=Paper=
{{Paper
|id=Vol-2718/paper27
|storemode=property
|title=Search for Optimization of a Real-time Neural Trading System
|pdfUrl=https://ceur-ws.org/Vol-2718/paper27.pdf
|volume=Vol-2718
|authors=Michal Pobucký
|dblpUrl=https://dblp.org/rec/conf/itat/Pobucky20
}}
==Search for Optimization of a Real-time Neural Trading System==
https://ceur-ws.org/Vol-2718/paper27.pdf
Search for optimization of a real-time neural trading system
Michal Pobucký
The Institute of Computer Science, Silesian University in Opava, Czech Republic
michal.pobucky@fpf.slu.cz
Abstract: We describe the development of a complex of individual indicators. Various analytical tools and the
modular system that should be able to predict the move- choice of right methods can reduce the effect of errors and
ment of the currency market (FOREX) using real-time also increase profitability [1][17].
neural networks and then automatically issue instructions With the help of the MetaTrader application, we gain
for buying or selling individual currencies. To access the access to the online trading system on currency markets.
currency market, we use the free MetaTrader 5 application, First, we programmed an automatic expert system, the
which is linked to the Thisifi Trading Terminal (TTT) ap- so-called ExpertAdvisor, which obtains the values of in-
plication that we created. TTT first reads the data needed dividual indicators from the market and stores them in a
for network training and then selects the training sets. The database. We programmed this automaton in a special lan-
samples serve as input for training the neural network. guage MQL. This ExpertAdvisor stores all types of indi-
Once the network is trained, the user can switch to on- cators, i.e. trend indicators [4], [8], [14], [25], [26], os-
line mode, where the network responds in real time to a cillators [3], [5], [6], [7], [10], [18], [22], [28], volume
change in the currency market and passes to MetaTrader indicators [2], [24] and Bill Williams indicators [9], [21].
instructions for the transaction. The values of individual indicators are sent to the in-
put of the neural network - indicators that take the values
1 (buy), -1 (sell) or 0 (no instruction). The broker, as a
1 Introduction human person, himself evaluates the information from in-
dividual indicators and, based on his own experience, sub-
Business trading tools aided with artificial intelligence sequently instructs the currency market. At this stage, our
methods, and artificial neural networks in particular, are trading system will evaluate the values of the indicators,
a frequent research topic. We aim at a development of a not the actual price position in the market. This procedure
a complex modular trading system which shoudl be fully is different from previous research. For instance, [29] fo-
autonomous. The article focuses on the development of cuses on eight indicators, four of which are simple mov-
that part of the system that should be able to trade auto- ing averages and four are exponential. [30] used LSTM
matically on the currency market - FOREX. We experi- model with 11 indicators, [23] then gives an overview of
ment with trading strategies using multilayer perceptrons the neural networks that were used for prediction on finan-
with the Backpropagation training and we present results cial time series and indicators are represented here rather
of this initial study. While this network model is standard, marginally. So far, the research has not focused on a train-
the novelty of our approach lies in the composition of the ing set comprising several hundred market indicators.
training dataset. Future extensions of our trading system We trade on the currency pair EUR/USD, because it has
are described in Conclusions. The reader is referred to [20] the greatest volatility. Trading is minute by minute, so
for a more detailed description of our trading system. we work with minute charts and data. Other authors use
The strategy of the trading system should not be based also 1-hour time frequencies [19], daily and monthly fre-
on the fact that based on the input data we will predict the quencies [11]. All fees (broker AdmiralMarkets), includ-
exact movement of the price by the predicted value, such ing slippage, are already included in the profit of individ-
as [27], where the input of the neural network is the set ual trades, so the net financial profit is really the profit that
of last 25 High or Low values. Currently, currency mar- comes to the client’s account. The lot size is set to 0.1.
ket traders perform either fundamental or technical analy-
sis. The vast majority of them use technical analysis based
on a specific set of mathematical tools – indicators. Each 2 Network settings
broker uses an application that displays a graph on which
the broker sees the current price of a currency pair and its We used a classical multilayer perceptron with the Back-
movement up or down. The broker can also add indicators propagation training [12] based on gradient descent
to this basic chart, which can indicate whether the price method [16], while the activation function is similar to a
will rise or fall, and tries to predict future movement of hyperbolic tangent of the form:
the market on the basis of his experience and knowledge
x
Copyright c 2020 for this paper by its authors. Use permitted un-
f (x) = √
1 + x2
der Creative Commons License Attribution 4.0 International (CC BY
4.0). and its derivation is
The numerals 2, 3 and 4 denote a trading system where
0 1 2, 3 or 4 neural networks were required to issue an order
f (x) = 3
(1 + x2 ) 2 to buy or sell.
Within each adaptation cycle we check whether the net- Letter K means correction. This makes it possible to
work has already been trained by at least 95% of samples. eliminate situations where neural networks generate, for
When this threshold is reached, we let the unlearned sam- example, a Buy signal and at least one network generates a
ples be trained more often and regularly by a special rule, Sell, then the output of the trading system will be the value
to ensure that the network learns the whole training set. 991 and not Buy. Similarly, if the networks generate a Sell
signal and at least one network indicates a Buy signal, the
movement interval signal system result will be 992 and not Sell.
0.00140 +∞ BUY The letter M indicates that we will only accept buy or
0.00120 0.00140 10th neuron sell signals if the same signal was generated in the previ-
0.00100 0.00120 11th neuron ous time period. This means at least two identical instruc-
0.00075 0.00100 12th neuron tions in a row.
0.00050 0.00075 5th neuron The letter N indicates that we will not trade at night be-
0.00030 0.00050 8th neuron tween 22:00 and 06:00. During this period, trades are the
0.00005 0.00030 4th neuron lowest and at the same time the most unusual signals ap-
-0.00005 0.00005 ZERO pear here.
-0.00030 -0.00005 7th neuron
-0.00050 -0.00030 9th neuron 3.2 Changes in network topology
-0.00075 -0.00050 6th neuron
-0.00100 -0.00075 13th neuron The search for the optimal network began with four indi-
-0.00120 -0.00100 14th neuron cators in the input layer, one output neuron and one or two
-0.00140 -0.00120 15th neuron hidden layers. The maximum allowed error is set to 0.1
−∞ -0.00140 SELL (Buy will be detected in the range from 1 to 0.9 and Zero
in the range from -1 to -0.9.), the total number of cycles
Table 1: Signal table to 100,000, with one cycle making 100 random selections
from the training data set. Data was used from one trading
day. None of the tested networks achieved 100% success,
For the gradual evaluation of the found patterns, we had so none of the networks fully adapted to the training data.
to define the scope of individual business instructions. For Similar results were obtained when expanding the number
each time T0 , the price movement at time T1 to T10 is evalu- of indicators to 7 and 13.
ated, and based on the size of the movement, it is assigned When the input layer expands range from 14 to 26 neu-
to one of the categories. By optimizing the network topol- rons, some of the four networks gradually adapt. While in
ogy, the output layer is extended to 15 neurons, three of networks with 14 and 15 inputs only two out of four were
which are Buy, Sell and Zero, the others are auxiliary (4th successful, after 25 thousand to 50 thousand cycles, for ex-
to 15th). If the maximum price value in period hT1 ; T10 i is ample ad for 21 inputs is already enough 4 to 20 thousand
within ± 0.00005, it is neither an increasing nor a decreas- cycles and in 26 input neurons all networks have already
ing trend. The price is holding almost unchanged and it is fully adapted and with one exception, 3 to 7 thousand cy-
a signal called Zero. On the other hand, everything that is cles were enough. It was similar to up to 40 indicators in
at least 0.00140 from the original value is the Buy signal the input layer.
and everything that is at least −0.00140 from the original We tested a larger range of the input layer and pro-
value is then the Sell signal. See Table 1 for an overview. grammed 151 indicators that will be on the input. The net-
work topology is now 151 – X – 1, X ∈ h10; 63i, maximum
allowed error at 0.25, total number of cycles 300,000, 146
3 Neural network optimization training samples from 3 trading days. From 16 neurons in
the inner layer, the networks are capable of full adaptation
3.1 Variants of the trading system
and 40 to 50 thousand iterations are needed. When the
The autonomous trading system consists of four separate training set is extended to 5 trading days and 187 training
neural networks, which were trained separately on their samples, the error increases and none of the tested net-
own data sets. Depending on the user settings, it is possi- works is able to fully adapt.
ble to specify in more detail how these networks will coop- We programmed the maximum of indicators and sent
erate with each other. At the same time, general conditions them to the input layer, which has now expanded to 236
for trading on the currency market can be set. We will neurons. We also test in parallel on 1,180 input neu-
compare results of trading systems with various settings rons, where information from five consecutive bars is in-
in the tables that follow. Setting of the trading system is put within one data set, with a clear tendency to overfit-
denoted by numbers and letters as follow. ting. Neural networks show similar results of adaptation,
with the topology 1,180 – X – 1 being able to learn the four separate neural networks), instead of 221 Sell sig-
same percentage of training samples in the previous cycle. nals 3,077 to 3,564 signals and instead of 209 Zero signals
From 110 to 130 neurons in the inner layer, we get to a net- 2,163 to 2,468 signals. Thus, instead of 637 recognized
work success of about 80%, slightly increasing from 240 signals, the networks generated 8,798 to 8,928 signals.
neurons onwards. Topology 708 – 240 – 4, we decided to add another neu-
Adaptation on 537 training samples shows a success ron to the output layer, which will be evaluated in all other
rate in the range of 61 to 63% in the 236 – X – 1 net- cases where the samples will not be evaluated as Buy, Sell
work, from the 110th internal neuron to 280. In the 1,180 or Zero. Within the training set, two cases were tested –
– X – 1 network, in the range of 67 to 73%, from 210 in- either 100 or 200 samples for the 4th neuron. The training
ternal neuron after 320. The average error is then in the set contains 737 and 837 samples.
range of 0.46 to 0.51 for the 236 – X – 1 topology and A neural network with four neurons in the output layer
in the range of 0.30 to 0.44 for the 1,180 – X – 1 topol- has significantly better results, as we can see in Table 2.
ogy, which means that we already have about twice the False signals (The neural network misinterpreted the Buy
average error of networks. These results indicate that the or Sell signal and there would be a financial loss in real
constant increase in the training set is already becoming trading.) for Buy fell by 36%, in the variant with 200
counterproductive and the reliability of the trained neural training samples even by 42 to 46%. There was a simi-
network is declining. Therefore, next we examined the ef- lar decrease in false signals for both Zero (by 28 to 30%
fect of changes in the initialization of the weights and the and 42 to 44%, respectively) and Sell (by 32 to 36% and
changes of the training rate coefficient α on the quality of 40 to 46%, respectively). By creating the 4th neuron, we
results. eliminated more than 2,000 erroneously detected signals,
We generated various hidden layer sizes and initial and even by simply increasing the set of training samples
weight settings. The best variant had 240 neurons in the from 100 to 200 for the 4th neuron, we moved the nearly
hidden layer and the initial value of the weights and bias a thousand misrecognized samples that were divided be-
is in the range from h− 1.0; 1.0i to h− 2.0; 2.0i. Subse- tween Buy, Zero, and Sell to the 4th neuron category.
quently, adaptation testing with a different magnitude of
Net 1 3N 4N 100 4N 200 Net 2 3N 4N 100 4N 200
the α coefficient was performed, as well as was sent to 1 3,369 2,125 1,812 1 3,230 2,147 1,942
0 2,352 1,574 1,328 0 2,290 1,612 1,287
the input layer from one to seven bars, so that the input -1 3,077 2,257 1,897 -1 3,324 2,074 1,851
4th sig. - 784 1,667 4th sig. - 690 1,499
layer was large from 236 to 1,652 neurons. The most im- Total 8,798 6,740 6,704 Total 8,844 6,523 6,579
portant aspect of adapting these networks is that none of Net 3
1
3N
3,201
4N 100
2,070
4N 200
1,743
Net 4
1
3N
3,177
4N 100
2,023
4N 200
1,709
these twenty networks was able to fully learn the complete 0
-1
2,163
3,564
1,557
2,101
1,237
1,860
0
-1
2,468
3,220
1,716
2,071
1,372
1,825
training set, so even resetting the input layer range, inner 4th sig.
Total
-
8,928
793
6,521
1,858
6,698
4th sig.
Total
-
8,865
687
6,497
1,549
6,455
layer, initializing weights and biases or training coefficient
did not lead to the goal. The best setting for the α training Table 2: Results of complex activation of networks with
coefficient is 0.001, unlike for example [15], where it is three and four neurons in the output layer
0.05.
Topology 708 – 240 – 6, the training set contains 937
3.3 Optimization of complex network activation and 1,237 samples (for 100 and 200 samples on the 4th to
6th neurons). The neural network again has significantly
The optimization will now take place in terms of the width better results, finding the Buy signal in 1,472 to 1,778
of the neural network output layer. We expanded the out- cases, and for the 200 variant in 1,067 to 1,117 cases. This
put layer from one output neuron, which recognizes three is an improvement of 17-27% (respectively 37-42%) in re-
output values to a set of three neurons, with the first neu- moving Buy, Sell and Zero spurious signals over the four
ron, when excited, indicates the Buy signal, the second neurons in the output layer. The overall improvement over
neuron the Zero signal, and the third neuron the Sell signal. the three neurons in the output layer is then 47-53% (re-
We also standardized the training set to 637 samples, with spectively 66%).
Buy signals 207, Zero signals 209 and Sell signals 221. There has been a change in the ratio between the signals
The search for the optimal topology was performed on we require and the auxiliary signals (4th to 6th neurons).
X – 240 – 3 networks, with the number of bars ranging For networks with four neurons in the output layer, the ra-
from three to seven, so that the size of the input layer tio of generated signals was only 11% in favor of auxiliary
was 708, 944, 1,180, 1,416 and 1,652 neurons. There are neurons (738.5 out of 6,570.25) and 24% (1,643.25 out
two conclusions – all networks have learned to recognize of 6,609), respectively, while for neural networks with six
a complete set of training samples, and less than 5,000 output neurons, this ratio is already 25% (1607.75 out of
repetitions, in some cases even less than 3,000 repetitions, 6328.75) and 47% (2906.5 out of 6102.25), respectively.
were almost always enough for the learning itself. This condition is desirable because the adaptation of the
Topology 708 – 240 – 3, after activating a complex network to Buy, Zero and Sell signals is being refined.
test data set that contained 14,382 samples, we received The adaptation rate with 708 neurons in the input layer
instead of 207 Buy signals 3,177 to 3,369 signals (on begins to reach 9,384 to 9,911 cycles in 1,237 training
samples, so an additional input bar was added and the The best weekly earnings results are highlighted in bold.
topology was changed to 1,180 – 240 – 6, where the adap- Strikethrough are settings that show worse or similar re-
tation rate is only 1,887 to 2,300 cycles. sults than their previous variation. We will no longer count
Topology 1,180 – 240 – 7, the training set contains on these settings. The underlined value of gross loss sig-
1,437 samples (200 samples on the 4th to 7th neurons), nals an excellent ratio of gross profit versus gross loss.
the testing set contains 14,375 samples as usual. The 7th
Settings Transactions Net profit Gross profit Gross loss Weekly profit
neuron was created by dividing the 4th neuron, so that only 4 55 265.36 310.20 -44.84 132.68
3 275 1,191.86 1,395.70 -203.84 595.93
one new neuron was added to the network topology in the 2 729 1,777.43 2,561.38 -783.95 888.715
3K 274 1,189.96 1,393.80 -203.84 594.98
output layer, so it is not surprising that the improvement 2K 713 1,784.93 2,540.78 -755.85 892.465
4M 2 17.30 17.30 0 8.65
in network results is not so great. The improvement for 3M 42 307.50 324.60 -17.10 153.75
2M 131 835.60 890.30 -54.70 417.80
Buy reaches a maximum of 5%, for Zero 10% and for Sell 4N 41 245.30 271.40 -26.10 122.65
3N 168 986.40 1,094.20 -107.80 493.2
15%. 2N 415 1,574.70 1,980.20 -405.50 787.35
3KM 42 307.50 324.60 -17.10 153.75
Topology 1,652 – 240 – 9, the training set contains 2KM 127 843.70 890.30 -46.60 421.85
3KN 167 984.50 1,092.30 -107.80 492.25
1,837 samples. The improvement for Buy is 31-32%, for 2KN 407 1,572.10 1,962.40 -390.30 786.05
4MN 2 17.30 17.30 0 8.65
Zero 26-30% and for Sell 32-34%. The total number of 3MN 26 223.10 228.90 -5.80 111.55
2MN 83 668.80 681.10 -12.30 334.4
identified signals then dropped below six thousand in three 3KMN 26 223.10 228.90 -5.80 111.55
2KMN 82 668.80 681.10 -12.30 334.4
cases out of four for the first time and remained in the
range from 5,894 to 6,008. The ratio of generated signals Table 3: Results of a trading system with a two-week set
is now 66% in favor of auxiliary neurons (3,946.25 out of of training data
5,954).
Topology 2,124 – 240 – 15 and topology 2,360 – 240 – In the figure 1 we see the balance on the trading account
15, the training set contains 3,067 samples. We decided to of The Two system with a correction.
add more output neurons at once – a total of 15, as shown
in Table 1. The adaptation rate was 610 to 800 cycles and
560 to 880 cycles, respectively, indicating that it deterio-
rated in some cases. We got to the edge of the suitability
of the topology, and therefore we chose the better one for
further optimization, i.e. the topology 2,124 – 240 – 15
with nine bars at the input. In terms of improving the net-
work’s results, we found that there was an improvement Figure 1: The Two system with a correction - 2K
again, but it is no longer a leap. With each increasing neu-
ron in the output layer, the network adaptation results will
The validation of trading systems was performed on
improve, however, this improvement will be smaller and
new data from the period from 15 to 21 October of the
smaller. That’s why we decided to change our optimiza-
same year. The success of a data set that the network has
tion strategy.
never dealt with is significantly lower, and neither setting
generates any gain but loss, as we can see in table 4. Al-
3.4 Business system optimization though the network is able to predict movements in the
currency market, it has not been able to generalize train-
We decided to perform a simulation on a time series and ing to data that was not part of the training set.
observe what results of adaptation and complex activation
Settings Transactions Net profit Gross profit Gross loss
we will get. We will now optimize the entire trading sys- 4 19 -31.20 11.90 -43.10
tem based on 4 separate neural networks. The division of 3 81 -87.11 52.50 -139.61
the training and test set is in a certain ratio (for example, 2 298 -255.75 239.30 -495.05
2K 282 -258.25 218.70 -476.95
[15] uses 70:30), we use a floating layout, where the train- 3M 3 -3.50 2.30 -5.80
ing set includes data with a minimum size of two weeks, a 3N 44 -31.20 47.80 -79.00
2N 145 -103.50 174.20 -277.70
maximum of eight and test data are always from one week. 2KM 19 -32.50 12.90 -45.40
The training set is generated by an automaton, which grad- 2KN 135 -104.40 162.80 -267.20
2MN 14 -21.60 12.10 -33.70
ually goes through the historical data after a minute, and
assigns them the corresponding values of the output signal. Table 4: Results of the trading system with a two-week set
of training data for the new week
Two-week data Training set: 2,543 samples (October 1 –
14, 2018), Buy 208 times, Zero 56 times, Sell 115 times,
duration: 503 - 702 cycles, test set: 14,371 data. As we Three-week data Training set: 2,888 samples (Septem-
can see in table 3, all trading systems show a weekly profit, ber 24 to October 14, 2018), Buy 299 times, Zero 72 times,
the best ones almost $ 900. The individual settings of trad- Sell 229 times, duration: 1,116 - 1,316 cycles, test set:
ing systems are based on the names from the chapter 3.1. 21,565 data. The maximum weekly profit is around $ 600
Settings Transactions Net profit Gross profit Gross loss Weekly profit
to $ 700, see table 5. None of the settings used reached 4 1,164 1,500.13 2,995.16 -1,495.03 500.04
3 4,390 1,610.37 7,763.35 -6,152.98 536.79
the same ratio between gross profit and gross loss as in 2 9,275 -716.29 13,167.30 -13,883.59 -238.76
3K 4,050 1,785.29 7,424.67 -5,639.38 595.09
the two-week data. For two-week data, this ratio was even 2K 7,378 418.79 11,157.34 -10,738.55 139.59
4M 187 573.48 757.78 -184,30 191.16
equal to 55.37, for three-week data it ranges from 3 to 5. 3M 1,226 1,914.46 3,423.11 -1,508.65 638.15
2M 3,604 1,901.36 6,931.41 -5,030.05 633.78
4N 645 1,447.70 2,295.60 -847,90 482.56
Settings Transactions Net profit Gross profit Gross loss Weekly profit 3N 2,339 2,390.50 5,907.90 -3,517.40 796.83
4 237 779.64 1,057.30 -277.66 259.88 2N 4,762 1,696.00 9,700.10 -8,004.10 565.33
3 823 1,816.95 2,779.45 -962.50 605.65 3KM 1,108 1,942.10 3,271.23 -1,329.13 647.36
2 1,742 2,018.73 4,284.10 -2,265.37 672.91 2KM 2,615 2,165.52 5,563.73 -3,398.21 721.84
2K 1,679 2,050.93 4,212.40 -2,161.47 683.64 3KN 2,167 2,454.40 5,662.20 -3,207.80 818.13
3M 116 618.86 717.70 -98.84 206.28 2KN 3,793 2,246.00 8,331.20 -6,085.20 748.66
3N 488 1,451.30 2,070.20 -618.90 483.76 4MN 118 409.20 533.40 -124.20 136.4
2N 988 1,766.30 3,189.10 -1,422.80 588.76 3MN 674 1,692.20 2,547.80 -855.60 564.06
2KM 260 972.95 1,247.69 -274.74 324.31 2MN 1,946 2,337.60 5,192.30 -2,854.70 792.53
2KN 950 1,783.50 3,136.70 -1,353.20 594.5 3KMN 618 1,686.80 2,433.80 -747.00 562.26
2MN 169 633.30 851.10 -217.80 211.10 2KMN 1,399 2,394.30 4,207.50 -1,813.20 798.1
Table 5: Results of a trading system with a three-week set Table 7: Results of a trading system with a three-week
of training data truncated set of training data
The validation of trading systems was performed on
new data from the period from 15 to 21 October of the Although the results of the trading system are very sat-
same year. Setting 4 is profitable, however $ 10.69 per isfactory within the training period, the validation on the
week is not a successful strategy. If we look at the prof- new week from 15 to 21 October of the same year proves
itability of trades by hour, then it is clear that this setting the exact opposite. None of the tested trading systems
has the largest losses during the night hours. The 3M setup shows a net profit.
is also profitable, however, only 11 transactions per week
and a net profit of less than $ 30 is still low. There was
an increase in the number of transactions for all trading
systems, ranging from 1.76 times to 3.66 times, compared
to trading systems with a two-week set of training data.
There was an increase in net profit for the 4, 3M, 3N,
2KM and 2MN settings, gross profit for all settings and
gross loss for all but 3M, where the gross loss decreased Figure 2: The Three and not night trading system - 3N
from $ 5.80 to $ 1.50. See table 6.
Settings Transactions Net profit Gross profit Gross loss
4 44 10.69 64.30 -53.61 Four-week data truncated Training set: 1,179 samples
3 206 -113.91 224.40 -338.31 (September 17 to October 14, 2018), Buy 381 times, Zero
2 526 -382.96 486.40 -869.36
2K 510 -363.56 471.80 -835.36 95 times, Sell 313 times, duration: 184 - 197 cycles, test
3M 11 28.60 30.10 -1.50 set: 28,758 data. Validation performed for a week from
3N 112 -28.20 179.70 -207.90
2N 264 -146.20 358.20 -504.40 15 to 21 October. Only one setting shows a net profit, the
2KM 51 5.00 72.20 -67.20 number of transactions has increased almost everywhere.
2KN 250 -129.00 343.60 -472.60
2MN 32 23.70 61.60 -37.90 See Table 8 for emergence of several categories:
1. improvement of net profit, increase of gross profit and
Table 6: Results of the trading system with a three-week
decrease of gross loss
set of training data for the new week
2. improvement of net profit, improvement of gross
Because the adaptation on all neural networks took profit, but also increase of gross loss
more than two days, we decided to shorten the training
data set, where in addition to the Buy, Sell and Zero sig- 3. a decrease in net profit, but at the same time there
nals, the signals of the 10th and 15th neurons remain, was an increase in gross profit, which increase is a
which are in the immediate vicinity of the Buy and Sell percentage higher than the increase in gross loss
signals. 4. a decrease in net profit, but at the same time there was
an increase in gross profit, which increase is already
Three-week data truncated Training set: 888 samples a percentage lower than the increase in gross loss
(September 24 to October 14, 2018), Buy 299 times, Zero
5. increase net loss, decrease gross profit and increase
72 times, Sell 229 times, duration: 106 - 151 cycles, test
gross loss
set: 21,565 data. The results of the three-week truncated
data set are remarkable, see table 7. The maximum weekly Trading system Three falls into the best categories, i.e.
profit is around $ 700 to $ 820, which is an improvement 1st and 2nd, trading system Two then into the worst cat-
over the full data set by more than $ 100. The number of egories, i.e. 4th and 5th. The course of validation of the
transactions increased rapidly from a maximum of 1,600 3KMN trading system and the amount of the balance on
to 1,700 to 7,000 and 9,000. the trading account can be seen in the figure 3.
Settings Transactions Net profit Gross profit Gross loss Cat
4 406 -158.86 442.30 -601.16 1. The best and worst trading systems have been reversed,
3 1,596 -1,048.69 1,430.30 -2,478.99 3.
3K 1,490 -862.48 1,382.70 -2,245.18 2. with the Three no longer the best, but falling into the 4th
3M 396 -167.12 428.10 -595.22 1.
2M 1,172 -891.84 1,008.10 -1,899.94 5. and 5th categories. In contrast, the trading system Two
3N 830 -386.30 1,104.70 -1,491.00 2.
2N 1,614 -1,164.90 2,029.00 -3,193.90 5. shows the best results - it moved to the 1st and 2nd cate-
3KM 358 -103.22 418.40 -521.62 1.
2KM 835 -576.83 773.50 -1,350.33 5. gory.
3KN 784 -290.70 1,073.30 -1,364.00 2.
2KN 1,297 -820.10 1,659.60 -2,479.70 4. Based on these results, we conclude that the adaptation
3MN 236 -44.80 359.40 -404.20 2.
2MN 588 -459.60 707.30 -1,166.90 5. of neural networks did not generalize in such a way as to
3KMN 218 6.70 355.00 -348.30 2.
2KMN 437 -291.20 561.80 -853.00 5. create a self-sufficient business system. We will now test
this hypothesis on an eight-week data set.
Table 8: Results of the trading system with a four-week
truncated set of training data for the new week
Eight-week data truncated Training set: 1,988 samples
(August 20 to October 14, 2018), Buy 751 times, Zero 185
times, Sell 652 times, duration: 750 - 903 cycles, test set:
57,538 data. Validation performed for a week from 15 to
21 October. As we can see in Table 10, all tested trad-
ing systems show worse results with an eight-week set of
training data than with a five-week or four-week set. We
Figure 3: The Three with correction, minimal 2 and not therefore confirmed our hypothesis that the neural network
night trading system - 3KMN did not generalize sufficiently.
Settings Transactions Net profit Gross profit Gross loss
4 1,229 -998.11 1,094.00 -2,092.11
Five-week data truncated Training set: 1,386 samples 3 2,690 -2,004.60 2,410.26 -4,414.86
(September 10 to October 14, 2018), Buy 464 times, Zero 3K 2,614 -1,942.39 2,347.26 -4,289.65
3M 892 -732.91 771.00 -1,503.91
121 times, Sell 401 times, duration: 253 - 283 cycles, test 2M 1,650 -1,190.95 1,506.20 -2,697.15
set: 35,951 data. Validation performed for a week from 15 3N 1,354 -783.70 1,803.60 -2,587.30
2N 1,963 -1,182.30 2,600.00 -3,782.30
to 21 October. As we can see in table 9, no setting shows 3KM 853 -662.50 747.60 -1,410.10
a net gain. There has been an increase in transactions ev- 2KM 1,379 -989.54 1,269.80 -2,259.34
3KN 1,313 -743.80 1,768.30 -2,512.10
erywhere. Again, there was a division into five categories: 2KN 1,763 -1,060.60 2,347.90 -3,408.50
3MN 444 -311.20 580.50 -891.70
Settings Transactions Net profit Gross profit Gross loss Cat 2MN 818 -451.50 1,113.40 -1,564.90
4 406 -158.86 442.30 -601.16 1. 3KMN 421 -252.60 565.40 -818.00
4 533 -374.80 486.80 -861.60 4. 2KMN 687 -335.50 949.60 -1,285.10
3 1,798 -1,188.91 1,694.60 -2,883.51 4.
3K 1,700 -1,049.11 1,640.20 -2,689.31 4.
3M 495 -303.42 527.30 -830.72 4.
2M 1,348 -787.61 1,374.00 -2,161.61 2. Table 10: Results of the trading system with an eight-week
3N 957 -455.10 1,293.10 -1,748.20 4.
2N 1,744 -663.60 2,486.60 -3,150.20 1. truncated set of training data for the new week
3KM 445 -242.72 502.80 -745.52 4.
2KM 1,020 -586.12 1,081.50 -1,667.62 3.
3KN 912 -390.40 1,254.10 -1,644.50 4.
2KN 1,482 -620.00 2,076.30 -2,696.30 2.
3MN 281 -111.10 406.40 -517.50 4.
2MN 732 -211.50 1,090.80 -1,302.30 2.
3KMN
2KMN
259
578
-83.40
-127.60
391.00
887.80
-474.40
-1,015.40
5.
2.
4 The best setting for the trading system
Table 9: Results of the trading system with a five-week As part of the search for the ideal trading system, we found
truncated set of training data for the new week only one that was able to generate a net profit even on un-
adapted data. It was a 3KMN system with a four-week
truncated data set – graphs in figures 4 and 5. It reported a
1. improvement of net profit, increase of gross profit and net weekly profit of $ 6.70 for the new week. In contrast, it
decrease of gross loss generates a net profit of $ 1,660.50 on the learned dataset,
2. improvement of net profit, improvement of gross which is a weekly average of $ 415.12. The weekly num-
profit, but also increase of gross loss ber of transactions is 218 in the first case and 216.75 in the
second.
3. a decrease in net profit, but at the same time there
was an increase in gross profit, which increase is a
percentage higher than the increase in gross loss
4. a decrease in net profit, but at the same time there was
an increase in gross profit, which increase is already
a percentage lower than the increase in gross loss
5. increase net loss, decrease gross profit and increase Figure 4: 3KMN with a four-week truncated set of train-
gross loss ing data for the new week
Figure 5: 3KMN with a four-week truncated set of train- Figure 9: 3KMN with a five-week truncated set of training
ing data data
The second best system was 3MN also on the four-week 5 Conclusion
data set (graphs in figures 4 and 5), which generated a net
loss of $ -44.80 on the new week and then reported a net Several conclusions emerged from the search for the opti-
gain of $ 1,724.80 on the learned data set, the weekly av- mal setting of the trading system. The first is the fact that
erage then $ 431.20. The weekly number of transactions the neural network was able to adapt to both unabridged
is 236 in the first case and 242.5 in the second. and abbreviated data sets, and this did not have a signif-
icant effect on trading in the currency market. On the
training dataset, the trading system worked relatively well.
Net profit for the best systems ranged from $ 382.16 to
$ 431.20 per week.
Another conclusion can be made that the constant ex-
pansion of the size of the time period for the training data
set is not the most suitable solution, because the two most
Figure 6: 3MN with a four-week truncated set of training successful trading systems are from a four-week period,
data for the new week one from a five-week period. The eight-week period did
not generate a single successful trading system, and even
resulted in worse than five and four weeks. We also per-
formed a test on a 2KMN trading system with nine bars in
the input layer, topology 2,124 – 420 – 15. When we ex-
tended the training set to an entire calendar year, the result
was a net profit of $ 22.2 on unknown data.
The neural network, although doing relatively well on
the trained data, was unable to perform well with the un-
trained test set due to overfitting. Of all the settings and
Figure 7: 3MN with a four-week truncated set of training
trading systems tested, only one variant reported a net
data
profit on the new unlearned week, with this net weekly
gain being only $ 6.70. This indicates that raw trading
The third best trading system was the 3KMN variant on data series may not have enough regularities that could be
the five-week dataset (graphs in figures 8 and 9), which successfully learned by the network. It will be necessary
reported a net profit of $ -83.40 on the new week, with a to optimize various types of input aggregated from the in-
net profit of $ 1,910.80 and a weekly average of $ 382.16 put dataset, and to include external information possibly
on the adapted dataset. The weekly number of transactions influencing the market. It will be necessary to use a larger
is 259 in the first case and 266.8 in the second. training data set, which the computing power of the GPU
will allow.
Another improvement will be the use of deep learning
methods [13] with various types of layers suitable for time
series prediction. Therefore, in the next phase, the appli-
cation will be migrated from the C# programming lan-
guage to Python, and the Keras and Tensorflow libraries
will be used to find the optimal topology of the trading sys-
Figure 8: 3KMN with a five-week truncated set of training tem. We will, e.g., compare results of convolutional neural
data for the new week networks and recurrent neural networks, including LSTM
or GRU networks, multimodal networks and fuzzy-neural
There is a clear difference between the application of networks. The transition to this technology using GPU
the trading system on adapted and non-adapted data. It is will also make it possible to work with a larger amount
obvious that the neural network could not be generalized of data. We tested that switching to Keras libraries would
enough to be able to generate profit even on unknown data. allow to work with a ten-fold larger data set, so the adap-
tation set can include data from the entire calendar year. [11] GALESCHUK, S. Neural networks performance in ex-
The complex modular system will be completed by change rate prediction. Neurocomputing, 172, pages 446-
a separate group of fuzzy-neural networks, which will 452, 2016.
form a set of rules, which will then be applied in on- [12] GOH, A. T. C. Back-propagation neural networks for mod-
line trading in the currency market. By combining all eling complex systems Artificial Intelligence in Engineering,
these separate modules, we will obtain an artificial intelli- vol. 9, num. 3, pages 143-151, 1995, issn 0954-1810.
gence tool that will incorporate many possible approaches [13] GOODFELLOW, I., BENGIO Y., COURVILLE
to time series prediction. A. Deep Learning. MIT Press, 2016.
http://www.deeplearningbook.org
[14] How to Use the Parabolic SAR Indicator to
Acknowledgements Trade the Trend [online]. [cit. 17. 6. 2019].
https://admiralmarkets.com/education/articles/forex-
This work was supported by the Ministry of Educa- indicators/parabolic-sar-indicator
tion, Youth and Sports Of the Czech Republic from the [15] KIPRUTO, G., MUNG’ATU, J., ORWA, G.,
National Programme of Sustainability (NPU II) project GAITHIMBA, N. Application of Artificial Neural Net-
work (Ann). In Modeling Foreign Currency Exchange
IT4Innovations Excellence in Science - LQ1602, and by
Rates. International Journal of Scientific Research and
the Silesian University in Opava under the Student Fund- Management. vol 6, issue 10, pages 111-123, 2018. ISSN:
ing Scheme, project SGS/11/2019. 2321-3418
[16] LAM, M. Neural network techniques for financial per-
formance prediction: integrating fundamental and techni-
References
cal analysis. In Decision Support Systems, vol. 37, Issue 4,
pages 567-581, 2004.
[1] ABRAHAM, A., CHOWDHURY, M. U. An Intelligent
[17] LAVANYA, V., PARVEENTAJ, M. Foreign Currency Ex-
FOREX Monitoring System. In Proceedings of IEEE Interna-
change Rate (FOREX) using Neural Network. In Interna-
tional Conference on Info-tech and Infonet, Beijing, China,
tional Journal of Science and Research (IJSR 2013) ISSN
page 523-528, 2001.
(Online): 2319-7064.
[2] Accumulation/Distribution (A/D) [online]. Copyright c
[18] The MACD Indicator In Depth [online]. [cit. 17. 6.
2009 - 2019 FXstreet.cz s.r.o. [cit. 17. 6. 2019].
2019]. https://admiralmarkets.com/education/articles/forex-
https://www.fxstreet.cz/accumulationdistribution-ad.html
indicators/macd-indicator-in-depth
[3] CMF – Chaikin Money Flow [online]. Copy-
[19] MANSOUR, F., AKAY, M. Predicting Exchange Rate by
right c eTradink.sk [cit. 17. 6. 2019]
Using Time Series Multilayer Perceptron. In Science and en-
https://www.etrading.sk/cz/technicka-analyza/116-
gineering congress (IMSEC), 2019.
indikatory/649-cmf-chaikin-money-flow
[20] POBUCKÝ, M. Neuronový obchodní systém pracující v
[4] Discover the Advantages of the MT4 Moving
reálném čase. Opava: Silesian University, The Institute of
Average Indicator [online]. [cit. 16. 6. 2019].
Computer Science, Rigorous thesis, 2020. 118 s.
https://admiralmarkets.com/education/articles/forex-
indicators/moving-average-indicator [21] Put the Pedal to the Metal With the Accel-
erator Oscillator [online]. [cit. 17. 6. 2019].
[5] Everything You Need to Know About the Oscillator of
https://admiralmarkets.com/education/articles/forex-
Moving Average (OsMA Indicator) [online]. [cit. 17. 6.
indicators/accelerator-oscillator
2019]. https://admiralmarkets.com/education/articles/forex-
indicators/osma-indicator [22] See Beneath the Surface of the Market With the
Bears and Bulls Power Indicators [online]. [cit. 17. 6.
[6] Find the Truth Behind the Trend with the MT4
2019]. https://admiralmarkets.com/education/articles/forex-
Momentum Indicator [online]. [cit 17. 6. 2019].
indicators/bears-and-bulls-power-indicator
https://admiralmarkets.com/education/articles/forex-
indicators/momentum-indicator-mt4-explained-in-detail [23] SEZER, O. B., GUDELEK, M. U., OZBAYOGLU, A. M.
Financial Time Series Forecasting with Deep Learning : A
[7] Force Index - FRC [online]. [cit. 17. 6. 2019].
Systematic Literature Review: 2005-2019. In Applied Soft
https://www.instaforex.eu/cz/forex_technical_indicators/frc
Computing, vol 90, 2020.
[8] Forex Trading Strategies With Envelopes Indicator [online].
[24] Track the Pulse of the Market with the Money
Copyright @ 2019 Dolphintrader.com. [cit. 17. 6. 2019].
Flow Index Indicator [online]. [cit. 17. 6. 2019].
https://www.dolphintrader.com/forex-trading-strategies-
https://admiralmarkets.com/education/articles/forex-
with-envelopes-indicator/
indicators/money-flow-index-indicator
[9] Get a Feel for Market Momentum with the Awe-
[25] Trading With the Cloud: Using the Ichimoku Kinko
some Oscillator Indicator [online]. [cit. 17. 6. 2019].
Hyo Indicator in MetaTrader 4 [online]. [cit. 17. 6.
https://admiralmarkets.com/education/articles/forex-
2019]. https://admiralmarkets.com/education/articles/forex-
indicators/awesome-oscillator-indicator
indicators/ichimoku-kinko-hyo-indicator
[10] Get Ahead of the Curve with the MT4 Commodity
[26] Vidya – Variable Index Dynamic Average [on-
Channel Index Indicator [online]. [cit. 17. 6. 2019].
line]. Copyright c eTradink.sk [cit. 16. 6.
https://admiralmarkets.com/education/articles/forex-
2019] https://etrading.sk/cz/technicka-analyza/116-
indicators/commodity-channel-index-indicator
indikatory/639-vidya-variable-index-dynamic-average
[27] VUKOVIC, D., VYKLYUK, V. Forex predicton with neu-
ral network: usd/eur currency pair. In Actual Problems of
Economics. vol 10, pages 251-261, 2013.
[28] %R: Williams Percent Range [online]. Copyright c
2009 - 2019 FXstreet.cz s.r.o. [cit. 17. 6. 2019].
https://www.fxstreet.cz/r-williams-percent-range.html
[29] YONG, Y. L., NGO, D. C. L., LEE, Y. Technical Indicators
for Forex Forecasting: A Preliminary Study. In Advances
in Swarm and Computational Intelligence, 6th International
Conference, ICSI, Lecture Notes in Computer Science book
series (LNCS, volume 9142), pages 87 - 97, 2015.
[30] ZHANG, X., TAN, Y. Deep stock ranker: A LSTM neu-
ral network model for stock selection. In Data Mining and
Big Data, pages 614–623. Springer International Publishing,
2018.