KR102171252B1 - Method and server for predicting exchange rate - Google Patents

Abstract

Provided is a method for predicting an exchange rate which can manage a risk. The method for predicting an exchange rate comprises the steps of: collecting a data set related to a plurality of foreign exchange data; creating a Bayesian network; learning the Bayesian network to input the data set and output distribution data for each period; collecting a plurality of real-time foreign exchange data; inputting the plurality of real-time foreign exchange data to the Bayesian network to obtain the distribution data for each period; and predicting an exchange rate based on the distribution data for each period.

Description

How to predict exchange rates and server {METHOD AND SERVER FOR PREDICTING EXCHANGE RATE}

The present invention relates to a method and a server for predicting exchange rates.

The share of foreign capital in the financial market is gradually increasing, and the dependence of each company on importing foreign raw materials and parts is increasing.

Usually, most companies have risks of foreign exchange and raw materials in import/export or overseas operations, and when they want to sell or buy foreign currency due to export negotiations or import settlement, they predict the risk of currency exchange based on their own expected 　 exchange rate. There is a situation.

However, even if the risk of currency exchange is predicted, it is difficult to make a decision to buy or sell foreign currency held or required. For this reason, companies simply entrust foreign exchange trading agencies or hire foreign exchange management specialists for foreign exchange transactions, but they have to pay a lot of additional economic costs for this.

Referring to FIG. 1A, when a currency exchange time point 103 is derived from the exchange rate movement graph 101 predicted through a conventional exchange rate prediction algorithm, more exchanges must be made at the derived exchange time point 103, but such a currency exchange time point. It is difficult to determine the best time to exchange money with only a single piece of information on (103). This is because there is no guarantee that the movement of the actual exchange rate will follow the exchange rate movement graph 101 predicted through the conventional exchange rate prediction algorithm, and there is uncertainty about the exchange rate movement graph 101.

If it is possible to predict the movement of the exchange rate, it is possible to obtain optimal results by intensively conducting currency exchange (buying) between weeks 2 and 3, but it is difficult to accurately predict the actual exchange rate and raw material price. You can also maximize the risk.

In particular, for most companies that do not pay attention to the management of foreign exchange and raw materials, there is a need to consider the volatility of exchange rates because there is a separate main business area that actually makes profits.

Referring to FIG. 1B, it is assumed that the uncertainty of the predicted value for the exchange rate movement graph 101 is displayed in a green area. Here, when the green area becomes wider, it means that the uncertainty of the predicted value for the exchange rate movement graph 101 increases.

If the exchange time point 103 is an uncertain predicted value, the time point at which the exchange time point 103 looks most advantageous on the exchange rate movement graph 101 should be approached conservatively.

Assuming that the exchange rate volatility predicted through the exchange rate movement graph 101 moves within the volatility period (green area), if all money is exchanged between weeks 2 and 3, it is difficult to predict the result because exchange rate volatility is large. In this regard, it can put a great burden on companies.

Korean Patent Application Publication No. 2019-0018950 (published on February 26, 2019)

The present invention is to solve the problems of the prior art described above, and an attempt is made to train a Bayesian network to input a plurality of foreign exchange data data sets to a Bayesian network to output distribution data for each period.

In addition, the present invention is to obtain distribution data for each period by inputting a plurality of real-time foreign exchange data to the learned Bayesian network, and to predict exchange rates based on the obtained distribution data for each period.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, a method for predicting an exchange rate according to a first aspect of the present invention comprises: collecting a plurality of data sets related to foreign exchange data; Creating a Bayesian network; Learning the Bayesian network to input the data set and output distribution data for each period; Collecting a plurality of real-time foreign exchange data; Inputting the plurality of real-time foreign exchange data to the Bayesian network to obtain distribution data for each period; And predicting an exchange rate based on the distribution data for each period.

In the second aspect of the present invention, the exchange rate prediction server includes a data set related to a plurality of foreign exchange data and a collection unit for collecting a plurality of real-time foreign exchange data, a Bayesian network generator that generates a Bayesian network, and inputs the data set. A learning unit for learning the Bayesian network to output distribution data for each period, a distribution data acquisition unit for obtaining distribution data for each period by inputting the plurality of real-time foreign exchange data to the Bayesian network, and predicting an exchange rate based on the distribution data for each period It may include an exchange rate prediction unit.

The above-described problem solving means are merely exemplary and should not be construed as limiting the present invention. In addition to the above-described exemplary embodiments, there may be additional embodiments described in the drawings and detailed description of the invention.

According to any one of the above-described problem solving means of the present invention, the present invention can obtain distribution data for each period by inputting a plurality of real-time foreign exchange data into the learned Bayesian network, and predict an exchange rate based on the obtained distribution data for each period. .

Through this, the present invention does not predict a single value for the currency exchange time as in the prior art, but predicts the exchange rate using distribution data for each period, so that the average moving value prediction for the exchange rate movement and the confidence interval information of the average moving value prediction are Can provide.

In addition, in the present invention, when the distribution data for each period is data with high uncertainty about the exchange rate, the risk can be managed through measures such as reducing the size of an action (eg, the amount of currency exchange) according to the exchange rate prediction.

1A to 1B are diagrams for explaining the difficulty in predicting a currency exchange time according to a conventional method.
2 is a block diagram of an exchange rate prediction server according to an embodiment of the present invention.
3A to 3D are diagrams for describing an Ensemble network and a Bayesian network according to an embodiment of the present invention.
4A to 4D are diagrams for explaining a method of learning a Bayesian network according to an embodiment of the present invention.
5 is a flowchart illustrating a method for predicting exchange rates according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

In the present specification, the term "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized using two or more hardware, or two or more units may be realized using one hardware.

In the present specification, some of the operations or functions described as being performed by the terminal or device may be performed instead by a server connected to the terminal or device. Likewise, some of the operations or functions described as being performed by the server may also be performed by a terminal or device connected to the server.

Hereinafter, with reference to the accompanying configuration diagram or processing flow chart, specific details for the implementation of the present invention will be described.

2 is a block diagram of an exchange rate prediction server according to an embodiment of the present invention.

2, the exchange rate prediction server 10 includes a collection unit 200, a Bayesian network generation unit 210, a learning unit 220, a distribution data acquisition unit 230, and an exchange rate prediction unit 240. I can. However, the exchange rate prediction server 10 shown in FIG. 2 is only an example of implementation of the present invention, and various modifications are possible based on the components shown in FIG. 2.

Hereinafter, FIG. 2 will be described with reference to FIGS. 3A to 4D together.

The collection unit 200 may collect data sets related to a plurality of foreign exchange data. Here, the data set for the plurality of foreign exchange data may include foreign exchange data and distribution data for each actual period of the foreign exchange data.

For example, the collection unit 200 includes exchange rate data, oil price data, raw material price data (eg, price data of gold, silver, aluminum, copper, gasoline, diesel, etc.) of each country for a predetermined period. It is possible to collect foreign exchange data of, and collect actual period-specific distribution data for a plurality of foreign exchange data for a preset period.

In addition, the collection unit 200 may collect macroeconomic data including price index data, national index data, and bond data of each country related to foreign exchange and raw materials, and collect actual distribution data for each period of the collected macroeconomic data. have.

In addition, the collection unit 200 may collect foreign financial market data related to foreign exchange and raw materials, and may collect actual distribution data for each period of the collected foreign financial market data. In addition, the collection unit 200 collects domestic financial market trend data, including domestic financial indices related to foreign exchange and raw materials, and specific individual stock price data, and collects actual period-specific distribution data for the collected domestic financial market trend data. can do.

The pre-processing unit (not shown) may pre-process a data set related to a plurality of foreign exchange data. For example, the preprocessor may extract financial feature information from a plurality of foreign exchange data. Here, the extracted financial feature information may be pre-processed information on a dataset related to the collected foreign exchange data.

For example, the preprocessor may extract moving average information (MA) and volatility information of foreign exchange data on a specific date (N days) from a plurality of foreign exchange data. In addition, the preprocessor may extract information on the closing price of today compared to the closing price on a specific date from the plurality of foreign exchange data, or information on a return rate for a preset period from the present day compared to a return rate for a preset period from a specific date. Alternatively, the preprocessor may extract a relative strength index (RSI) that intuitively shows the strength of the flow and trend of the stock price from the plurality of foreign exchange data.

The Bayesian network generator 210 may generate a Bayesian network. Here, the Bayesian network is a network that outputs a directional acyclic graph of a node representing a variable and an arc representing the dependency relationship between the variables, and is a joint distribution for all variables represented by nodes on the directional acyclic graph. distribution) can be printed.

3A to 3C, the Bayesian network generation unit 210 may generate a Bayesian network 32 including an ensemble network 30 including a plurality of sub-networks for ensemble learning. Here, ensemble learning is a machine learning technique that combines several decision trees to achieve better performance than one decision tree. The core of ensemble learning is to create a strong classifier by combining several weak classifiers, so the accuracy of learning can be improved.

The ensemble network 30 may include a plurality of sub-networks 301, 303, and 305. Here, the structure of each of the plurality of sub-networks 301, 303, 305 is composed of multiple layers including a Self Attention Layer 307, a Cross Section Layer 309, and a Fully Connected Layer 311. Here, the Self Attention Layer (307), Cross Section Layer (309) and Fully Connected Layer (311) refer to the characteristics of foreign exchange and commodity price fluctuations without losing the value of the distribution data for each period in the time series. It is a machine learning layer that extracts and derives specific values for predicting distribution data for each period (prediction data on foreign exchange and raw material price fluctuations) corresponding to the extracted features.

)을 출력할 수 있다. 제 2 하위 네트워크(303)는 복수의 외환 데이터에 관한 데이터셋 중 제 2 외환 데이터(예컨대, 외환과 관련된 해외 금융 시장 데이터) 및 제 2 금융 특징 정보를 입력값으로 받고, 제 2 특정값(

)을 출력할 수 있다. 제 N 하위 네트워크(305)는 복수의 외환 데이터에 관한 데이터셋 중 제 N 외환 데이터(예컨대, 외환과 관련된 금융시장 동향 데이터) 및 제 N 금융 특징 정보를 입력값으로 받고, 제 N 특정값(

)을 출력할 수 있다. Each of the plurality of sub-networks 301, 303, and 305 is not a duplicate of the same sub-network, but is a multi-task network that derives different output values (distribution data by period) according to different input values (multiple foreign exchange data). . For example, the first sub-network 301 receives, as input values, first foreign exchange data (eg, macroeconomic data related to foreign exchange) and first financial characteristic information among a plurality of data sets related to foreign exchange data, and value(

) Can be printed. The second sub-network 303 receives second foreign exchange data (eg, foreign financial market data related to foreign exchange) and second financial characteristic information from among the plurality of data sets related to foreign exchange data as input values, and the second specific value (

) Can be printed. The N-th sub-network 305 receives the N-th foreign exchange data (eg, financial market trend data related to foreign exchange) and the N-th financial characteristic information among the plurality of foreign exchange data-related data sets as input values, and the N-th specific value (

) Can be printed.

)과 제 2 하위 네트워크(303)를 통해 출력된 제 2 특정값(

) 및 제 N 하위 네트워크(305)를 통해 출력된 제 N 특정값(

)을 평균 분산처리하여 기간별 분포 데이터를 예측하여 출력할 수 있다. As described above, the ensemble network 30 may infer (predict) the distribution data for each period by synthesizing the distribution data for each period through a plurality of sub-networks instead of inferring (predicting) the distribution data for each period as a single sub-network. That is, the ensemble network 30 has a first specific value output through the first sub-network 301 (

) And a second specific value output through the second sub-network 303 (

) And the N-th specific value output through the N-th sub-network 305 (

) Can be averaged to predict and output distribution data for each period.

Meanwhile, referring to FIG. 3D, the Bayesian network generator 210 may generate a Bayesian network 32 including an ensemble network 30 and a Gaussian Process Regression model 34. Here, the Bayesian network may include a Gaussian process regression model 34 that outputs a log likelihood by inputting the output of the ensemble network 30.

Hereinafter, a method of learning a Bayesian network will be described with reference to FIGS. 4A to 4C.

The learning unit 220 may train the Bayesian network to output distribution data for each period by inputting a data set about a plurality of collected foreign exchange data into the Bayesian network. For example, the learning unit 220 inputs a dataset including the first foreign exchange data and the actual distribution data for each period of the first foreign exchange data into the Bayesian network, and when the first foreign exchange data is input to the Bayesian network, 1 A Bayesian network can be trained so that actual distribution data for each period of foreign exchange data is output.

Referring to FIG. 4A, the learning unit 220 inputs a data set related to a plurality of collected foreign exchange data into the ensemble network 30 of the Bayesian network 32 and outputs a specific value to the Gaussian process regression model 34. By inputting, the Bayesian network 32 can be trained in the direction of increasing the log likelihood.

The learning unit 220 may learn the Bayesian network to increase the log likelihood based on the loss function of [Equation 1]. That is, the learning unit 220 may learn the Bayesian network in the direction of decreasing -logp(y|X).

[Equation 1]

은 노이즈 제곱값이고, K는 커널 함수이다. Here, m(X) is the output value of the ensemble network 30, N is the total number of foreign exchange data included in the dataset,

Is the squared noise value, and K is the kernel function.

The kernel function K used in [Equation 1] is used in [Equation 2].

[Equation 2]

Referring to FIG. 4B, the learning unit 220 selects one of a plurality of parameterized distributions (that is, distributions that can be described by parameters), and selects a parameterized distribution (eg, a Gaussian normal distribution), and the ensemble network 30 The Bayesian network 32 may be trained to follow a parameter of a parameterized distribution in which a specific value output from is selected (eg, a Gaussian normal distribution parameter is a μ value and a σ value).

For example, the learning unit 220 learns the Bayesian network 32 to reduce heteroscedastic losses for the Gaussian normal distribution by estimating the parameter of the Gaussian normal distribution based on the loss function of [Equation 1]. can do. Here, the heteroscedasticity loss means taking a negative number to the log likelihood of a Gaussian normal distribution, and can be expressed as in Equation 3.

[Equation 3]

The learning unit 220 is a Gaussian normal distribution based on the loss of heteroscedasticity for the Gaussian normal distribution in order to reduce the difference between the distribution data for each period and the actual distribution data for each period output from the Bayesian network 32 trained to follow the Gaussian normal distribution parameter. The parameter of can be improved.

On the other hand, when the value output from the Bayesian network 32 is configured in a form in which it is difficult to specify the distribution data, the specific value output from the ensemble network 30 is estimated as a parameter of a parameterized distribution (eg, normal distribution). The learning method to do can be inefficient. In this case, the specific value output from the ensemble network 30 is not to estimate the parameter, but is output as a single value, but learning using a dropout so that a different specific value is output probabilistically even when the same foreign exchange data is input. You need a way.

Referring to FIG. 4C, each of a plurality of sub-networks may include a dropout neural network. In this case, the learning unit 220 may learn the Bayesian network 32 so that each of the plurality of sub-networks outputs a plurality of random values for predicting exchange rates using the dropout neural network 36. In this case, since the output value of the Bayesian network 32 no longer estimates the parameter of the parameterized distribution, but is output as a random value, distribution data for each period may be obtained based on the variance of the random values.

The dropout neural network 36 is a neural network that does not train the entire neural network, but randomly selects and learns some nodes. Dropout calculates the output of the neural network by randomly selecting nodes at a certain ratio between learning and making the output zero. When the dropout neural network 36 is applied to the Bayesian network 32, nodes and weights to be learned are changed each time, so that the neural network can be effectively prevented from overfitting.

Returning to FIG. 2, the collection unit 200 may collect a plurality of real-time foreign exchange data.

The distribution data acquisition unit 230 may acquire distribution data for each period by inputting a plurality of real-time foreign exchange data into the learned Bayesian network.

For example, referring to FIG. 4D, when real-time won-dollar data is collected, the distribution data acquisition unit 230 provides real-time won-dollar data to a Bayesian network (the Bayesian network of FIG. 3D) composed of an ensemble network and a Gaussian process regression model. By inputting, distribution data 401 for each period of real-time won dollar data may be obtained.

As another example, in the case of collecting real-time renminbi dollar data, the distribution data acquisition unit 230 inputs real-time renminbi dollar data to a Bayesian network consisting of only an ensemble network (the Bayesian network in FIG. (403) can be obtained.

As another example, when real-time yen dollar data is collected, the distribution data acquisition unit 230 inputs real-time yen dollar data to a Bayesian network (Bayesian network in Fig. 4c) composed of an ensemble network and a dropout neural network to provide real-time yen dollar data. Distribution data 405 for each period may be obtained.

At this time, the distribution data acquisition unit 230 obtains a plurality of random values from the Bayesian network and variance of the plurality of random values when obtaining distribution data for each period using a Bayesian network consisting of an ensemble network and a dropout neural network. Distribution data 405 for each period may be obtained based on.

The exchange rate prediction unit 240 may predict the exchange rate based on distribution data for each period.

In another embodiment, the learning unit 220 may learn the Bayesian network to reduce a difference between the distribution data for each period obtained by inputting real-time foreign exchange data into the Bayesian network and the actual distribution data for each period of the real-time foreign exchange data.

The learning unit 220 may perform a heteroscedasticity loss calculation using an average rising predicted value and a volatility index (eg, variance, standard deviation, etc.) for specific values obtained from the Bayesian network. The learning unit 220 may learn the Bayesian network by reflecting the risk aversion tendency of the client to the Bayesian network based on [Equation 4].

[Equation 4]

Here, μ is the average upward predicted value, σ is the volatility index, ra (risk-averseness) is the risk aversion tendency of the client, and z is the Scale Normalizer index.

In [Equation 4], the scale adjustment index is an index that converts 0.8 to 1.2 numbers into 800 to 1200 won when an average of 1000 won needs to be exchanged.

In [Equation 4], the output value (logit) is proportional to the average rising predicted value, inversely proportional to variance, and it can be seen that the greater the risk avoidance tendency, the more sensitive the variance is.

The exchange rate execution unit (not shown) may provide a guide to exchange currency or purchase raw materials or hedging through the output value (logit) of [Equation 4]. For example, in the case of currency exchange, the larger the output value in [Equation 4], the higher the foreign exchange price of the next day (or n days later), so a guide to proceed with the purchase is provided, and in the case of hedging, [Equation 4] ], it can provide a guide to increase the proportion of foreign exchange holdings as the output value increases.

On the other hand, for those skilled in the art, each of the collection unit 200, the Bayesian network generation unit 210, the learning unit 220, the distribution data acquisition unit 230, and the exchange rate prediction unit 240 may be implemented separately, or at least one of them. It will be fully understood that this can be implemented in an integrated manner.

5 is a flowchart illustrating a method for predicting exchange rates according to an embodiment of the present invention.

Referring to FIG. 5, in step S501, the exchange rate prediction server 10 may collect data sets related to a plurality of foreign exchange data.

In step S503, the exchange rate prediction server 10 may generate a Bayesian network.

In step S505, the exchange rate prediction server 10 may train the Bayesian network to input data sets related to a plurality of foreign exchange data and output distribution data for each period.

In step S507, the exchange rate prediction server 10 may collect a plurality of real-time foreign exchange data.

In step S509, the exchange rate prediction server 10 may obtain distribution data for each period by inputting a plurality of real-time foreign exchange data into the Bayesian network.

In step S511, the exchange rate prediction server 10 may predict the exchange rate based on distribution data for each period.

In the above description, steps S501 to S511 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, and the order between steps may be changed.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. .

10: Exchange rate prediction server
200: collection unit
210: Bayesian network generator
220: Learning Department
230: distribution data acquisition unit
240: exchange rate prediction unit

Claims (14)

In the method for the exchange rate prediction server to predict the exchange rate,
Collecting, by a collection unit of the exchange rate prediction server, a data set related to a plurality of foreign exchange data;
Generating a Bayesian network by a Bayesian network generator of the exchange rate prediction server;
Learning the Bayesian network so that a learning unit of the exchange rate prediction server inputs the data set and outputs distribution data for each period;
Collecting a plurality of real-time foreign exchange data by the collection unit;
Obtaining distribution data for each period by inputting the plurality of real-time foreign exchange data into the Bayesian network by a distribution data acquisition unit of the exchange rate prediction server; And
Predicting an exchange rate based on the distribution data for each period by an exchange rate prediction unit of the exchange rate prediction server
That includes, the exchange rate prediction method.
The method of claim 1,
The Bayesian network includes an ensemble network including a plurality of sub-networks for ensemble learning.
The method of claim 2,
The Bayesian network includes a Gaussian Process Regression model for inputting the output of the ensemble network and outputting a log likelihood.
delete delete The method of claim 2,
Each of the plurality of sub-networks includes a dropout neural network,
Learning the Bayesian network
And learning the Bayesian network so that each of the plurality of sub-networks outputs a plurality of random values for exchange rate prediction using the dropout neural network.
The method of claim 6,
The step of obtaining distribution data for each period by inputting the plurality of real-time foreign exchange data into the Bayesian network
Obtaining, by the distribution data acquisition unit, the plurality of random values from each of the plurality of sub-networks; And
And obtaining, by the distribution data acquisition unit, the distribution data for each period based on the variance of the plurality of random values.
In the exchange rate prediction server,
A collection unit for collecting a data set of a plurality of foreign exchange data and a plurality of real-time foreign exchange data;
A Bayesian network generator that generates a Bayesian network;
A learning unit for learning the Bayesian network to input the data set and output distribution data for each period;
A distribution data acquisition unit for inputting the plurality of real-time foreign exchange data into the Bayesian network to obtain distribution data for each period; And
Exchange rate prediction unit for predicting exchange rates based on the distribution data for each period
That includes, the exchange rate prediction server.
The method of claim 8,
The Bayesian network includes an ensemble network including a plurality of sub-networks for ensemble learning.
The method of claim 9,
The Bayesian network includes a Gaussian Process Regression model for inputting the output of the ensemble network and outputting a log likelihood.
delete delete The method of claim 9,
Each of the plurality of sub-networks includes a dropout neural network,
Wherein the learning unit learns the Bayesian network so that each of the plurality of sub-networks outputs a plurality of random values for predicting the exchange rate using only a dropout nerve.
The method of claim 13,
The distribution data acquisition unit inputs the plurality of real-time foreign exchange data to the Bayesian network to obtain distribution data for each period, obtains the plurality of random values from each of the plurality of sub-networks, and based on the variance of the plurality of random values. To obtain the distribution data for each period by, exchange rate prediction server.

Images