract features from raw data. This motivates us to investigate the ability

of deep convolution neural networks to predict the direction of change in forex rates. Exchange

rates for the currency pairs EUR/USD, GBP/USD and JPY/USD are used in experiments. Results

demonstrate that trained deep networks achieve satisfactory out‐of‐sample prediction accuracy.

KEYWORDS

deep network, featureengineering, financial prediction,foreign exchange

1|THE PROBLEM

The foreign exchange (forex) market is the largest financial market in

the world, facilitating trades worth trillions of dollars a day (BIS,

2013). Forex rates are quoted in terms of a base‐quote currency pair;

it represents the number of units of quote currency to be exchanged

for each unit of the base currency. Organizations and individuals par-

ticipate in the forex market for speculative purposes and as a hedge

to reduce risks due to adverse currency fluctuations. Regardless of

the objectives, accurate forecasting of forex rate trends is an essential

element of any effective strategy for trading in the forex market (Lukas

and Taylor, 2007).

Econometric models to forecast exchange rates based on funda-

mental analysis have had limited success, especially when the forecast

horizon is less than a year (Meese and Rogoff, 1983). Time series

models produce acceptable point estimates in forex rate prediction

tasks but are poor at predicting the direction in which the rates move.

Machine learning methods, such as shallow artificial neural networks

(ANNs) and support vector machines (SVMs), may be marginally better

at predicting the direction of change, but their success depends criti-

cally on the input features used to train the models. However, this

improvement comes at a considerable cost; obtaining a good set of

features from raw input data may require significant efforts from

domain experts.

Deep neural networks have proven effective for difficult predic-

tion problems in a variety of domains. Their success is attributable to

their ability to learn abstract features from raw data (LeCun et al.,

2015). This motivates us to investigate the effectiveness of deep net-

works to predict the direction of change in forex rates. Our main con-

tribution is in demonstrating that deep convolution neural networks

(CNNs) are significantly better at predicting directions of change in

forex rates than time series models and shallow networks when raw

exchange rate data are used as inputs to the models. Deep networks

also outperform traditional machine learning classifiers, such as shal-

low networks and SVMs, that are trained on derived features. We

use the exchange rates between the US dollar and three major curren-

cies: Euro, British pound, and Japanese yen in our computational

experiments.

Our prediction problem may be formalized as follows. Let y

t

and

y

t+k

denote the values of an exchange rate between a pair of currencies

in periods tand t+krespectively for some k> 0. Define the direction of

change z

k

(t) = 1 if the rate increases in kperiods; that is, if y

t+k

−y

t

>0;

otherwise, z

k

(t) = 0. Our goal is to learn a function f

k

:ℝ

p

→{0, 1} such

that f

k

(y

t

,y

t−1

,…,y

t−p+1

)=z

k

(t). We train models to predict the

Received: 6 April 2016 Revised: 14 December 2016 Accepted: 9 February 2017

DOI: 10.1002/isaf.1404

100 Copyright © 2017 John Wiley & Sons, Ltd. Intell Sys Acc Fin Mgmt. 2017;24:100–110.wileyonlinelibrary.com/journal/isaf