We describe the development of a complex modular system that should be able to predict the movement of the currency market (FOREX) using real-time neural networks and then automatically issue instructions for buying or selling individual currencies. To access the currency market, we use the free MetaTrader 5 application, which is linked to the Thisifi Trading Terminal (TTT) application that we created. TTT first reads the data needed for network training and then selects the training sets. The samples serve as input for training the neural network. Once the network is trained, the user can switch to online mode, where the network responds in real time to a change in the currency market and passes to MetaTrader instructions for the transaction.
Business trading tools aided with artificial intelligence
methods, and artificial neural networks in particular, are
a frequent research topic. We aim at a development of a
a complex modular trading system which shoudl be fully
autonomous. The article focuses on the development of
that part of the system that should be able to trade
automatically on the currency market - FOREX. We
experiment with trading strategies using multilayer perceptrons
with the Backpropagation training and we present results
of this initial study. While this network model is standard,
the novelty of our approach lies in the composition of the
training dataset. Future extensions of our trading system
are described in Conclusions. The reader is referred to [
The strategy of the trading system should not be based
on the fact that based on the input data we will predict the
exact movement of the price by the predicted value, such
as [
With the help of the MetaTrader application, we gain
access to the online trading system on currency markets.
First, we programmed an automatic expert system, the
so-called ExpertAdvisor, which obtains the values of
individual indicators from the market and stores them in a
database. We programmed this automaton in a special
language MQL. This ExpertAdvisor stores all types of
indicators, i.e. trend indicators [
The values of individual indicators are sent to the
input of the neural network - indicators that take the values
1 (buy), -1 (sell) or 0 (no instruction). The broker, as a
human person, himself evaluates the information from
individual indicators and, based on his own experience,
subsequently instructs the currency market. At this stage, our
trading system will evaluate the values of the indicators,
not the actual price position in the market. This procedure
is different from previous research. For instance, [
We trade on the currency pair EUR/USD, because it has
the greatest volatility. Trading is minute by minute, so
we work with minute charts and data. Other authors use
also 1-hour time frequencies [
We used a classical multilayer perceptron with the
Backpropagation training [
x 1 + x2 and its derivation is f 0(x) = (1 + x2) 23
Within each adaptation cycle we check whether the network has already been trained by at least 95% of samples. When this threshold is reached, we let the unlearned samples be trained more often and regularly by a special rule, to ensure that the network learns the whole training set.
For the gradual evaluation of the found patterns, we had to define the scope of individual business instructions. For each time T0, the price movement at time T1 to T10 is evaluated, and based on the size of the movement, it is assigned to one of the categories. By optimizing the network topology, the output layer is extended to 15 neurons, three of which are Buy, Sell and Zero, the others are auxiliary (4th to 15th). If the maximum price value in period hT1; T10i is within 0.00005, it is neither an increasing nor a decreasing trend. The price is holding almost unchanged and it is a signal called Zero. On the other hand, everything that is at least 0:00140 from the original value is the Buy signal and everything that is at least 0:00140 from the original value is then the Sell signal. See Table 1 for an overview. The autonomous trading system consists of four separate neural networks, which were trained separately on their own data sets. Depending on the user settings, it is possible to specify in more detail how these networks will cooperate with each other. At the same time, general conditions for trading on the currency market can be set. We will compare results of trading systems with various settings in the tables that follow. Setting of the trading system is denoted by numbers and letters as follow.
The numerals 2, 3 and 4 denote a trading system where 2, 3 or 4 neural networks were required to issue an order to buy or sell.
Letter K means correction. This makes it possible to eliminate situations where neural networks generate, for example, a Buy signal and at least one network generates a Sell, then the output of the trading system will be the value 991 and not Buy. Similarly, if the networks generate a Sell signal and at least one network indicates a Buy signal, the system result will be 992 and not Sell.
The letter M indicates that we will only accept buy or sell signals if the same signal was generated in the previous time period. This means at least two identical instructions in a row.
The letter N indicates that we will not trade at night between 22:00 and 06:00. During this period, trades are the lowest and at the same time the most unusual signals appear here. 3.2
The search for the optimal network began with four indicators in the input layer, one output neuron and one or two hidden layers. The maximum allowed error is set to 0.1 (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 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, so none of the networks fully adapted to the training data. Similar results were obtained when expanding the number of indicators to 7 and 13.
When the input layer expands range from 14 to 26 neurons, some of the four networks gradually adapt. While in networks with 14 and 15 inputs only two out of four were successful, after 25 thousand to 50 thousand cycles, for example ad for 21 inputs is already enough 4 to 20 thousand cycles and in 26 input neurons all networks have already fully adapted and with one exception, 3 to 7 thousand cycles were enough. It was similar to up to 40 indicators in the input layer.
We tested a larger range of the input layer and programmed 151 indicators that will be on the input. The network topology is now 151 – X – 1, X 2 h10; 63i, maximum allowed error at 0.25, total number of cycles 300,000, 146 training samples from 3 trading days. From 16 neurons in the inner layer, the networks are capable of full adaptation and 40 to 50 thousand iterations are needed. When the training set is extended to 5 trading days and 187 training samples, the error increases and none of the tested networks is able to fully adapt.
We programmed the maximum of indicators and sent them to the input layer, which has now expanded to 236 neurons. We also test in parallel on 1,180 input neurons, where information from five consecutive bars is input within one data set, with a clear tendency to overfitting. Neural networks show similar results of adaptation, with the topology 1,180 – X – 1 being able to learn the same percentage of training samples in the previous cycle. From 110 to 130 neurons in the inner layer, we get to a network success of about 80%, slightly increasing from 240 neurons onwards.
Adaptation on 537 training samples shows a success rate in the range of 61 to 63% in the 236 – X – 1 network, from the 110th internal neuron to 280. In the 1,180 – X – 1 network, in the range of 67 to 73%, from 210 internal neuron after 320. The average error is then in the range of 0.46 to 0.51 for the 236 – X – 1 topology and in the range of 0.30 to 0.44 for the 1,180 – X – 1 topology, which means that we already have about twice the average error of networks. These results indicate that the constant increase in the training set is already becoming counterproductive and the reliability of the trained neural network is declining. Therefore, next we examined the effect of changes in the initialization of the weights and the changes of the training rate coefficient a on the quality of results.
We generated various hidden layer sizes and initial
weight settings. The best variant had 240 neurons in the
hidden layer and the initial value of the weights and bias
is in the range from h 1.0; 1.0i to h 2.0; 2.0i.
Subsequently, adaptation testing with a different magnitude of
the a coefficient was performed, as well as was sent to
the input layer from one to seven bars, so that the input
layer was large from 236 to 1,652 neurons. The most
important aspect of adapting these networks is that none of
these twenty networks was able to fully learn the complete
training set, so even resetting the input layer range, inner
layer, initializing weights and biases or training coefficient
did not lead to the goal. The best setting for the a training
coefficient is 0.001, unlike for example [
The optimization will now take place in terms of the width of the neural network output layer. We expanded the output layer from one output neuron, which recognizes three output values to a set of three neurons, with the first neuron, when excited, indicates the Buy signal, the second neuron the Zero signal, and the third neuron the Sell signal. We also standardized the training set to 637 samples, with Buy signals 207, Zero signals 209 and Sell signals 221.
The search for the optimal topology was performed on X – 240 – 3 networks, with the number of bars ranging from three to seven, so that the size of the input layer was 708, 944, 1,180, 1,416 and 1,652 neurons. There are two conclusions – all networks have learned to recognize a complete set of training samples, and less than 5,000 repetitions, in some cases even less than 3,000 repetitions, were almost always enough for the learning itself.
Topology 708 – 240 – 3, after activating a complex test data set that contained 14,382 samples, we received instead of 207 Buy signals 3,177 to 3,369 signals (on four separate neural networks), instead of 221 Sell signals 3,077 to 3,564 signals and instead of 209 Zero signals 2,163 to 2,468 signals. Thus, instead of 637 recognized signals, the networks generated 8,798 to 8,928 signals.
Topology 708 – 240 – 4, we decided to add another neuron to the output layer, which will be evaluated in all other cases where the samples will not be evaluated as Buy, Sell or Zero. Within the training set, two cases were tested – either 100 or 200 samples for the 4th neuron. The training set contains 737 and 837 samples.
A neural network with four neurons in the output layer has significantly better results, as we can see in Table 2. False signals (The neural network misinterpreted the Buy or Sell signal and there would be a financial loss in real trading.) for Buy fell by 36%, in the variant with 200 training samples even by 42 to 46%. There was a similar decrease in false signals for both Zero (by 28 to 30% and 42 to 44%, respectively) and Sell (by 32 to 36% and 40 to 46%, respectively). By creating the 4th neuron, we eliminated more than 2,000 erroneously detected signals, and even by simply increasing the set of training samples from 100 to 200 for the 4th neuron, we moved the nearly a thousand misrecognized samples that were divided between Buy, Zero, and Sell to the 4th neuron category.
Net 1 1 -01 4th sig.
Total Net 3 1 -01 4th sig.
Total 3N 3,369 2,352 38,,07-7978 3N 3,201 2,163 38,,59-6248 4N 100 2,125 1,574 2,257 784 6,740 4N 100 2,070 1,557 2,101 793 6,521 4N 200 1,812 1,328 1,897 1,667 6,704 4N 200 1,743 1,237 1,860 1,858 6,698
Net 2 1 -01 4th sig.
Total Net 4 1 -01 4th sig.
Total 3N 3,230 2,290 38,,38-2444 3N 3,177 2,468 38,,28-2605 4N 100 2,147 1,612 2,074 690 6,523 4N 100 2,023 1,716 2,071 687 6,497 4N 200 1,942 1,287 1,851 1,499 6,579 4N 200 1,709 1,372 1,825 1,549 6,455
Topology 708 – 240 – 6, the training set contains 937 and 1,237 samples (for 100 and 200 samples on the 4th to 6th neurons). The neural network again has significantly better results, finding the Buy signal in 1,472 to 1,778 cases, and for the 200 variant in 1,067 to 1,117 cases. This is an improvement of 17-27% (respectively 37-42%) in removing Buy, Sell and Zero spurious signals over the four neurons in the output layer. The overall improvement over the three neurons in the output layer is then 47-53% (respectively 66%).
There has been a change in the ratio between the signals we require and the auxiliary signals (4th to 6th neurons). For networks with four neurons in the output layer, the ratio of generated signals was only 11% in favor of auxiliary neurons (738.5 out of 6,570.25) and 24% (1,643.25 out of 6,609), respectively, while for neural networks with six output neurons, this ratio is already 25% (1607.75 out of 6328.75) and 47% (2906.5 out of 6102.25), respectively. This condition is desirable because the adaptation of the network to Buy, Zero and Sell signals is being refined.
The adaptation rate with 708 neurons in the input layer begins to reach 9,384 to 9,911 cycles in 1,237 training samples, so an additional input bar was added and the topology was changed to 1,180 – 240 – 6, where the adaptation rate is only 1,887 to 2,300 cycles.
Topology 1,180 – 240 – 7, the training set contains 1,437 samples (200 samples on the 4th to 7th neurons), the testing set contains 14,375 samples as usual. The 7th neuron was created by dividing the 4th neuron, so that only one new neuron was added to the network topology in the output layer, so it is not surprising that the improvement in network results is not so great. The improvement for Buy reaches a maximum of 5%, for Zero 10% and for Sell 15%.
Topology 1,652 – 240 – 9, the training set contains 1,837 samples. The improvement for Buy is 31-32%, for Zero 26-30% and for Sell 32-34%. The total number of identified signals then dropped below six thousand in three cases out of four for the first time and remained in the range from 5,894 to 6,008. The ratio of generated signals is now 66% in favor of auxiliary neurons (3,946.25 out of 5,954).
Topology 2,124 – 240 – 15 and topology 2,360 – 240 – 15, the training set contains 3,067 samples. We decided to 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 deteriorated 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 network’s results, we found that there was an improvement again, but it is no longer a leap. With each increasing neuron in the output layer, the network adaptation results will improve, however, this improvement will be smaller and smaller. That’s why we decided to change our optimization strategy. 3.4
We decided to perform a simulation on a time series and
observe what results of adaptation and complex activation
we will get. We will now optimize the entire trading
system based on 4 separate neural networks. The division of
the training and test set is in a certain ratio (for example,
[
In the figure 1 we see the balance on the trading account of The Two system with a correction. The validation of trading systems was performed on new data from the period from 15 to 21 October of the same year. The success of a data set that the network has never dealt with is significantly lower, and neither setting generates any gain but loss, as we can see in table 4. Although the network is able to predict movements in the currency market, it has not been able to generalize training to data that was not part of the training set. Three-week data Training set: 2,888 samples (September 24 to October 14, 2018), Buy 299 times, Zero 72 times, Sell 229 times, duration: 1,116 - 1,316 cycles, test set: 21,565 data. The maximum weekly profit is around $ 600 to $ 700, see table 5. None of the settings used reached the same ratio between gross profit and gross loss as in the two-week data. For two-week data, this ratio was even equal to 55.37, for three-week data it ranges from 3 to 5.
Settings 4 3 22K 3M 3N 2N 2KM 2KN 2MN
The validation of trading systems was performed on new data from the period from 15 to 21 October of the same year. Setting 4 is profitable, however $ 10.69 per week is not a successful strategy. If we look at the profitability of trades by hour, then it is clear that this setting has the largest losses during the night hours. The 3M setup 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 from $ 5.80 to $ 1.50. See table 6.
Because the adaptation on all neural networks took more than two days, we decided to shorten the training data set, where in addition to the Buy, Sell and Zero signals, the signals of the 10th and 15th neurons remain, which are in the immediate vicinity of the Buy and Sell signals.
Three-week data truncated Training set: 888 samples (September 24 to October 14, 2018), Buy 299 times, Zero 72 times, Sell 229 times, duration: 106 - 151 cycles, test set: 21,565 data. The results of the three-week truncated data set are remarkable, see table 7. The maximum weekly profit is around $ 700 to $ 820, which is an improvement over the full data set by more than $ 100. The number of transactions increased rapidly from a maximum of 1,600 to 1,700 to 7,000 and 9,000.
Although the results of the trading system are very satisfactory within the training period, the validation on the new week from 15 to 21 October of the same year proves the exact opposite. None of the tested trading systems shows a net profit. Four-week data truncated Training set: 1,179 samples (September 17 to October 14, 2018), Buy 381 times, Zero 95 times, Sell 313 times, duration: 184 - 197 cycles, test set: 28,758 data. Validation performed for a week from 15 to 21 October. Only one setting shows a net profit, the number of transactions has increased almost everywhere. See Table 8 for emergence of several categories: 1. improvement of net profit, increase of gross profit and decrease of gross loss 2. improvement of net profit, improvement of gross profit, but also increase of gross loss 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 gross loss
Trading system Three falls into the best categories, i.e. 1st and 2nd, trading system Two then into the worst categories, i.e. 4th and 5th. The course of validation of the 3KMN trading system and the amount of the balance on the trading account can be seen in the figure 3. Settings 4 33K 3M 2M 3N 2N 3KM 2KM 3KN 2KN 3MN 2MN 3KMN 2KMN Five-week data truncated Training set: 1,386 samples (September 10 to October 14, 2018), Buy 464 times, Zero 121 times, Sell 401 times, duration: 253 - 283 cycles, test set: 35,951 data. Validation performed for a week from 15 to 21 October. As we can see in table 9, no setting shows a net gain. There has been an increase in transactions everywhere. Again, there was a division into five categories: Settings 4 4 33K 3M 2M 3N 2N 3KM 2KM 3KN 2KN 3MN 2MN 3KMN 2KMN 1. improvement of net profit, increase of gross profit and decrease of gross loss 2. improvement of net profit, improvement of gross profit, but also increase of gross loss 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 gross loss
The best and worst trading systems have been reversed, with the Three no longer the best, but falling into the 4th and 5th categories. In contrast, the trading system Two shows the best results - it moved to the 1st and 2nd category.
Based on these results, we conclude that the adaptation of neural networks did not generalize in such a way as to create a self-sufficient business system. We will now test this hypothesis on an eight-week data set.
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 trading systems show worse results with an eight-week set of training data than with a five-week or four-week set. We therefore confirmed our hypothesis that the neural network did not generalize sufficiently.
Settings 4 3 3K 3M 2M 3N 2N 3KM 2KM 3KN 2KN 3MN 2MN 3KMN 2KMN
As part of the search for the ideal trading system, we found only one that was able to generate a net profit even on unadapted data. It was a 3KMN system with a four-week truncated data set – graphs in figures 4 and 5. It reported a net weekly profit of $ 6.70 for the new week. In contrast, it generates a net profit of $ 1,660.50 on the learned dataset, which is a weekly average of $ 415.12. The weekly number of transactions is 218 in the first case and 216.75 in the second.
The second best system was 3MN also on the four-week 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 gain of $ 1,724.80 on the learned data set, the weekly average then $ 431.20. The weekly number of transactions is 236 in the first case and 242.5 in the second.
The third best trading system was the 3KMN variant on the five-week dataset (graphs in figures 8 and 9), which reported a net profit of $ -83.40 on the new week, with a net profit of $ 1,910.80 and a weekly average of $ 382.16 on the adapted dataset. The weekly number of transactions is 259 in the first case and 266.8 in the second.
There is a clear difference between the application of the trading system on adapted and non-adapted data. It is obvious that the neural network could not be generalized enough to be able to generate profit even on unknown data. 5
Several conclusions emerged from the search for the optimal setting of the trading system. The first is the fact that the neural network was able to adapt to both unabridged and abbreviated data sets, and this did not have a significant 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 expansion of the size of the time period for the training data set is not the most suitable solution, because the two most successful trading systems are from a four-week period, 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 performed a test on a 2KMN trading system with nine bars in the input layer, topology 2,124 – 420 – 15. When we extended 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 untrained test set due to overfitting. Of all the settings and trading systems tested, only one variant reported a net profit on the new unlearned week, with this net weekly gain being only $ 6.70. This indicates that raw trading data series may not have enough regularities that could be successfully learned by the network. It will be necessary to optimize various types of input aggregated from the input dataset, and to include external information possibly influencing the market. It will be necessary to use a larger training data set, which the computing power of the GPU will allow.
Another improvement will be the use of deep learning
methods [
The complex modular system will be completed by a separate group of fuzzy-neural networks, which will form a set of rules, which will then be applied in online trading in the currency market. By combining all these separate modules, we will obtain an artificial intelligence tool that will incorporate many possible approaches to time series prediction.
This work was supported by the Ministry of Education, Youth and Sports Of the Czech Republic from the National Programme of Sustainability (NPU II) project IT4Innovations Excellence in Science - LQ1602, and by the Silesian University in Opava under the Student Funding Scheme, project SGS/11/2019.