KR20230027903A - Apparatus for predicting fluctuation of stock price based on learning model - Google Patents

Abstract

The present invention relates to an apparatus for predicting fluctuations in stock price based on a learning model. The apparatus for predicting fluctuations in stock price, according to the present invention, comprises: a learning item input unit; a learning data generation module; a learning model unit; an item-to-be-evaluated input unit; an evaluation data generation module; and a prediction classification unit. Therefore, the accuracy of prediction results regarding the fluctuation and extent of stock price trends in a specific period of time in the future can be improved.

Description

Stock price fluctuation prediction device based on learning model {APPARATUS FOR PREDICTING FLUCTUATION OF STOCK PRICE BASED ON LEARNING MODEL}

The present invention relates to a stock price fluctuation prediction device based on a learning model that performs machine learning-based supervised learning for predicting and analyzing the stock price volatility of a stock item that a user intends to trade in a short period of time.

Securities companies are entrusted with large-volume securities orders from institutions such as asset management companies, execute securities transactions, and receive transaction fees.

If a securities company sells or buys a large number of entrusted securities orders at once, the impact of a sharp drop or a sharp rise in stock prices in the stock market occurs, making it difficult to trade at an appropriate price. For this reason, securities companies use an algorithm 　 trading system that divides and executes a large number of securities orders.

However, it is difficult to respond appropriately to the changing stock market because the existing algorithm-trading system divides and executes orders according to pre-set rules.

For example, in the case of a trading system using a TWAP (Time Weighted Average Price) algorithm, since orders are executed with uniformly distributed quantities over time, adaptability to the market and efficiency are inevitably reduced.

Meanwhile, with the recent introduction of IT technology in the financial field, artificial intelligence (AI) technology is being applied to various fields such as sales, marketing, investment management, and trading, and accordingly, the use of robo-advisors is increasing.

RoboAdvisor is a compound word of robot and advisor, which combines IT technologies such as big data, machine learning, and algorithms with financial theories such as modern portfolio theory (MPT) to reflect individual investment propensity. It refers to an online asset management service that automatically composes, rebalances (reconstructs), and operates portfolios.

Currently, most commercial banks and securities companies are providing products or services using robo-advisors, and startup companies that provide robo-advisors are also growing significantly.

However, portfolio services provided using these robo-advisors have structural limitations in that they do not consider various variables that may occur during actual asset management, and the scope of advice is limited to portfolio allocation among financial products such as ETFs. .

KR 10-2021-0003029 A KR 10-1909706 B1

An object of the present invention is to solve the above problems, and to provide a stock price fluctuation prediction device based on a learning model that performs machine learning-based supervised learning for predicting and analyzing the stock price volatility of a stock item to be traded by a user in a short time period. aims to

In order to achieve the above object, an apparatus for predicting stock price fluctuations based on a learning model according to an aspect of the present invention is a learning item input unit that receives a plurality of stock items listed on the stock market as learning items to perform machine learning-based supervised learning And, based on the stock data of the past specific period for the input learning item, a plurality of learning item binary codes representing the change status of each item for each stock item included in the learning item are generated, and then according to a preset order A learning data generation module that generates imaged first learning data by assigning pixel values corresponding to the listed results, and a prediction result for the degree of stock price fluctuation in a specific period in the future based on the generated first learning data. A learning model unit that performs supervised learning using a plurality of pre-set supervised learning values for each class, and an evaluation that receives at least one stock item included in the learning item as an evaluation target item to predict stock price volatility from a user Based on the target item input unit and the securities data of the past specific period for the subject item to be evaluated, a plurality of evaluation item binary codes indicating the change status of each item for the item are generated, and then the results are listed in a predetermined order. An evaluation data generation module that generates imaged first evaluation data by assigning a corresponding pixel value; and when receiving the first evaluation data after the supervised learning is completed, the first evaluation data is converted to the supervised learning value. It is characterized in that it includes a prediction classification unit for classifying into one of the plurality of classes according to the comparison.

Preferably, the learning data generating module provides stock data including at least one of stock price information by time slot, trading volume information, supply and demand trend information, and daily market capitalization ranking information for each of a plurality of stock issues included in the learning issue. A stock data acquiring unit that acquires in real time, and a plurality of bit values consisting of a preset number of bits for each stock item based on the acquired stock data corresponding to the learning item, a preset item code, a preset type code, and the item Learning to create a learning item binary code by allocating a state code indicating a state value according to a comparison according to a current value or a preset standard corresponding to the code or the type code, and a content code indicating a detailed value corresponding to the state code. When a plurality of learning item binary codes are generated by an item binary code generation unit and the learning item binary code generation unit, they are arranged in a predetermined order, and pixels corresponding to the binary values of the plurality of learning item binary codes listed. It is characterized in that it includes a first learning data generation unit that generates imaged first learning data as a value is assigned.

Preferably, the learning data generation module indicates a change in the amount of mention for each stock item based on unstructured data including the amount of keyword mention during a predetermined learning period for each of a plurality of stock items included in the learning item. After generating a plurality of first buzz amount binary codes, pixel values corresponding to the results arranged in a predetermined order are assigned to generate imaged second learning data, and the learning model unit comprises the first learning data or It is characterized in that supervised learning is performed using a plurality of pre-set supervised learning values for each class corresponding to each of the second learning data.

In addition, the learning data generating module includes unstructured data including the amount of keyword mentions for each stock item as a result of searching for a web page on the Internet created during the learning period using a keyword corresponding to each of a plurality of stock items included in the learning item. an unstructured data acquisition unit that obtains in real time, and a plurality of bit values consisting of a preset number of bits for each stock item based on the unstructured data acquired in correspondence with the learning item, a first item code set in advance, and the first item Corresponding to the code, a first status code indicating a state value according to comparison according to a current value or a preset standard and a first content code indicating a detailed value corresponding to the first state code are assigned to the first buzz quantity binary code. When a plurality of the first buzz amount binary codes are generated by a first buzz amount binary code generator that generates a buzz amount and the first buzz amount binary code generator, they are arranged in a predetermined order, and a plurality of the listed first buzz amount binary codes are generated. It is characterized in that it includes a second learning data generation unit that generates imaged second learning data by assigning a pixel value corresponding to the binary value of the 1-buzz binary code.

Preferably, the evaluation data generation module includes a stock data acquisition unit that acquires stock data including at least one of stock price information, trading volume information, supply and demand trend information, and market capitalization ranking information for each time period for the evaluation target stock in real time and a plurality of bit values composed of a preset number of bits based on the securities data acquired corresponding to the evaluation target issue, a preset item code, a preset type code, a current value corresponding to the item code or the type code, or An evaluation item binary code generation unit for generating an evaluation item binary code by allocating a status code representing a status value according to comparison according to a predetermined standard and a content code representing a detailed value corresponding to the status code, and the evaluation item binary code When a plurality of binary codes for evaluation items are generated by the code generation unit, they are arranged in a predetermined order, and pixel values corresponding to the binary values of the plurality of binary codes for evaluation items are assigned to the imaged first evaluation data. It is characterized in that it comprises a first evaluation data generation unit for generating.

Preferably, the evaluation data generation module includes a plurality of second buzz amount binaries representing changes in the amount of mention of the item based on unstructured data including the amount of mention of keywords for a predetermined evaluation period for the item to be evaluated. After generating a code, pixel values corresponding to the results listed in a predetermined order are assigned to generate imaged second evaluation data, and the prediction classifier generates the first evaluation data and the second evaluation data after the supervised learning is completed. When at least one of the second evaluation data is received, the received evaluation data is compared with the supervised learning value and classified into one of the plurality of classes.

In addition, the evaluation data generating module acquires, in real time, unstructured data including the amount of keyword mentions by searching web pages on the Internet created during the evaluation period using keywords corresponding to the evaluation target items. A data acquisition unit converts a plurality of bit values consisting of a preset number of bits based on unstructured data obtained corresponding to the evaluation target item to a preset second item code, a current value corresponding to the second item code, or a preset A second buzz amount binary code for generating a second buzz amount binary code by allocating a second status code indicating a state value according to comparison according to a standard and a second content code indicating a detailed value corresponding to the second state code. When a plurality of second buzz amount binary codes are generated by a generation unit and the second buzz amount binary code generation unit, they are arranged in a predetermined order and correspond to the binary values of the plurality of second buzz amount binary codes listed. It is characterized in that it includes a second evaluation data generation unit that generates imaged second evaluation data by assigning one pixel value.

Preferably, a learning period setting unit for setting a first period and a second period for a predetermined time based on each of the first and second time points that are different from each other prior to the current time point without overlapping, and the input of the evaluation target item It further includes an evaluation period setting unit for setting a third period and a fourth period for a predetermined time based on each of the third and fourth time points prior to the point in time without overlap, and the learning item binary code generation unit and the The evaluation item binary code generation unit generates the learning item binary code or the evaluation item binary code based on the stock data corresponding to at least one stock item input through the learning item input unit or the evaluation item item input unit, A first binary code generated based on the stock price information for each time period collected for a stock item, a second binary code generated based on the trading volume information collected for the stock item, and a supply and demand trend collected for the stock item. It is characterized in that at least one of a third binary code generated based on information and a fourth binary code generated based on the market capitalization ranking collected for the stock item is generated.

In addition, the plurality of classes are set by classifying whether the stock price trend for the fifth period after the current time has risen or fallen and the degree of separation accordingly based on the supervised learning results based on the first period and the second period. and, when the first evaluation data is received after the supervised learning is completed, a probability distribution for each of the plurality of classes based on the first evaluation data is calculated and a clustering unit for clustering according to a preset clustering algorithm is further included. It is characterized by doing.

In addition, based on the clustering result, a trading decision to determine whether to buy or sell the evaluation target item according to a result of comparing the degree of separation from the reference learning value corresponding to the class with the highest similarity with a preset threshold It is characterized in that it further includes wealth.

According to the present invention, when generating learning data for a learning item to perform supervised learning, various index information on the past stock price fluctuation state of stock data and buzz volume information according to various issues are reflected, so that the stock price trend of a specific period in the future There is an effect of improving the accuracy of the predicted result value for the variation and degree.

In addition, according to the present invention, since a binary code value in which necessary information is compressed when generating learning data or evaluation data for each stock item is used, there is an advantage in simplifying and implementing a stock price fluctuation prediction device based on a learning model.

In addition, according to the present invention, when generating learning data used for supervised learning or evaluation data used for predictive analysis, setting values for stock price-related indicators are easily set according to a manager's input when assigning a binary code as a basis. By being able to set and change, there is an effect of providing an asset management portfolio service capable of individual predictive analysis tailored to customers.

1 is a diagram for schematically explaining an apparatus for predicting stock price fluctuations based on a learning model according to an embodiment of the present invention;
2 is a block diagram showing the configuration of an apparatus for predicting stock price fluctuations based on a learning model according to an embodiment of the present invention;
FIG. 3 is a diagram showing the binary codes for learning items or the binary codes for evaluation items generated by the binary code generation unit for learning items or the binary code for evaluation items in FIG. 2 divided into assigned areas;
4 is a diagram showing an example of supervised learning values for each of a plurality of classes used when performing supervised learning in the learning model unit of FIG. 2;
5 is displayed on a graph of stock price and trading volume by hour to explain a method of classifying a method of classifying into a class corresponding to the predicted degree of stock price change in a specific period in the future based on past evaluation data for an item to be evaluated in the prediction classification unit of FIG. 2 is a drawing,
6 is a flowchart illustrating an operation flow of an apparatus for predicting stock price fluctuation based on a learning model according to an embodiment of the present invention.

The specific details, including the problems to be solved, the solutions to the problems, and the effect of the invention for the present invention as described above are included in the embodiments and drawings to be described below. Advantages and features of the present invention, and methods of achieving them, will become clear with reference to the detailed description of the following embodiments taken in conjunction with the accompanying drawings. Like reference numbers designate like elements throughout the specification.

1 is a diagram for schematically explaining an apparatus for predicting stock price fluctuations based on a learning model according to an embodiment of the present invention, and FIG. 2 shows a configuration of an apparatus for predicting stock price fluctuations based on a learning model according to an embodiment of the present invention. FIG. 3 is a diagram showing the learning item binary code or the evaluation item binary code generated by the learning item binary code generator or the evaluation item binary code generator of FIG. It is a diagram showing an example of a plurality of supervised learning values for each class used when supervised learning is performed in the learning model unit of FIG. 2, and FIG. 5 is a future specific period based on past evaluation data for an item to be evaluated in the prediction classification unit of FIG. It is a diagram displayed on the hourly stock price and trading volume graph to explain a method of classifying into a class corresponding to the predicted degree of stock price change, and FIG. Here is a flow chart showing the flow.

Hereinafter, an apparatus for predicting stock price fluctuation based on a learning model according to a preferred embodiment of the present invention will be described with reference to the above-described drawings.

Referring to FIG. 1, the stock price fluctuation prediction apparatus 1000 based on the learning model according to an embodiment of the present invention includes a learning item input unit 10, a learning period setting unit 20, an evaluation target item input unit 30, It is composed of an evaluation period setting unit 40, a learning data generation module 100, a learning model unit 200, an evaluation data generation module 300, and a prediction module 400.

The learning item input unit 10 is for receiving an input of a learning item ( _IL ) to perform machine learning-based supervised learning (S610).

The learning item input unit 10 receives a plurality of stock items listed on the stock market as a learning item ( _IL ) from a manager.

The learning item input unit 10 may be connected to the manager terminal 1 by data communication, and receives the result of the manager selecting and inputting from all stock items currently listed on the stock market from the manager terminal 1 and receiving the received selection The stock item corresponding to the input can be input as the learning item ( _IL ).

The learning period setting unit 20 specifies the acquisition period of unstructured data and stock data that is the basis for generating learning data for performing supervised learning for the learning item ( _IL ) input by the learning item input unit 10 to set the learning period (P _S ).

In setting the learning period (P _S ), the learning period setting unit 20 sets the first time period (T ₁ ) and the second time period (T ₂ ) prior to the current time period during a preset time based on each other. The first period (P ₁ ) and the second period (P ₂ ) are set without overlapping, but the second period (P ₂ ) may be set to have a predetermined time interval after the first period (P ₁ ).

The learning period setting unit 20 can newly set and update the learning period (P _S ) at every predetermined time period so that supervised learning can be repeatedly performed using a large amount of learning data for the learning item ( _IL ). there is.

The evaluation target item input unit 30 is for receiving an evaluation target item (I _E ) to predict stock price volatility (S640).

The evaluation target item input unit 30 inputs at least one stock item among a plurality of stock items included in the learning item ( _IL ) input by the learning item input unit 10 as the evaluation target item (I _E ) from the user. receive

The evaluation target item input unit 30 may be connected to the user terminal 3 by data communication, and the user selects and inputs a result from a plurality of stock items included in the learning item _IL from the user terminal 3. Thus, the stock issue corresponding to the received selection input can be input as the evaluation target issue (I _E ).

The evaluation period setting unit 40 sets the evaluation period ( _PE ) by specifying the acquisition period of unstructured data and securities data that are the basis for generating evaluation data for the evaluation target item (I _E ). It may be to perform the same function as the period setting unit 20 .

In setting the evaluation period ( _PE ), the evaluation period setting unit 40 sets the evaluation target item (I _E ) at a different third time point (T ₃ ) and fourth time point (T C ) prior to the input time point (T _C ). ₄ ) Set the third period (P ₃ ) and the fourth period (P ₄ ) without overlapping during the predetermined time based on each, but the fourth period (P ₄ ) is the third period (P ₃ ). You can set a time interval.

At this time, the time interval from the first time point (T ₁ ) and the second time point (T ₂ ) based on the learning period ( _PS ) to the current time point and the third time point (T 3 ) based on the evaluation period ( _PE ) and the second time point (T ₃ ) The time interval from the 4th time point (T ₄ ) to the input time point (T _C ) of the evaluation target item (I _E ) can be set identically.

The learning data generation module 100 serves to generate learning data based on supervised learning based on the learning item ( _IL ) and the learning period ( _PS ) (S620).

The learning data generation module 100 generates at least one of first learning data based on stock data reflecting actual stock price fluctuations and second learning data based on unstructured data reflecting investor sentiment according to various issues. .

Specifically, the learning data generation module 100 is based on the stock data of the past specific period for the learning item ( _IL ) input by the learning item input unit 10, each stock item included in the learning item ( _IL ) Imaged first learning data may be generated by generating a plurality of learning item binary codes representing the change state for each item and then assigning pixel values corresponding to the results arranged in a predetermined order.

In addition, the learning data generating module 100 is a keyword reference amount for a predetermined learning period (P _S ) for each of a plurality of stock items included in the learning item ( _IL ) input by the learning item input unit 10 . After generating a plurality of first buzz amount binary codes representing the trend of change in the amount of mention for each stock item based on unstructured data including data can be generated.

In this regard, the learning data generation module 100, as shown in FIG. 2, includes a securities data acquisition unit 110 involved in generating the first learning data, a learning item binary code generation unit 120, and the first learning data It may include a generator 130, an unstructured data acquisition unit 140 involved in generating second learning data, a first buzz amount binary code generator 150, and a second training data binary code generator 160. can

The stock data acquisition unit 110 provides stock data including at least one of stock price information, trading volume information, supply and demand trend information by time slot, and daily market capitalization ranking information for each of a plurality of stock items included in the learning item ( _IL ). is obtained in real time.

The stock data acquisition unit 110 may acquire stock data for each stock item in real time using a web scraping technique based on the Python language by accessing a corresponding stock office server through a plurality of stock company APIs. .

The learning item binary code generation unit 120 generates stock price fluctuation state information for each stock item in the form of a binary code based on the acquired stock data corresponding to the learning item ( _IL ).

The learning item binary code generation unit 120 converts a plurality of bit values consisting of a preset number of bits for each stock item into a preset item code, a preset type code, a current value or a preset value corresponding to the item code or the type code. Binary codes for learning items can be generated by assigning a state code indicating a state value according to comparison according to a standard and a content code indicating a detailed value corresponding to the state code.

Here, the learning item binary code generation unit 120 generates a learning item binary code based on stock data corresponding to each of a plurality of stock items included in the learning item ( _IL ) input through the learning item input unit 10 However, the first binary code generated based on the stock price information for each time period collected for the stock item, the second binary code generated based on the trading volume information collected for the stock item, and the collection for the stock item At least one of a third binary code generated based on the collected supply and demand trend information and a fourth binary code generated based on a market capitalization ranking collected for the stock item may be generated.

At this time, as shown in FIG. 3, the learning item binary code generation unit 120 converts a plurality of data areas except for the header area 50 into a first binary code area 51 according to a preset number of bits, the first A second binary code area 52 allocated to a lower bit area than the binary code area 51, a third binary code area 53 allocated to a lower bit area than the second binary code area 52, Binary code values corresponding to the plurality of data areas may be generated by dividing into a fourth binary code area 54 allocated to a lower bit area than the binary code area 53 .

In this case, the first binary code is generated in the first binary code area 51, the second binary code is generated in the second binary code area 52, and the third binary code is generated in the third binary code area 53. A code may be generated and the fourth binary code may be generated in the fourth binary code area 54 .

Here, the first binary code uses stock price information for each time period for each stock item acquired by the stock data acquisition unit 110 corresponding to the learning item ( _IL ), as shown in Table 1 below, Trend direction information based on a moving average for each stock item for each predetermined unit is assigned to the status code, and the unit division value and reference unit value of the moving average line are respectively assigned to the type code and the content code. may be assigned.

At this time, the moving average line represents a line created by sequentially connecting stock price moving averages, which are arithmetic averages of stock prices over a certain period of time. Therefore, a detailed description thereof will be omitted.

item code type code status code content code unit value more below moving average period momentum
(1100) minute
city
Day
main
month 001
010
011
101
110 001 010 Reference unit value for the moving average to be referenced

Here, the second _binary code is the learning period (P _S ) The first average transaction volume ( _V1 ), which is the average of the accumulated transaction volumes during the first period (P 1 ) belonging to the learning period (P S ), and the second average, which is the average of the accumulated transaction volumes during the second period (P ₂ ) belonging to the learning period (P _S ) Based on the transaction volume (V2), the state value and the detailed value corresponding to whether and the degree of increase or decrease of the second average transaction volume (V2) compared to the first average transaction volume (V1) may be assigned to the status code and the content code, respectively. there is.

item code type code status code 101 000 001 trading volume
(0001) Not applicable
(000) V1 > V2 V1 = V2
(Drop by -1 or rise by less than 1) V1 < V2

Here, the third binary code corresponds to the learning item ( _IL ) by using the supply and demand trend information for each time period acquired by the stock data acquisition unit 110, as shown in Table 3 below, from a predetermined point in the past. Status values and detailed values corresponding to occurrence and degree of continuous buying or selling during the period S up to the present time may be assigned to the status code and the content code, respectively.

At this time, Ps ₁ and Ps ₂ represent a first supply and demand period and a second supply and demand period that are set differently from each other, and G represents a preset percentage.

item code type code status code subject name value 101 000 001 supply and demand I
(0101) Agency
foreigner
pension fund 001
010
011 S > Ps ₁
(continuous selling) Not applicable S > Ps ₁
(continuous purchase) supply and demand II
(0110) Agency
foreigner
pension fund 001
010
011 S ⊂ Ps ₂
(Sell above G) Not applicable S ⊂ Ps ₂
(Buy more than G)

Here, the fourth binary code corresponds to the learning item ( _IL ), using the daily total ranking information obtained by the stock data acquisition unit 110, as shown in Table 4 below, a plurality of predetermined rankings Among the ranges, a status value corresponding to a ranking range to which the stock item belongs may be assigned as the status code, and market type information in which the stock item is listed may be assigned as the type code.

item code
type code status code market value 000 001 010 011 Market cap ranking
(1001) KOSPI
KOSDAQ 001
010 1st scope
(301st and below) 2nd scope
(101st - 300th) 3rd scope
(51st to 100th) 4th scope
(within top 50)

Table 5 below shows an example of each of the first to fourth binary codes generated as described above in a table.

item code type code status code content code 1100
(momentum) 011 001 0101 0001
(trading volume) 000 001 0101 0101
(supply and demand I) 001 101 0101 0110
(supply and demand II) 001 101 0111 1001
(Market cap ranking) 001 001 0000

For example, referring to Table 5, the first binary code includes item code "1100" corresponding to 'momentum', type code "011" indicating that the unit division value of the moving average line is 'day (unit)', It may include the status code "001" indicating that the trend direction of the current stock price is higher than the moving average line (uptrend), and the content code "0101" indicating that the reference unit value of the moving average line is '5 (day)', This indicates that the current stock price is above the 5-day moving average.

In addition, the second binary code includes an item code "0001" corresponding to 'transaction volume', a status code "001" indicating a state in which the first average transaction volume (V1) is smaller than the second average transaction volume (V2), and the average of the two It may include content code "0101" indicating that the difference in trading volume is '5[%]', which indicates a case in which the average trading volume of the second period is increased by 5% compared to the average trading volume of the first period.

In addition, when the first supply and demand period (P _S1 ) is set to '3 days', the third binary code is the item code "0101" corresponding to 'supply and demand I' and the type indicating that the supply and demand subject is an 'institution'. The code "001", the status code "101" indicating the state of continuous selling during the supply and demand period (S) greater than the preset first supply and demand period (P _S1 ), and the detailed value of the supply and demand period (S) are '5 (days) )', which indicates that the institution has sold for 5 consecutive days.

In addition, when the second supply and demand period (P _S2 ) is '10 days' and the reference percentage (G) is preset to '50%', the third binary code is item code "0110" corresponding to 'supply and demand II' and "001", a type code indicating that the supply and demand subject is an 'institution', and a status code "101" indicating a state of continuous sales of 50% or more during a supply and demand period (S) greater than the preset second supply and demand period (P _S2 ) , The content code "0111" indicating that the detailed value of the supply period (S) is '7 (day)', which indicates that the institution has sold 7 days out of the last 10 days.

In addition, the fourth binary code includes an item code "1001" corresponding to 'market capitalization ranking', a type code "001" indicating that the type of market in which stock item A is listed is 'KOSPI', and a ranking to which stock item A belongs. The status code "001" indicating that the range is 'second range' may be included, which indicates a case where the market capitalization ranking of KOSPI A stock is 101 to 300.

When a plurality of the learning item binary codes are generated by the learning item binary code generator 120, the first learning data generation unit 130 lists them in a predetermined order, and the binary code of the plurality of the learning item binary codes listed. By assigning a pixel value corresponding to the value, imaged first learning data is generated.

Here, the first learning data generation unit 130 divides the learning item binary code into predetermined bit units to represent one pixel per unit, so that an image that meets the specifications of the input layer of the learning model, That is, first learning data in the form of an image having the same ratio as the ratio of rows and columns of the matrix of the input layer may be generated.

The unstructured data acquisition unit 140 searches web pages on the Internet created during the learning period ( _PS ) using keywords corresponding to each of a plurality of stock items included in the learning item ( _IL ), and keywords for each stock item. Acquire unstructured data including reference amount in real time.

The unstructured data acquisition unit 140 obtains a keyword string corresponding to each stock item through a web crawling and parsing technique according to a query-based approach using a predetermined programming language, for example, Python. It may be to acquire unstructured data in real time according to calculating the amount of keyword mentions as a result of performing a web search based on the web search.

The first buzz amount binary code generation unit 150 generates buzz amount information according to an issue for each stock item in the form of a binary code based on unstructured data obtained in correspondence with the learning item _IL .

The first buzz amount binary code generation unit 150 generates a plurality of bit values consisting of a preset number of bits for each stock item according to a preset first item code, a current value corresponding to the first item code, or a preset standard. The first buzz amount binary code may be generated by allocating a first state code indicating a state value according to comparison and a first content code indicating a detailed value corresponding to the first state code.

When a plurality of first buzz amount binary codes are generated by the first buzz amount binary code generator 150, the second learning data generation unit 160 arranges them in a predetermined order, and lists the plurality of first buzz amount binary codes listed. Imaged second learning data is generated by assigning a pixel value corresponding to the binary value of the buzz amount binary code.

Here, the second learning data generation unit 160 divides the first buzz amount binary code into predetermined bit units to represent one pixel per unit, thereby meeting the specifications of the input layer of the learning model. Second learning data in the form of an image, that is, an image having the same ratio as the ratio of rows and columns of the matrix of the input layer may be generated.

The learning model unit 200 performs supervised learning to predict the degree of stock price change of a specific future period of the learning item ( _IL ) based on the learning data generated by the learning data generation module 100 (S630). .

The learning model unit 200 is a learning item (I) based on the first learning data generated by the first learning data generator 130 or the second learning data generated by the second learning data generator 160, respectively. It is possible to perform supervised learning using a plurality of pre-set supervised learning values for each class to correspond to the predicted result of the future stock price fluctuation of _L ).

The learning model unit 200 includes an input layer, a plurality of convolution layers that perform convolution operations, a plurality of pooling layers that reduce the size of the calculated data, and non-linearly transforming the pooled data for each class. It may include a CNN learning model based on a convolution neural network (CNN) consisting of a neural network structure including a plurality of activation function layers to facilitate classification.

In this case, when the learning model unit 200 receives image-type learning data and a corresponding supervised learning value, it may be learned as the parameters of the neural network structure are arranged.

Here, the plurality of classes, based on the results of supervised learning based on the first period (P ₁ ) and the second period (P ₂ ), the rise or fall of the stock price trend for the fifth period (P _F ) after the current point in time It may be set by classifying whether or not and the degree of separation accordingly.

Table 6 below shows the direction of the trend and the degree of separation for the supervised learning values corresponding to each of the plurality of classes.

supervised learning value Fifth Period Data Forecast Trends degree of separation One C1 degradation 3rd scope 2 C2 degradation 2nd scope 3 C3 degradation 1st range 4 C4 bohab ­ 5 C5 Increase 1st range 6 C6 Increase 2nd range 7 C7 Increase 3rd scope

For example, referring to Table 6 and FIG. 4, the first supervised learning value C1 is currently based on the supervised learning result based on the past specific period (the first period P ₁ and the second period P ₂ ). Corresponds to a case where the stock price trend for the fifth period (P _F ) after the point in time is predicted to fall within the third range, and the fourth supervised learning value (C4) is the fifth period (P F ) based on the supervised learning result. _F ) corresponds to a case where the stock price trend is predicted to remain flat without fluctuation, and the sixth supervised learning value (C6) is the stock price trend for the fifth period (P _F ) based on the supervised learning result. It may be set to correspond to the case where it is predicted to rise within the range.

The evaluation data generation module 300 serves to generate evaluation data based on an evaluation target issue (I _E ) and an evaluation period ( _PE ) for which stock price volatility is to be predicted (S650).

The evaluation data generation module 300 generates at least one of first evaluation data based on stock data reflecting actual stock price fluctuations and second evaluation data based on unstructured data reflecting investor sentiment according to various issues.

At this time, the evaluation data generation module 300 may perform the same function as the learning data generation module 100 described above, only with different values of stock data and unstructured data.

Specifically, the evaluation data generation module 300 represents the change status of each item for the item based on the stock data of the past specific period for the evaluation target item (I _E ) input by the evaluation target item input unit 30 Imaged first evaluation data may be generated by generating a plurality of evaluation item binary codes and then assigning pixel values corresponding to the results arranged in a predetermined order.

In addition, the evaluation data generation module 300 indicates the change in the amount of mention of the item on the basis of unstructured data including the amount of keyword mention for a predetermined evaluation period (PE ₎ for the item to be evaluated (I _E ). After generating a plurality of second buzz amount binary codes, pixel values corresponding to the results arranged in a predetermined order are assigned, thereby generating imaged second evaluation data.

In this regard, the evaluation data generation module 300 includes a securities data acquisition unit 310 involved in generating the first evaluation data, an evaluation item binary code generation unit 320, and an evaluation data generation unit, as shown in FIG. 330, an unstructured data acquisition unit 340 involved in generating second evaluation data, a second buzz amount binary code generator 350, and a second evaluation data binary code generator 360. .

The stock data acquisition unit 310 acquires stock data including at least one of stock price information, trading volume information, supply and demand trend information, and market capitalization ranking information for each time slot for the stock to be evaluated (I _E ) in real time.

The securities data acquisition unit 310 accesses the securities company server through a plurality of securities company APIs and uses a web scraping technique based on the Python language to acquire stock data for the evaluation target item (I _E ) in real time it may be

The evaluation issue binary code generation unit 320 generates stock price change state information for the evaluation target issue in the form of a binary code based on the acquired stock data corresponding to the evaluation target issue (I _E ).

The evaluation item binary code generator 320 compares a plurality of bit values composed of a preset number of bits according to a preset item code, a preset type code, a current value corresponding to the item code or the type code, or a preset standard. The evaluation item binary code can be generated by assigning a state code indicating a state value according to the process and a content code indicating a detailed value corresponding to the state code.

Here, the evaluation item binary code generation unit 320 generates the evaluation item binary code based on the securities data corresponding to the evaluation target item (I _E ) input through the evaluation target item input unit 30, and the evaluation item binary code is generated. A first binary code generated based on stock price information for each time period collected for the item (I _E ), and a second binary code generated based on the trading volume information collected for the item (I _E ) to be evaluated; Of the third binary code generated based on the collected supply and demand trend information for the evaluation target item (I _E ) and the fourth binary code generated based on the market capitalization ranking collected for the evaluation target item (I _E ) It may be to create at least one.

At this time, the evaluation item binary code generation unit 320 may generate the evaluation item binary code according to the same binary code generation algorithm as the above-described learning item binary code generation unit 120, and a detailed description of the corresponding binary code Since is redundant with the above-described binary code for the learning item, it is omitted.

When a plurality of evaluation item binary codes are generated by the evaluation item binary code generator 320, the first evaluation data generation unit 330 arranges them in a predetermined order, and the binary codes of the plurality of evaluation item binary codes listed. By assigning a pixel value corresponding to the value, imaged first evaluation data is generated.

Here, the first evaluation data generation unit 330 divides the evaluation item binary code into predetermined bit units to represent one pixel per unit, so that an image that meets the specifications of the input layer of the learning model, That is, first evaluation data in the form of an image having the same ratio as the ratio of rows and columns of the matrix of the input layer may be generated.

The unstructured data acquisition unit 340 retrieves unstructured data including the amount of keyword mentions as a result of searching web pages on the Internet created during the evaluation period ( _PE ) using keywords corresponding to the evaluation target item (I _E ) in real time. Acquire with

The unstructured data acquisition unit 340 obtains a keyword string corresponding to each stock item through a web crawling and parsing technique according to a query-based approach using a predetermined programming language, for example, Python. It may be to acquire unstructured data in real time according to calculating the amount of keyword mentions as a result of performing a web search based on the web search.

The second buzz quantity binary code generation unit 350 converts buzz quantity information according to issues related to the evaluation target item (I _E ) into a binary code form based on unstructured data obtained in correspondence with the evaluation target item (I _E ). generate

The second buzz amount binary code generation unit 350 compares a plurality of bit values composed of a preset number of bits according to a preset second item code, a current value corresponding to the second item code, or a preset standard. A second buzz amount binary code may be generated by allocating a second status code representing a value and a second content code representing a detailed value corresponding to the second status code.

The second evaluation data generation unit 360, when a plurality of second buzz amount binary codes are generated by the second buzz amount binary code generation unit 350, arranges them in a predetermined order and sets the plurality of second buzz Imaged second evaluation data is generated by assigning a pixel value corresponding to the binary value of the quantity binary code.

Here, the second evaluation data generation unit 360 divides the second buzz amount binary code into predetermined bit units to represent one pixel per unit, thereby meeting the specifications of the input layer of the learning model. Second evaluation data in the form of an image, that is, an image having the same ratio as the ratio of rows and columns of the matrix of the input layer may be generated.

The prediction module 400 is an evaluation target item corresponding to the evaluation data generated by the evaluation data generation module 300 based on the result of supervised learning using the learning data generated by the learning data generation module 100 (I _E ) It is to predict stock price volatility for .

In this regard, the prediction module 400 may include a prediction classification unit 420, a clustering unit 440, and a trading determination unit 460 as shown in FIG. 2 .

When the prediction classification unit 420 receives predetermined evaluation data from the evaluation data generation module 300 after the supervised learning is completed, it is classified into one of the plurality of classes by comparing it with the supervised learning value (S660).

After the supervised learning is completed, the prediction classification unit 420 selects at least one of the first evaluation data generated by the first evaluation data generator 330 and the second evaluation data generated by the second evaluation data generator 360. When one input is received, the input evaluation data may be classified into one of the plurality of classes by comparing the input evaluation data with the supervised learning value for each class.

For example, referring to FIG. 5 , when the time point at which the evaluation target item (I _E ) is input from the user through the evaluation target item input unit 30 is "T _C ", the first evaluation data of the evaluation data generating module 300 Securities data and unstructured data corresponding to the third period (P ₃ ) and the fourth period (P ₄ ) before the input time point (T _C ), respectively, by the generation unit 330 and the second evaluation data generation unit 360 When the first evaluation data or the second evaluation data based thereon is generated, the prediction classification unit 420 receives the first evaluation data or the second evaluation data and compares them with a plurality of preset teaching learning values for each class, and as a result, the most similar The supervised learning value can be output as a classification value ( _CP ) representing the predicted trend and degree of the 5th period (P _F ) of the evaluation target item (I _E ).

The clustering unit 440 selects at least one of the first evaluation data generated by the first evaluation data generator 330 and the second evaluation data generated by the second evaluation data generator 360 after the supervised learning is completed. When is received, a probability distribution for each of the plurality of classes based on the input evaluation data is calculated and clustered according to a preset clustering algorithm (S670).

The clustering unit 440 may estimate the probability distribution state of each class for the input evaluation data and cluster based on the similarity according to the comparison result with the supervised learning value of each class. For example, the clustering algorithm may perform K- Means clustering, Mean Shift clustering, Gaussian Mixture Model (GMM), and Density Based Clustering (DBSCAN) may be included.

Based on the result of the clustering by the clustering unit 440, the trading decision unit 460 compares the degree of separation with respect to the reference learning value corresponding to the class having the highest degree of similarity with a preset threshold, and determines the item to be evaluated ( It is determined whether to buy or sell I _E ) (S680).

When the degree of separation from the reference learning value is higher than the threshold value, the trading decision unit 350 determines the evaluation target item (I _E ) when the trend direction of the evaluation target item (I _E ) based on the reference learning value shows an upward trend. It can be decided to buy a certain number of.

When the degree of separation from the reference learning value is higher than the threshold value, the trading decision unit 350 determines the evaluation target item (I _E ) when the trend direction of the evaluation target item (I _E ) based on the reference learning value shows a downward trend. may decide to sell a certain number of

Meanwhile, the apparatus for predicting stock price fluctuations based on the learning model according to the present invention may further include a standard setting unit 60 as shown in FIG. 1 .

The standard setting unit 60 sets the interval between the first period (P ₁ ) and the second period (P ₂ ) set by the learning period setting unit 20 according to the manager's input, or sets the first period (P ₁ ) and the second period P ₂ , an interval between the first time point T ₁ and the second time point T ₂ serving as a reference may be set.

The standard setting unit 60 allows the administrator to set the interval between the first time point (T ₁ ) and the second time point (T ₂ ) or the first period (P ₁ ) and the second period (P ₂ ) through the manager terminal (1). A value for setting an interval between values may be received from the manager terminal 1 and transmitted to the learning period setting unit 20 and the evaluation period setting unit 40 .

In this case, the learning period setting unit 20 determines the interval between the first time point (T ₁ ) and the second time point (T ₂ ) or the first period (P ₁ ) and the second period ( P ₂ ), and the evaluation period setting unit 40 determines the interval between the third time point T ₃ and the fourth time point T ₄ or the third period P according to the received interval setting value. ₃ ) and the fourth period P ₄ may be set.

The standard setting unit 60 is a learning item binary code generator 120 or an evaluation item binary code generator 320 based on the supply and demand trend information collected in response to the learning item ( _IL ) or the evaluation target item (I _E ). ), the first supply and demand period (P _S1 ) and the second supply and demand period (P _S2 ), which are comparison standards when allocating the third binary code generated by .

The criterion setting unit 60 is a learning item binary code generator 120 or an evaluation item binary code generator 320 based on the stock price information collected in response to the learning item ( _IL ) or the evaluation target item (I _E ). A reference unit value for a moving average line serving as a reference when allocating the first binary code generated by may be set according to a manager's input.

Accordingly, according to the present invention described above, when generating learning data for learning items to be supervised learning, various index information on the past stock price fluctuation state of stock data and buzz information according to various issues are reflected, so that the future It is possible to improve the accuracy of prediction result values for the change and degree of stock price trend in a specific period.

In addition, according to the present invention, since a binary code value in which necessary information is compressed when generating learning data or evaluation data for each stock issue is used, a stock price fluctuation prediction device based on a learning model can be simplified and implemented.

In addition, according to the present invention, when generating learning data used for supervised learning or evaluation data used for predictive analysis, setting values for stock price-related indicators are easily set according to a manager's input when assigning a binary code as a basis. By being able to set and change, it is possible to provide asset management portfolio services that enable individual predictive analysis tailored to customers.

Above, the present invention has been described in detail through preferred embodiments, but the present invention is not limited thereto and may be variously practiced within the scope of the claims.

In particular, the foregoing has outlined rather broadly the features and technical strengths of the present invention so that the claims of the invention to be described later may be better understood, so that the concept and specific embodiments of the present invention described above are intended to serve similar purposes to the present invention. It should be recognized by those skilled in the art that it can be readily used as a basis for designing or modifying other shapes for the purpose.

In addition, the embodiment described above is only one embodiment according to the present invention, and can be implemented in various modified and changed forms within the scope of the technical idea of the present invention by those skilled in the art. You will understand. Therefore, the disclosed embodiments should be considered from an explanatory point of view rather than a limiting point of view, and these various modifications and changes are also shown in the claims of the present invention described above as belonging to the scope of the technical idea of the present invention, and the scope equivalent thereto. Any differences within them should be construed as being included in the present invention.

10: learning item input unit 20: learning period setting unit
30: evaluation target item input unit 40: evaluation period setting unit
60: standard setting unit 100: learning data generation module
110: Stock data acquisition unit 120: Learning item binary code generation unit
130: first learning data generation unit 140: unstructured data acquisition unit
150: first buzz quantity binary code generator 160: second learning data generator
200: learning model unit 300: evaluation data generation module
310: Stock data acquisition unit 320: Evaluation item binary code generation unit
330: first evaluation data generation unit 340: unstructured data acquisition unit
350: second buzz quantity binary code generator 360: second evaluation data generator
400: prediction module 420: prediction classification unit
440: clustering unit 460: trading decision unit
1000: stock price fluctuation predictor

Claims (10)

A learning item input unit that receives a plurality of stock items listed on the stock market as learning items to perform machine learning-based supervised learning;
Based on the stock data of the past specific period for the input learning item, a plurality of learning item binary codes representing the change status of each item for each stock item included in the learning item are generated and then listed in a predetermined order. A learning data generation module for generating imaged first learning data by assigning a pixel value corresponding to ;
a learning model unit that performs supervised learning using a plurality of pre-set supervised learning values for each class, based on the generated first learning data;
an evaluation target input unit for receiving at least one stock issue included in the learning issue as an evaluation target issue for which stock price volatility is to be predicted from a user;
Based on the stock data of the past specific period for the received evaluation target item, a plurality of evaluation item binary codes representing the change status of each item for the item are generated, and pixel values corresponding to the results listed in a predetermined order are generated. Evaluation data generation module for generating the imaged first evaluation data according to the grant; and
When receiving the first evaluation data after the supervised learning is completed, a prediction classification unit that classifies the first evaluation data into one of the plurality of classes according to comparing the first evaluation data with the supervised learning value; Stock price fluctuation predictor based on learning model. According to claim 1,
The learning data generation module,
a stock data acquiring unit that acquires stock data including at least one of stock price information by time period, trading volume information, supply and demand trend information, and daily market capitalization ranking information for each of a plurality of stock items included in the learning item;
Based on the stock data obtained in correspondence with the learning item, a plurality of bit values consisting of a preset number of bits for each stock item is a preset item code, a preset type code, a current value corresponding to the item code or the type code or a learning item binary code generation unit for generating a learning item binary code by allocating a state code indicating a state value according to a comparison according to a predetermined criterion and a content code indicating a detailed value corresponding to the state code; and
When a plurality of binary codes for learning items are generated by the learning item binary code generation unit, they are arranged in a predetermined order, and pixel values corresponding to the binary values of the listed plurality of binary codes for learning items are given. A stock price fluctuation prediction device based on a learning model comprising a; first learning data generation unit for generating first learning data. According to claim 1,
The learning data generation module,
Based on unstructured data including the amount of keyword mention during a predetermined learning period for each of the plurality of stock items included in the learning item, a plurality of first buzz amount binary codes representing the change in the amount of mention for each stock item are generated Then, by assigning pixel values corresponding to the results listed in a predetermined order, imaged second learning data is generated,
The learning model unit,
A stock price fluctuation prediction device based on a learning model, characterized in that performing supervised learning using a plurality of pre-set supervised learning values for each class corresponding to each of the first learning data or the second learning data. According to claim 3,
The learning data generation module,
An unstructured data acquisition unit that obtains, in real time, unstructured data including the amount of keyword mentions for each stock item by searching web pages on the Internet created during the learning period using keywords corresponding to each of a plurality of stock items included in the learning item. ;
Based on the unstructured data obtained in correspondence with the learning item, a plurality of bit values consisting of a preset number of bits for each stock item are converted to a preset first item code, a current value corresponding to the first item code, or a preset standard. A first buzz amount binary code generator for generating a first buzz amount binary code by allocating a first state code indicating a state value according to comparison and a first content code indicating a detailed value corresponding to the first state code, respectively. ; and
When a plurality of the first buzz amount binary codes are generated by the first buzz amount binary code generator, they are arranged in a predetermined order, and pixel values corresponding to the binary values of the plurality of the first buzz amount binary codes are arranged. A stock price fluctuation prediction device based on a learning model comprising a; According to claim 1,
The evaluation data generation module,
a stock data acquisition unit that acquires stock data including at least one of stock price information, trading volume information, supply and demand trend information, and market capitalization ranking information for each time slot for the subject stock to be evaluated in real time;
A plurality of bit values composed of a preset number of bits based on the securities data acquired corresponding to the evaluation target issue are a preset item code, a preset type code, a current value or a preset value corresponding to the item code or the type code. an evaluation item binary code generation unit for generating an evaluation item binary code by allocating a status code indicating a status value according to comparison according to a standard and a content code indicating a detailed value corresponding to the status code; and
When a plurality of evaluation item binary codes are generated by the evaluation item binary code generation unit, they are arranged in a predetermined order, and pixel values corresponding to the binary values of the listed plurality of evaluation item binary codes are assigned. A stock price fluctuation prediction device based on a learning model, comprising: a first evaluation data generating unit for generating first evaluation data. According to claim 1,
The evaluation data generation module,
Based on unstructured data including the amount of keyword mentions for a predetermined evaluation period for the evaluation target item, a plurality of second buzz amount binary codes representing the change in the amount of mention of the item are generated and arranged in a predetermined order By assigning a pixel value corresponding to the result, imaged second evaluation data is generated,
The prediction classification unit,
When at least one of the first evaluation data and the second evaluation data is received after the supervised learning is completed, the input evaluation data is compared with the supervised learning value and classified into one of the plurality of classes. Stock price fluctuation prediction device based on a learning model to be. According to claim 6,
The evaluation data generation module,
a second buzz amount data acquisition unit that acquires, in real time, unstructured data including the amount of references to keywords as a result of searching for web pages on the Internet created during the evaluation period using keywords corresponding to the subject items to be evaluated;
A plurality of bit values composed of a preset number of bits based on the unstructured data obtained corresponding to the evaluation target item are compared according to a preset second item code, a current value corresponding to the second item code, or a preset standard a second buzz amount binary code generation unit for generating a second buzz amount binary code by allocating a second status code indicating a state value corresponding to a second state code and a second content code indicating a detailed value corresponding to the second state code; and
When a plurality of second buzz amount binary codes are generated by the second buzz amount binary code generation unit, they are arranged in a predetermined order, and pixel values corresponding to the binary values of the plurality of second buzz amount binary codes are arranged. A stock price fluctuation prediction device based on a learning model comprising: a second evaluation data generation unit that generates imaged second evaluation data according to the application. According to claim 1,
a learning period setting unit for setting a first period and a second period for a predetermined time based on different first and second times prior to the current time, respectively, without overlap; and
An evaluation period setting unit for setting a third period and a fourth period for a predetermined time based on different third and fourth times points before the input point of the evaluation target item, respectively, without overlap,
The learning item binary code generator and the evaluation item binary code generator,
The learning item binary code or the evaluation item binary code is generated based on the stock data corresponding to at least one stock item input through the learning item input unit or the evaluation target item input unit, and the time collected for the stock item A first binary code generated based on stock price information for each unit, a second binary code generated based on trading volume information collected for the stock item, and a second binary code generated based on supply and demand trend information collected for the stock item. 3 binary code and a fourth binary code generated based on the market capitalization ranking collected for the stock issue. According to claim 8,
The plurality of classes,
Based on the supervised learning results based on the first period and the second period, whether the stock price trend for the fifth period after the current point of time rises or falls and the degree of separation accordingly is classified and set,
When the first evaluation data is received after the supervised learning is completed, a clustering unit that calculates a probability distribution for each of the plurality of classes based on the first evaluation data and clusters them according to a preset clustering algorithm; further comprising Stock price fluctuation prediction device based on the learning model, characterized in that. According to claim 9,
a trading decision unit that determines whether to buy or sell the evaluation target item according to a result of comparing the degree of separation with respect to the reference learning value corresponding to the class having the highest similarity with a preset threshold based on the clustered result; Stock price fluctuation prediction device based on the learning model, characterized in that it further comprises.

Images