International Journal of Engineering & Technology, 7 (4.

15) (2018) 400-404

International Journal of Engineering & Technology

Website: www.sciencepubco.com/index.php/IJET

Research paper

FoRex Trading Using Supervised Machine Learning

Thuy Nguyen Thi Thu1*, Vuong Dang Xuan2
1
ThuongMai University, Mai Dich, Cau Giay, Hanoi, Vietnam
2
Sonla Gifted HighSchool, Sonla, Vietnam
*Corresponding author E-mail: [email protected]

Abstract
The exchange rate of each money pair can be predicted by using machine learning algorithm during classification process. With the help
of supervised machine learning model, the predicted uptrend or downtrend of FoRex rate might help traders to have right decision on
FoRex transactions. The installation of machine learning algorithms in the FoRex trading online market can automatically make the
transactions of buying/selling. All the transactions in the experiment are performed by using scripts added-on in transaction application.
The capital, profits results of use support vector machine (SVM) models are higher than the normal one (without use of SVM).
Keywords: Classification; Foreign Exchange rate; Supervised Machine Learning; Transaction.

1. Introduction FoRex rate in the trading market. These specific terms are the
basic explanations for the classification output results.
Supervised machine learning can apply into many areas in com- To be more generality, a framework of supervised machine learn-
puter science such as decision making, forecasting, and specially ing (SML) techniques in predicting FoRex rate has been shown in
is to predict stock price or money exchange rate, and so on. Su- section 4. Based on this framework, a representative SVM model,
pervised machine learning means the teaching and monitoring the which is proposed from research experiment in [27], is chosen to
computer to classify or cluster data from observed data set. With install into expert advisor (robotics) in FoRex transaction software
the explosion of data nowadays, particular with business big data, to make the actual FoRex transactions with demo account.
supervised machine learning techniques can support people to
look depth and analyze their data in order to extract useful infor-
mation for their company’s purposes. In stock market, the transac-
2. Related Works
tions of money exchange can be predicted rise or fall with helping
of machine learning and the observed data set collected in the past. The Foreign Exchange problems can be seen as the ones of pre-
The artificial intelligent systems are used with supervised machine dicting FoRex rates (up or down) of each currency pairs. There-
learning methods like Logic, Perceptron, Statistics methods fore, they can be seen as classification problems where their out-
(Bayes Networks, Instance based) [2]. The goals are the classifica- puts (FoRex rates) are the binary values. There are some previous
tion or regression dataset into alternative classes so-called as out- techniques such as Auto Regressive Integrated Moving Average
puts. These target variable (y) is usually the nominal (discrete) or (ARIMA) used for predicting time series data. According to [6],
continuous values [1]. For simplicity, the output of y in the classi- ARIMA, however, is an unvaried model in general. Moreover, the
fication problem can be seen as binary values such as 0 or 1. techniques are performed with the assumption of the linear and
stationary time series.
Support Vector Machines (SVMs), which is one of Supervised
Machine Learning techniques, are the techniques that separate According to [7], the noisy date in the FoRex prediction problems
data in hyperplane space into two data classes [3]. These tech- made them to be challenged issues in time series forecasting field.
niques create maximized margins and the distance between them Some publications in [7], [8], [6], [9] used RBF, and MLP, and
to separate data into alternative sides of hyperplane [2]. So, SVMs gene to predict the FoRex rates.
can be a possible candidate of use to predict Foreign Exchange The works, for instance, in [6] used Multilayer Perceptron Net-
(Forex) trend (Up or Down) in money exchange rate problems. work (MLP) with the training set, testing set of 1600 and 225 over
The predicting FoRex price is performed by using historical data 1825 instances. The disadvantage in the work is a use of small
such as Open", "Close", "Low" and "High" in different time rang- data set. This might cause the quick regression or mis-predicting
es of FoRex transactions. output result. Another work of Rumani stock with MLP models
Alternative experiments with different support vector machine can be seen in [10]. The publication works shown the advantages
models are performed in this paper. These are to show the ad- in use supervised learning models in forecast a value of rise or fall
vantage of using them in FoRex rate prediction. The detail re- for stock price.
search on SVM in FoRex market can be seen in [27]. The analysis There are some ANN and decision tree models which have been
of FoRex rate prediction can be seen in section 3. This shows used with forecast the uptrend or downtrend in stock market [11-
some specific concepts which contribute to the predicted trend of 14]. Although these works reported about their abilities to work

Copyright © 2018 Authors. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted
use, distribution, and reproduction in any medium, provided the original work is properly cited.
International Journal of Engineering & Technology 401

with real time data, however, their experiments have reported of that a given exchange rate can make. Every money pair has a cor-
lacking of using high frequency data [15]. responding value (a quote). For example, a quote of 1.2776 is the
exchange rate of every EUR is USD 1.2776 (it costs the trader of
The nonlinear regression problems, especially in predicting things, 1.2776 to buy a single EUR). If the rate is lightly moved from
can be solved with SVMs models. There are some publications 1.2776 to 1.2777, it shows 1 PIP increase.
which have been successful in use nonlinear regression predictions
such as in 16], [17] [23], [24], [25]. However, there are limited The PIP is calculated as:
publications used SVMs models for financial time-series forecast-
ing [19], [20] although the SVMs models can perform with non- 𝑃𝐼𝑃𝑣𝑎𝑙𝑢𝑒 = 𝑃𝐼𝑃/𝑟𝑎𝑡𝑒 × 𝐴𝑚𝑜𝑢𝑛𝑡 𝑏𝑒𝑖𝑛𝑔 𝑝𝑢𝑟𝑐ℎ𝑎𝑠𝑒𝑑
linear problems and they have been reported as the advantage For example, if a trader purchase 10,000 units with the rate of
techniques. Therefore, continue to the research in [27], the repre- EUR/USD of 1.2776. Therefore, PIPvalue= 0.0001/1.2776 x 10,000
sentative supervised machine learning method of SVM is selected = 0.7827 EUR.
to estimate the uptrend or downtrend of each currency rates. It is
because of SVMs can avoid some issues such as over-fitting prob- The concept of the transaction trend describes the direction of the
lems (like ANN in [1]), or no parameters to be harmony except the market moving as uptrend or downtrend. In the uptrend, a re-
upper bound C [21]. sistance level means there is a change in supply and demand
(more sellers than buyers) whereas in a downtrend, there is a con-
cept of support level, in which there are more buyers than sellers.
3. Research Background When the trend is ended, there are two possible outcomes of oppo-
site starting trend or in range market (stop at certain resistance or
3.1. Supervised Support Vector Machine support levels). The task for the Expert Advisor (in the experiment
in later section) is to determine the trend, calculate the time of
The Support Vector Machines (SVMs) can be divided in to linear lasting and decide if the timing is right to place a good transaction
or nonlinear SVMs. These are based on linear or nonlinear hyper- of Bid/Ask.
plane lines which separate the training data set in hyperplane
space. The linear form in linear hyperplane in explained as follow-
ing:
𝐻: 𝑤. 𝑥 + 𝑏 = 0

Where w = (w1 , w2 , … , wd ) ∈ ℝd and x = (x1 , x2 , … , xd ) ∈

ℝd then w. x = ∑di=1 wi . xi
Assume that 𝑦𝑖 ∈ {+1, −1} , 𝑖 = ̅̅̅̅̅
1, 𝑛 . 𝑦𝑖 is labelled as output
class. Therefore, outputs can be formed as:
1 𝑖𝑓 𝑤. 𝑥𝑖 + 𝑏 ≥ 0 Fig.1: Illustration of Uptrend or Downtrend
𝑦𝑖 = {
−1 𝑖𝑓 𝑤. 𝑥𝑖 + 𝑏 < 0
To predict target of exchange rate as uptrend or downtrend, the
The hyperplane H need separate the data onto alternative classes. following parameters are defined.
So, SVMs have to build two hyperplane as: 𝐻1 : 𝑤. 𝑥 + 𝑏 = 1 ;  Target Profit (TP) shows the target profit (positive num-
and 𝐻2 : 𝑤. 𝑥 + 𝑏 = −1 with no existing 𝑥𝑖 in 𝐻1 and 𝐻2 , & ber).
margin of 𝐻1 and 𝐻2 is max.
 Stop Loss (SL) indicates the acceptable loss (positive
In some cases, the SVM need to use one special margin so-called number).
as soft margin to separate the data set into two classes including
some mis-classified training instances [4]. This is because SVM  Time Trend Detection (T) indicates the time starting
can not find any separating hyperplane which can separate exactly from present stage (positive number).
the data set. Hence, current rate (P), the trend of rate can be as:
It seems to be difficult to solve with nonlinear training dataset.  up if the rate of (P+TP) is increased, and there is a de-
The solution is that the data is mapped onto a higher dimensional
creasing of an amount of (P-SL). All these are calculated
space so-called as feature space in where the linear hyperplane can
in the time T.
separate the data set in to two classes.
 Low when the (P-TP) is decreased, and an amount of
By using kernel function, any new points in feature space are rep-
P+SL is increased. All these are calculated in the time T.
resented for the data in the given data set [5].
Assume that 𝐸(𝑥, 𝑦) is predicting error rate for each pair of (𝑥, 𝑦), 4. Supervised Machine Learning Prediction
selection of 𝛼𝑖 , 𝛼𝑗 will be produced in two heuristic loops so that
|𝐸𝑖 − 𝐸𝑗 | will max.
Framework
The supervised machine learning prediction frame work can be
3.2. Foreign Exchange Rate Prediction
seen in Figure 2, which is based on data mining process.
The FoRex market has its own characteristics of high benefit. In
there, the transaction can be performed online via the internet
services. This means that the investors can request their transac-
tions at any time and from anywhere in the world. The buying and
selling can be done with any currencies pairs such as EUR/USD,
EUR/JPY, or GBP/EUR etc. in order to make profit.
Fig.2: Supervised machine learning prediction framework
A term of PIP (Price Interest Point) is defined in order to measure
the profit or loss in the FoRex. A PIP is the smallest price change The detail as follows:
402 International Journal of Engineering & Technology

Step 1: Feature Selection. Collected data set might include un- #166-
Bollinger Bands 9 3 Lines x 3 Timeframe
necessary features for predicting process. Therefore, these features 174
are removed. #175- 2 values x 3
Average True Range 6
180 Timeframe
Step 2: Select dataset of currency pair. In this step, a defined pair #181- 4 values x 4
Highest, Lowest M1, M5 16
of money eg. EUR vs USD is chosen. Also in this step, the data is 196 Timeframe
taken from alternative candles, time frames etc. in order to be
easily used in installation with Robotics. Model configurations
Step 3: Processing data with supervised machine learning (SML) The SVMs models are chosen in the experiments. The models
models (here is support vector machine models). Dataset is divid- parameters are used the same as using in [27]. Therefore, the
ed into training, testing and validation sets. All data in mapped on model parameters are:
the range of [0,1].
‖x−y‖2 2
Step 4: Evaluation models with assessment metrics of mean  RBF: k(x, y) = e− 2σ = e−γ‖x−y‖ with
square errors (MSE). γ ∈ [0; 5]

Step 3,4 is repeated many time in order to choose a model with  Polynomial: k(x, y) = (x. y + θ)d, θ ∈ ℝ, d ∈ ℕ∗
minimum MSE. with d = 2,3,4 and θ ∈ [0; 1].

The output is a certain definition of trend (up or down).  The parameter of C has been used with C ∈ [1; 10].
Table 2: Experimental Model Configurations
5. Experiment Models Poly Poly Poly GsRBF GsRBF GSRBF
1 2 3 1 2 3
5.1. Data analysis and Model configuration C 1.0 1.0 1.0 2.0 1.0 2.0
Kernel Poly Poly Poly RBF RBF RBF
Parame- Pow- Pow- Pow- =5.0 =2.0 =2.0
The experimental data is collected from 01/01/2013 to 30/09/2016 er 2 er 3 er 3
ters
by using MetaTrader 4 [22]. For general use, the pair of No of 2500 3000 3972 3000 4500 5000
EUR/USD has been chosen. Two data sets are created with the vectors
one of training of D (including the instances in a time of
01/01/2013 to 31/12/2015), and one of testing D’ (including data 5.2. Classification results
in a period 1/01/2016 to 30/09/2016).
The features are chosen basing on the profits. In this paper, the Data set D data can be separated to two parts Dpos and Dneg posi-
profit is taken of bout 10 – 15 Pips in each transaction. These se- tive and negative output respectively. Which can be further split to
1 2 𝑘 1 2 𝑘
lections might effect to the experimental results. The input fea- 𝐷𝑝𝑜𝑠 , 𝐷𝑝𝑜𝑠 , … , 𝐷𝑝𝑜𝑠 and 𝐷𝑛𝑒𝑔 , 𝐷𝑛𝑒𝑔 , … , 𝐷𝑛𝑒𝑔 sub-sets with each
tures for samples exchange rates "Open", "Close", "Low" and 1
sub-sets has 𝑘 |𝐷𝑝𝑜𝑠/𝑛𝑒𝑔 | samples. D_Test set is taken with each
"High". These rates are taken in alternative candles in the window
sub-set. D_Train includes k-1 remaining samples in the sub-sets.
time (time frame). There are 1 minute, 5 minutes, 15 minutes,1
hour, 1 day according to M1, M5, M15, H1, D1 respectively. Oth- In here, k is 5.
er features are taken from Custom Indicators, Bollinger Bands, Following Table 3 & 4 show experimental results with 6 alterna-
RSI and MA. tive models. The performance between alternative models.is eval-
To continue with the experiment in [27], 196 features are chosen uated and compared by using the Accuracy Exactness Rate, Preci-
sion of Positive, Negative, Micro-Average and Macro-Average
(in Table 1). The addition more features (increase from 137 to 196)
individually.
for the classification is for expecting to have better results in the
classification process and having better transactions. This means Table 3: Results of RBF Configuration Model
data is taken from an extension of candlesticks and timeframe. Kernel: Gaussian RBF GsRBF1 GsRBF2 GSRBF3
This is to concentrate more detail in specific historical transaction Features: 196 D_Trai D_Tes D_Trai D_Tes D_Trai D_Tes
data. n t n t n t
Accurate 55.80 58.14 58.17
84.22% 83.14% 83.81%
Table 1: Feature selections % % %
Fea- N Positive 55.59 57.55 57.56
Data Explanations 84.05% 82.59% 83.12%
ture# o % % %
04 value X 5 nearest Negative 55.98 58.70 58.76
#1-20 O , H , L , C on M1 20 84.39% 83.71% 84.52%
candles Precisio- % % %
04 value X 5 nearest n Micro- 55.80 58.14 58.17
#21-40 O , H , L , C on M5 20 84.22% 83.14% 83.81%
candles Average % % %
04 value X 4 nearest Macro- 55.79 58.13 58.16
#41-56 O , H , L , C on M15 16 84.22% 83.15% 83.82%
candles Average % % %
04 value X 4 nearest
#57-72 O , H , L , C on H1 16 Table 4: Poly Configuration Models Results
candles
4 values x 5 nearest Kernel: Polynomial Poly1 Poly2 Poly3
#73-92 O, H, L, C on D1 20 Features:196 D_Train D_Test D_Train D_Test D_Train D_Test
candles
#93-97 Time data 5 Time attributes Accurate 86.15% 74.20% 86.15% 74.35% 86.25% 74.32%
#98- 2 values x 4 Positive 86.73% 80.29% 86.58% 80.18% 86.80% 80.21%
RSI(7) on M5, M15 8 Negative 85.59% 70.36% 85.73% 70.62% 85.72% 70.56%
105 Timeframe Precisio-n
Micro-Average 86.15% 74.20% 86.15% 74.35% 86.25% 74.32%
#106- 2 values x 4
RSI(14) on M5, M15 8 Macro-Average 86.16% 75.32% 86.16% 75.40% 86.26% 75.38%
113 Timeframe
#114- MA(9), MA(12), MA(100), 8 values x 4
32 5.3. Expert Advisor experimental results
145 MA(200) Timeframe
#146-
Custom Indicator 20 8 PAX + 12 MKC
165 The Poly2 SVM model is chosen for installation into Robotics
(Expert Advisor- EA) in MetaTrader 4 [30]. The script (written in
International Journal of Engineering & Technology 403

EA) has performed with a collection of observed data set. The Robotics’ transactions. In other words, more features are used in
information incorporates some specialized markers and exchanges SVM, the profit loss can be better control.
of purchase or offer FoRex match of USD and EUR.
Drawdown Comparison
For correlation with the utilization directed machine learning
60.00% 50.67%
model in Robotics, the typical exchange (test of Robotics without
utilizing backing of SVM) additionally has been performed. 50.00%
34.69% 34.50%
40.00%
The results can be seen in Table 5. The bold numbers in the table 30.00%
24.31%
show the comparison between the transactions use or not use
20.00%
SVM.
10.00%

Table 5: Results Indicators’. 0.00%

In- Indicators No_SVM SVM 137_NoSVM 137_SVM 196_NoSVM 196_SVM
dex
1 Net Profit 328.73 462.5 Fig. 4: Drawdown rate comparison between 137 features model [27] and
2 Gross Profits 1238.6 713.93 196 features model.
3 Gross Loss - -
909.87 251.44 6. Conclusion
4 Profit rate 1.36 2.84
5 Growth 32.9% 46.3% The support vector machine can help to forecast FoRex trend of up
6 Average Growth (in a 3.21% 4.31% or down. This is because the outputs of FoRex rate are received as
month) binary values (uptrend or downtrend). The FoRex problems now
7 Strict Drawdown 155.07 17.49
can be seen as classification ones which can be solved by supervised
8 Fair Drawdown 436.2 34.05 315.6 24.31 learning particular with SVMs models. The experiment has shown
3 % 2 %
the power of SVMs in performing a huge, complex and time series
9 Volume 188 94
data in order to help the investors with their transactions decisions.
10 Profit Transactions 139 73.9% 76 80.9%
11 Loss Transactions 49 26.1% 18 19.2% The experiment has been performed with alternative kernel func-
12 Ask Profit Transactions 86 95.4% 49 100.0 tions (Gaussian and Polynomial). Elective models with elective
% parameters additionally have been per-framed to win the higher
13 Bid Profit Transactions 102 55.9% 45 60.0% outcomes. The consequences of utilizing Polynomial Kernel
demonstrated the best outcomes contrasted with models utilized
5.4. Discussion Gaussian when all is said in done.
The picked model to introduce in Robotic exchange has a place with
Alternative models results with different kernel functions (Gaussi-
the best Poly show in above analyses. Its exchange results can ac-
an RBF, Polynomial) can be seen in Table 3 and Table 4. There
quire a larger number of benefits than the conventional one (without
are big gap results of test set compared to the results of training set
use of SVM). Therefore, an increase number of features (from 137
(see in table 3, about 57% and 83.8 % respectively). The ad-
to 196) compared to [27] has shown the better profit lost models can
vantage use of SVM in the transaction Robotics can be seen in
get (reduce the drawdown rate). This means that although there has
Table 5. The indicators’ results show that the net profit of using
been no different in classification accuracy rates (Figure 3), but the
SVM in add-on robot is higher than the one without using SVM.
more concentrating on specific data the better controlling of loss we
For instances, the number of samples for the script using support
can get.
vector machine is half of number of SVM (94 & 188) and its prof-
it rate can achieve of 2.84 compared to 1.36.
Accuracy
Acknowledgement
90.00%

80.00% Many thanks to ThuongMai University and research group for

D_train_137 supporting the research.
70.00%
D_Test_137
60.00% D_train_196 References
50.00% D_Test_196
[1] Witten IH, Frank E, Hall MA (2011), Data mining: practical
machine learning tools and techniques, Morgan Kaufmann
Publishers.
[2] Kotsiantis SB (2007), Supervised Machine Learning: A Review of
Fig 3: Comparison between 196 features-models and 137 features-models Classification Techniques, Informatica: 31, 249-268.
in [27]. [3] Vapnik V (1995), The Nature of Statistical Learning Theory,
Springer Verlag.
The accuracy rates of models using with 137 features in [27] are [4] Veropoulos K, Campbell C, Cristianini N (1999), Controlling the
Sensitivity of Support Vector Machines, In Proceedings of the
used to compare to the ones using with 196 features in this re-
International Joint Conference on Artificial Intelligence,
search (see in Figure 3). In general, there are not big gap of rates (IJCAI99).
between these models. This means the increase of features does [5] Scholkopf C, Burges JC, Smola AJ (1999), Advances in Kernel
not have better improvement in the accuracy rate. Methods, MIT Press.
[6] Baasher Aa, Mohamed WF (2010), FOREX Trend Classification
For more comparison, the fair drawdown values, which shows the using Machine Learning Techniques, Arab Academy for Science
rate of capital reduction, taken from [27] and this research are and Technology, 41-47.
represented on the chart (Figure 4). [7] Deboeck GJ (1994), Trading on the Edge: Neural, Genetic and
Fuzzy Systems for Chaotic Financial Markets, New York Wiley.
It is clear from Figure 4 that the Robotics using SVM with 196 [8] Box GEP, Jenkins GM (1970), Time Series Analysis: Forecasting
features can prevent the capital reduction compared to the normal and Control, Holden- Day, San Francosco, CA.
404 International Journal of Engineering & Technology

[9] Punniyamoorthy M, Thoppan JJ (2013), ANN-GA based model for

stock market surveillance, Journal of Financial Crime, vol. 20, no.
1, 52-66.
[10] Nemeş Dm, Butoi A (2013), Data Mining on Romanian Stock
Market Using Neural Networks for Price Prediction, Informatica
Economică, vol. 17, no. 3,125-136.
[11] Huang W, Nakamori Y, Wang SY (2004), Forecasting foreign
exchange rates with artificial neural networks, a review.
International Journal of Information Technology & Decision
Making, Vol 3, No 1, 145-165.
[12] Huang CL, Tsai CY (2010), A hybrid SOFM-SVR with a filter
based feature selection for stock market forecasting, Expert Systems
with Applications, 36 (2, Part 1), 1529 – 1539.
[13] Tsai CF, Hsiao YC (2010), Combining multiple feature selection
methods for stock prediction: Union, intersection, and multi-
intersection approaches, Decision Support Systems, 50(1), 258 –
269.
[14] Shen W, Guo X, Wu C, Wu D (2011), Forecasting stock indices
using radial basis function neural networks optimized by artificial
fish swarm algorithm, Knowledge-Based Systems, 24(3), 378–385.
[15] Reboredo M, Rubio JJ, Nonlinearity in forecasting of high-
frequency stock returns, Computational Economics, 40(3), 245–
264.
[16] Müller KR, Smola A, Rätsch G, Schölkopf B, Kohlmorgen J, and
Vapnik V, (1997), Predicting time series with support vector
machines, In proceedings of the IEEE workshop on neural networks
for signal processing 7, 511 – 519.
[17] Mukherjee S, Osuna E, Giroso F (1997), Nonlinear prediction of
chaotic time series using SVM. In proceedings of the IEEE
workshop on neural networks for signal processing 7, 511 – 519.
[18] Kim K (2003), Financial time series forecasting using Support
Vector Machines, Neurocomputing 55, 307- 319.
[19] Tay EFH, Cao L (2001), Application of support vector machines in
Financial time series forecasting, Omega 29, 309–317.
[20] Liu L, Wang W (2008), Exchange Rates Forecastiing with Least
Squares SVM, International Conference on Computer Science and
Software Engineering, 2008.
[21] Drucker H, Wu D, Vapnik VN, Support vector machines for spam
categorization, IEEE Trans. Neural Networks 10 (5),1048–1054
[22] Metatrader 4, (2016). Website: http://www.metaquotes.net/
[23] Kamruzzaman J, Sarker RA, Ahmad I (2003), SVM Based Models
for Predicting Foreign Currency Exchange Rates, Third IEEE
International Conference on Data Mining, ICDM 2003, 557 – 560.
[24] Sidehabi SW, Indrabayu, Tandungan S (2016), Statistical and
Machine Learning approach in forex prediction based on empirical
data, IEEE International Conference on Computational Intelligence
and Cybernetics, CYBERNETICSCOM 2016, 63 – 68.
[25] Sharma DK, Hota HS, Handa R (2017), Prediction of foreign
exchange rate using regression techniques” Review of Business and
Technology Research, Vol. 14, No. 1, 29-33.
[26] Ramakrishnan SK, Butt S, Chohan MA, Ahmad H (2017),
Forecasting Malaysian exchange rate using machine learning
techniques based on commodities prices, IEEE International
Conference on Research and Innovation in Information Systems,
ICRIIS 2017.
[27] Thuy NTT, Vuong DX (2018), Using Support Vector Machine in
FoRex Predicting, IEEE International Conference on Innovative
research and development, ICIRD 2018.