KR101458004B1 - System and method for predicting change of stock price using artificial neural network model - Google Patents

Abstract

Disclosed are a system for predicting the fluctuations of stock prices using an artificial neural network model and a method thereof. According to the present invention, the system for predicting the fluctuations of stock prices using an artificial neural network model includes: an input unit which receives a plurality of data related to a stock market as input data; a modeling unit which assigns the inputted stock market-related data to a plurality of first nodes, applies a first weight to each data assigned to the first nodes, and assigns the first weight-applied data to a plurality of second nodes, thereby modeling an artificial neural network model; and an output unit which applies a second weight to the data assigned to the second nodes to compute a weighted sum, inputs the data to an activation function to calculate a single value, compares the single value with a pre-defined threshold, outputs a high value as an output value if the single value is greater than the threshold, and outputs a low value as an output value if the single value is smaller than the threshold.

Description

TECHNICAL FIELD The present invention relates to a system and method for predicting stock price fluctuation using an artificial neural network model,

The present invention relates to a stock price fluctuation predicting system and a stock price predicting method using an artificial neural network model, and more particularly, to a stock price fluctuation predicting system and a stock price predicting method using an artificial neural network model .

The stock market has its own pricing mechanism, and this pricing mechanism is influenced by myriad variable data. In other words, the fundamentals that affect stock prices are so diverse and complex that these factors fluctuate from time to time, so that past cycle fluctuations affect the current share price.

Recently, due to the development of Internet technology, almost all information newly created by human beings is spread, shared and utilized through the Internet. In addition, since the information available through the Internet is digitized information, it can be readily utilized as input data for stock price prediction. In order to predict the stock price fluctuation by using a lot of data shared on the Internet, so-called big data, a big data analyzing using a statistical model is performed together.

However, the prediction of stock price through the conventional big data analysis can not sufficiently consider the influence between input data. Therefore, there is a need for a new kind of statistical model in order to improve the reliability of the stock price fluctuation prediction result.

According to the above-described necessity, the stock price fluctuation prediction system using the artificial neural network model according to the present invention models a neural network model as a computer-executable artificial neural network model, and then, in order to improve the predictability of the input data, And to provide a stock price fluctuation predicting system and a stock price fluctuation predicting method using a model.

According to an aspect of the present invention, there is provided a system for predicting a stock price change using an artificial neural network model, including an input unit for inputting a plurality of stock-related data as input data, A modeling unit for modeling an artificial neural network model by assigning a first weight to data allocated to a plurality of nodes and assigning a first weight to data allocated to the first plurality of nodes to a second plurality of nodes; And outputs the result as a single value. If the single value is greater than the threshold value, the single value is output as a high output value if the single value is greater than the threshold value And an output unit outputting a low value as the output value when the single value is smaller than the threshold value.

In this case, the modeling unit may model the artificial neural network model in a multi-layer structure by applying a third weight to data allocated to the second plurality of nodes and assigning the third weight to a third plurality of nodes.

In this case, the multi-layer structure may be configured such that the first plurality of nodes is an input node layer, the second plurality of nodes is a hidden node layer, and the third plurality of nodes constitute a multi-layer structure with an output node layer .

In this case, the hidden node layer may be configured as a multi-layer, and the multi-layer may be two to six.

Meanwhile, the stock price fluctuation prediction system calculates an error value by comparing the single value and an actual measurement value, and calculates an error value by using the error value to calculate a difference value between the first plurality of nodes, the second plurality of nodes, And an adjustment unit for performing a Back Propagation algorithm to adjust the first weight, the second weight, and the third weight applied to the data.

According to another embodiment of the present invention, there is provided a method for predicting a stock price change using an artificial neural network model, comprising the steps of: inputting a plurality of stock-related data as input data; assigning the inputted plurality of stock- The method comprising: applying a first weight and a first bias to data assigned to the first plurality of nodes, assigning the first weighted data to a second plurality of nodes, Calculating a weighted sum by applying a weighted sum to the activation function and calculating the weighted sum as a single value by comparing the single value with a predefined threshold value, And outputting a low value when the single value is smaller than the threshold value.

In this case, after the step of allocating to the second plurality of nodes, applying a third weight to each of the data allocated to the second plurality of nodes and assigning the third weight to the third plurality of nodes further comprises forming a multi-layer structure .

In this case, the step of forming the multi-layer structure may include: the first plurality of nodes is the input node layer, the second plurality of nodes is the hidden node layer, and the third plurality of nodes is the output node layer, .

In this case, the hidden node layer may be configured as a multi-layer, and the multi-layer may be two to six.

The method of predicting the stock price fluctuation may further include calculating an error value by comparing the single value and an actual measured value after the outputting step and calculating an error value using the error value, And performing a Back Propagation algorithm to adjust a first weight, a second weight, and a third weight applied to the data allocated to the third plurality of nodes.

According to various embodiments of the present invention, by modeling a statistical model as a computer-executable artificial neural network model for predicting stock price fluctuation, it is possible to perform mechanical learning on a computer-executable artificial neural network model, so that even if limited input data is used, And the reliability of the prediction result is improved.

1 is a block diagram for explaining a configuration of a stock price fluctuation prediction system using an artificial neural network model according to an embodiment of the present invention;
FIG. 2 is a diagram for explaining an artificial neural network model having one hidden layer according to an embodiment of the present invention;
3 is a schematic view for explaining a stock price fluctuation prediction system using an artificial neural network model according to an embodiment of the present invention;
4 is a diagram for explaining a multi-layered artificial neural network model according to an embodiment of the present invention.
FIG. 5 illustrates exemplary activation functions of various kinds according to an exemplary embodiment of the present invention, and FIG.
FIG. 6 is a flowchart for explaining a stock price fluctuation prediction method using an artificial neural network model according to another embodiment of the present invention.

Hereinafter, various embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention is not limited or limited by the embodiments.

1 is a block diagram illustrating a configuration of a stock price fluctuation prediction system using an artificial neural network model according to an embodiment of the present invention.

1, a stock price fluctuation prediction system 100 using an artificial neural network model according to an embodiment of the present invention includes an input unit 110, a modeling unit 130, a calculation unit 150, an output unit 170, And an adjustment unit 190.

The stock price fluctuation prediction system 100 according to an embodiment of the present invention may receive a plurality of stock market related data and improve the reliability of a predicted value of an increase or decrease of a closing price of a specific item. The stock fluctuation prediction system 100 may be implemented as a computer capable of communicating with an external server via an Internet network.

First, the input unit 110 receives a plurality of stock-related data as input data. At this time, a plurality of stock-related data may be stored in a database (not shown) installed separately from the stock price fluctuation prediction system 100 or may be stored in a storage unit (not shown) included in the stock price fluctuation prediction system 100 . Or the stock price fluctuation prediction system 100 can receive a plurality of stock market data from a remote server via the Internet.

The input unit 100 receives a plurality of stock-related data stored in a database unit or a storage unit according to a user's input command.

At this time, the multiple stock market data are related to various stock markets including KOSPI 200 index, closing prices of 200 individual items belonging to KOSPI, 30 individual items belonging to KOSDAQ, 11 industry indexes, social media data related to stock market Lt; / RTI > In addition, the plurality of stock market related data may include emotional data on social media data related to the stock market. Sensitivity related data will be separately described below.

The modeling unit 130 allocates a plurality of stock-related data input through the input unit 110 to the first plurality of nodes. Specifically, the modeling unit 130 analyzes the number of input data and generates a first plurality of nodes corresponding to the number of analyzed data. For example, when there are n number of items of stock-related data input, the modeling unit 130 generates n first plurality of nodes as input nodes and n number of market-related data to n first plurality of nodes.

Next, the modeling unit 130 generates an appropriate number of second plural nodes connected with the first plurality of nodes. That is, the modeling unit 130 may generate p second plurality of nodes connected to n first plurality of nodes. Then, the modeling unit 130 applies the first weight to the data allocated to the first plurality of nodes, and assigns the first weight to the second plurality of nodes. In this case, the first weight means a predefined weight applied when connecting the first plurality of nodes and the second plurality of nodes, respectively.

The modeling unit 130 may model a single hierarchical neural network model by generating a first plurality of nodes with an input node layer and a second plurality of nodes with an output node layer. In this way, a single hierarchical structure consisting of only the input node layer and the output node layer assigns a second weight to the data allocated to the output node layer, and calculates a single value. The detailed description thereof will be described in more detail below with reference to the calculation unit 150. [

The artificial neural network structure of the single hierarchical structure described above means a basic unit of the artificial neural network model. In various embodiments of the present invention, an artificial neural network model may be modeled in a multi-layer structure including at least one hidden node layer between an input node layer and an output node layer.

According to an embodiment, the modeling unit 130 generates an artificial neural network model of a basic unit composed of an input node layer (first plural nodes) and an output node layer (second plural nodes) By creating a third plurality of nodes that are assigned data as input, an artificial neural network model of a multi-layer structure can be modeled. That is, the modeling unit 130 generates m third plural nodes connected to p second plural nodes. The modeling unit 130 applies the third weight to data allocated to the second plurality of nodes, and assigns the third weight to the third plurality of nodes. In this case, the third weight means a predefined weight value applied when the second plurality of nodes and the third plurality of nodes are connected to each other. These predefined weight values can be entered by the user and can be changed through machine learning on neural network modeling.

 As described above, the modeling unit 130 may generate a first plurality of nodes as an input node layer, a second plurality of nodes as a hidden node layer, and a third plurality of nodes as an output node layer. That is, the output node layer (second plural nodes) in a single hierarchical structure can be a hidden node layer in a multi-layer structure. It is desirable that the number of such hidden node layers is designed to be two to six.

The calculating unit 150 calculates weighted sum by applying output weights to data allocated to the output node layer, and inputs the weighted sum to the activation function to calculate a single value. Here, the weighted sum refers to a value obtained by applying each output weight to data assigned to each of a plurality of output nodes and then adding them arithmetically. The result obtained by arithmetically combining a plurality of data to which the output weights are applied is a single value.

The output unit 170 compares the single value calculated by the calculation unit 150 with a predetermined threshold value. The output unit 170 outputs a high value as an output value when the single value is larger than the threshold value and outputs a low value as the output value when the single value is smaller than the threshold value.

The adjustment unit 190 compares the single value and the actual measurement value to calculate an error value. The adjustment unit 190 adjusts the first weight and the second weight applied to the data allocated to the first plurality of nodes, the second plurality of nodes using the error value, and adjusts the third weight assigned to the third plurality of nodes do. At this time, the adjustment unit 190 may adjust the first weight, the second weight, and the third weight using a back propagation algorithm.

The back propagation algorithm according to an embodiment of the present invention includes a mechanical learning process for estimating a weight applied to a plurality of nodes and a predefined bias value for each layer in order to improve the reliability of the stock price fluctuation prediction using an artificial neural network model it means. This will be described in more detail below.

The stock price fluctuation prediction system 100 according to an embodiment of the present invention predicts fluctuation of stock prices based on input data. The predicted result is compared with the actual measured value to calculate the predicted rate.

Table 1 below compares the prediction rates for various artificial neural network models.

Table 1 is a table summarizing the fluctuation of the closing price by inputting various stock price related data as input data in the stock price fluctuation prediction system 100 using the ANN model according to an embodiment of the present invention.

The stock price fluctuation prediction system 100 according to an embodiment of the present invention includes the emotion-related data RATIO, the AMT, the 5-day moving average MA 5, the 20-day moving average MA 20, (MA 60), a 100 day moving average (MA 100), a moving average oscillator (OSC), and an on balance volume (OBV) as input data.

Referring to Table 1, the model having a good prediction rate is the second artificial neural network model with a prediction rate of 0.674797. It is possible to perform mechanical learning on an artificial neural network model designed by a back propagation algorithm according to an embodiment of the present invention.

Hereinafter, the mechanical learning of the artificial neural network model based on the back propagation algorithm will be described in more detail.

Back propagation  algorithm( Back Propagation Algorithm )

The objective function of an artificial neural network model is a nonlinear function with respect to connection strength, and a backpropagation algorithm can be applied to optimize it.

The back propagation algorithm can be performed in the following steps. The first step calculates the predicted value using the link strength between nodes. The second step calculates the error between the actual output value and the predicted value. In the third step, the computed error is propagated back to the hidden layer and the input layer to newly adjust the connection strength. The fourth step repeats the previous three steps using the initial connection strength, and stops repeating if the connection strength value remains constant.

At this time, the inverse propagation algorithm requires a learning rate, which is preferably set to be large at first, and to be designed to gradually decrease as the number of iterations increases.

In order to derive the best prediction result, the mechanical learning of the artificial neural network model according to an embodiment of the present invention can estimate the bias (? _J ) of each node and the weight ( _wij ) . For each observation value, the ANN model outputs the predicted value compared with the actual response value. The difference means an error in the output node.

The connection strength estimation process of an artificial neural network model according to an embodiment of the present invention uses an error to repeatedly update the estimated connection strength and the error of the output node is distributed to all nodes of the middle layer connected thereto, Is responsible for a portion of the error. These node-specific errors can be used to update each corresponding connection strength.

In the following, we will look at a concrete method of updating the corresponding connection strength using the error of the output node.

)과 실제 값 사이의 오류(err_k)는 다음의 수학식 1과 같이 정의될 수 있다.In the case where the above-mentioned third plurality of nodes is the output node layer, the output value of the output node K

) And the actual value can be defined by the following Equation (1). &Quot; ( ₁ ) "

이지만, 오류에 대한 정의에 보정요소인

가 곱해진 형태로 오류(err_k)가 정의될 수 있다.Here, the definition of error is

However, in the definition of error,

(Err _k ) can be defined in the form multiplied by.

When such an error occurs, the connection strength (weight and bias) can be updated as shown in Equation (2).

θ _j ^new is the updated bias of the j node, and w _ij ^new is the updated weight between the i and j nodes. Also, θ _j ^old means a bias before updating, and w _ij ^old means a pre-update weight between i and j nodes. And err _j means an error value. And, η is a constant indicating the learning rate, which usually has a value between 0 and 1, and means a real value that adjusts the speed of learning, that is, the amount of change in the connection strength.

Hereinafter, an operation process of updating connection strength when an association error of an output node with respect to a first observation value in an artificial neural network model is as follows will be described in more detail with reference to FIG.

2 is a diagram for explaining an artificial neural network model having one hidden layer according to an embodiment of the present invention.

Referring to FIG. 2, the artificial neural network model is composed of six nodes. The input node layer includes node # 1 and node # 2. The hidden node layer includes nodes 3, 4, and 5. The output node layer includes node 6.

In the artificial neural network model shown in Fig. 2, two pieces of data (first data, second data) are inputted. The input unit 110 receives the first data and the second data. The modeling unit 130 generates a node 1 to receive the first data and a node 2 to receive the second data.

The modeling unit 130 generates three nodes as a hidden node layer. The nodes 3 to 5 are allocated to the input node layer to the hidden node in consideration of the connection strength. When allocating data from an input node to a concealed node, data is allocated in consideration of connection strength. More specifically, it is as follows.

Node 1 can be directly connected to nodes 3, 4 and 5. At this time, the connection between the node # 1 and 3 to 5 times, the node weight is defined as a group W _13, W ₁₄ and W _15. The third node can be predefined as θ ₃ , the fourth node can define the bias as θ ₄ , and the fifth node can define the bias as θ ₅ .

Nodes 3, 4, and 5, which are hidden nodes, can be directly connected to node 6 of the output node layer. At this time, the connection strength between the hidden node layer and the output node layer can be predefined as the connection weights W ₃₆ , W ₄₆ , and W ₅₆ connecting the respective nodes. And the node ₆ can be defined with a bias of? ₆ .

Hereinafter, the back propagation algorithm will be described with specific data.

For example, assume that 0.2 is input for the first data and 0.9 is input for the second data. W ₁₃ = 0.05, W ₁₄ = -0.01, W ₁₅ = 0.05. θ ₃ = -0.3, θ ₄ = 0.2, and θ ₅ = 0.05.

First, the output value of the middle node 3, which is the hidden node layer, can be calculated by Equation (3) below.

That is, the data (Z ₁ ) allocated to the third node considering the connection strength for each input data (the first data and the second data) is 0.430. In the same manner, the assignment data (Z ₂ ) is 0.507 for the node 4 and the data (Z ₃ ) assigned to the node 5 is 0.511. Data can be allocated to the three nodes (nodes 3, 4 and 5) constituting the hidden node layer in the above-described manner.

Data assigned to each of the hidden node layers are each assigned to the output node layer. That is, data of the hidden node layers (nodes 3, 4, and 5) may be allocated to the output node (node 6), respectively. At this time, the three data are 0.430, 0.507 and 0.511, respectively. The weight between the hidden node layer (nodes 3, 4 and 5) and the output node layer (node 6) can be defined as W ₃₆ = 0.01, W ₄₆ = 0.05, and W ₅₆ = 0.015. The bias of node ₆ can be predefined as θ ₆ = -0.015.

The data allocated to the node 6, which is an output node, can be calculated as shown in Equation (4).

That is, the data of the three hidden nodes become input data, and 0.506 is assigned to the data of the output node (node # 6). If we use the boundary value 0.5 to classify the data of the output node, 0.506> 0.5, the result of the neural network model can be classified as the first class.

The error value for the output value of the output node is calculated using Equation (1) as follows.

The error value of the artificial neural network model is calculated as 0.123. By using this error value, the backpropagation algorithm for predefined connection strength is applied to perform the mechanical learning, so that the prediction rate of the artificial neural network model can be increased. The mechanical learning process will be discussed in detail below.

First, when the learning rate (eta = 0.5) is used, the connection strength can be updated as shown in Equation (6).

Mechanical learning is thus interrupted when three general conditions are met while performing mechanical learning to update the connection strength:

First, mechanical learning can be discontinued when the new connection strengths differ by a predetermined amount from the connection strengths obtained in the previous iteration.

Second, mechanical learning can be stopped when the misclassification rate reaches the required target value.

Third, mechanical learning can be stopped when the limit of the number of iterations is reached.

According to an embodiment of the present invention, an artificial neural network model is modeled and a mechanical learning is applied to improve the prediction rate of the output value with respect to input data. Through mechanical learning, we can update the connection weights and node biases between the nodes that make up the neural network model. Also, the mechanical learning can be applied to determine the number of hidden node layers and the number of hidden node layers.

That is, according to one embodiment of the present invention, when the stock price related data is input as input data, the increase or decrease of the closing price of the corresponding stock can be predicted, and the prediction rate can be improved by performing mechanical learning on the conventional prediction result Can be remarkably effective.

According to an embodiment of the present invention, a stock price fluctuation prediction system capable of predicting stock price fluctuation can be implemented using an artificial neural network model. Hereinafter, the structure of the stock price fluctuation prediction system using the artificial neural network model will be described in detail.

3 is a schematic diagram for explaining a stock price fluctuation prediction system using an artificial neural network model according to an embodiment of the present invention.

3, the stock fluctuation prediction system 100 may include an input unit 110, a modeling unit 130, and an output unit 150. The input unit 110 may receive various types of stock price related data as input data. Including the previous day's closing prices, KOSPI 200 index, KOSPI 200, KOSDAQ 30 individual stocks, and 11 industry indexes.

In addition, the input unit 110 may receive emotion-related data as input data in addition to various stock market related data. At this time, the emotion-related data can be supplied from the database unit 120.

A method of calculating emotion-related data stored in the database unit 120 will be briefly described below. The method for calculating emotion-related data based on the social media data and the stock market related web data is limited and understood to limit the scope of the present invention.

First, the stock fluctuation prediction system 100 collects a large amount of documents from social media data and stock market related web data. The collected documents are stored in the database 120 for individual companies, and morphemes are analyzed for expressions or sentences included in a plurality of documents for each individual company.

The stock price fluctuation prediction system 100 analyzes the data of all the plurality of documents by evaluating the sensitivity of all the plurality of documents by performing an emotional evaluation with positive or negative for each keyword extracted from the analyzed morpheme.

The stock price fluctuation prediction system 100 generates analytical data from the correlation between the stock index data and the economic index data together with the sensitivity related evaluation data selected by the predetermined condition among the accumulated emotion evaluation data.

The stock fluctuation prediction system 100 collects a large amount of documents related to at least one individual item from social media data and stock market related web data, and receives stock index data. Here, the individual item is a company listed on the stock market, and the collected document can be implemented in the form of at least one of html, PDF (Portable Document Format), image and moving image.

In this case, the social media data is media data input through a fixed computer or a mobile device connected to a network such as the Internet, and is data that can be mutually shared with other users connected to the network. For example, the social media data may be blog sites that are operated by social media sites operated by a social media server and various portal sites and include personalized contents. Social media sites are so-called social networks, which can be social media served on twitter, facebook and various portal sites.

Web-related data on the stock market is web-based data provided by media companies, national broadcasters, cable broadcasting companies, portal site news, financial corporations, and stock market related organizations, and is more professional or credible market-related data than social media data. These web-related data on the stock market are related to stock market news sites that are served from news agencies, broadcasters, portal site news, financial institutions such as banks, securities firms, insurance companies, and so on. Related statistical sites that provide analysis information related to the stock market.

Stock index data are stock information for each individual stock listed on a stock, and may include stock price information such as market value, high price, low price, close price, closing price, transaction amount, transaction price, transaction source, upper limit price, lower limit price, declaration price, have.

 In the case of collecting a large amount of documents from social media data and stock market related web data, the stock advance prediction system 100 does not collect all the documents but refers to the keyword database (not shown) It is to collect documents.

The keyword database may include a keyword group categorized for each company corresponding to each item. Concretely, in addition to the main keyword related to the company name of each individual item, a keyword related to a product / service related to a product or a service to be released from the company, a keyword related to a person related to the management of the company, Keywords related to the company status, and the like can be stored. The sub keyword may exist in a categorized form classified for each company as a specific word, context or the like of the corresponding company.

For example, the main keyword may be a company name of individual stocks listed on the stock market such as Samsung Electronics, LG Electronics, KT, etc. In the case of Samsung Electronics, keywords related to products / services are "Galaxy", "smartphone", " It can be "tablet" or "appmarket." Human-related keywords can be key executives of Samsung Electronics, executives of companies dealing with Samsung Electronics, and keywords related to company situation. , And may include various words such as "ever-ever", "performance", "goodbye", "apple", "

The stock price fluctuation prediction system 100 extracts the documents including the main keyword, the product / service related keyword, and the human related keyword among the keywords described above in the expressions included in the collected plurality of documents, .

The stock fluctuation prediction system 100 can store extracted documents in a form suitable for morphological analysis. That is, they can be distributedly stored in formats of the extracted documents for each individual item group, for example, html, pdf, image, moving image, and the like.

The stock price fluctuation prediction system 100 is a preprocessing for processing in a form suitable for emotion evaluation, and performs processing for analyzing morphemes which are the smallest language units having a meaning with respect to the format of a plurality of stored documents to specify each part of speech.

In addition, the stock price fluctuation prediction system 100 can classify sentences, contexts, and the like in the expressions included in the document format in units of words, and can parse keywords adjacent to the keywords related to individual items. For example, in a case where a sentence relating to Samsung Electronics and a sentence related to the LG Electronics coexist in a blog site of a particular person, the stock price fluctuation prediction system 100 calculates the price of the blog site And then searches for keywords such as the names of Samsung Electronics or LG Electronics, keywords of goods / services, personal information, etc., parses the words and phrases adjacent thereto, and searches for keywords by Samsung Electronics and LG Electronics And stores them.

The stock fluctuation prediction system 100 evaluates emotion for all of a plurality of documents by performing emotional evaluation with affirmative or negative for each processed keyword, and statistically processes the morphologically analyzed keyword.

The stock price fluctuation prediction system 100 refers to an emotion dictionary database storing an evaluation of affirmative, neutral or negative for each morpheme analyzed keyword and a score associated with the evaluation, and determines whether positive, neutral or negative One is evaluated and the other is scored. In this case, the scoring algorithm may be Naive bayes algorithm, Simple voter algorithm, K Nearest Neighborhood (KNN), SVM (Support Vector Machine). As an example of the simple voter algorithm, the emotion dictionary database can store the keywords of affirmative, neutral, and negative in table form as the emotion evaluation for the keyword. Most of the parts of a keyword related to such emotional evaluation can be composed of nouns and adjectives.

The emotion dictionary database can include a table of affirmative and neutral negative evaluations. For example, in the affirmative evaluation table, there are keywords such as "rise "," maximum ever ", "ascend ", and the score" 1 " In the negative evaluation table 166, keywords such as "recession "," down ", and "complaint" exist and the score "-1" The score given to the keyword stored in the neutral evaluation table is "0 ". This score is exemplified to distinguish between positive and negative, but in contrast, the score associated with positive or negative evaluation may be calculated by differentiating the weight of the keyword according to the degree of sensitivity of the market participants to the keyword, .

The stock fluctuation prediction system 100 can calculate the emotion related evaluation data such as the emotion index for all the plurality of documents by summing the scores given for the keywords determined as positive, neutral, and negative by the emotion dictionaries database.

Here, the stock price fluctuation prediction system 100 performs the sensitivity evaluation on the keywords of all the documents, and does not perform the affirmative, neutral, and negative evaluation on a document-by-document basis. If an emotional evaluation is performed on a document to determine the sensitivity nuances of the document, each document may be given a negative evaluation of an equivalent score, even though there are many more negatively valued keywords compared to other documents . According to this, the ratio of the positive or negative elements of the individual items existing from all the plurality of documents extracted from the social media data and the stock market related web data can be analyzed to be distorted.

 Therefore, the stock price fluctuation prediction system 100 can perform the emotion evaluation without grouping the morphologically analyzed keywords from all the plurality of documents, by document, thereby preventing distortion of the analysis.

The stock price fluctuation prediction system 100 may perform statistical analysis on the morphologically analyzed keywords, such as the number of collected data per period and correlation analysis between the keywords, and display the results. In addition, the stock fluctuation prediction system 100 may select keywords that are not registered in the keyword database among the analyzed keywords, and may update and store the newly selected keywords in the keyword database. The user can separately store the keywords to be reflected in the sensitivity evaluation among the new keywords in the sensitivity dictionary database.

On the other hand, the stock price fluctuation prediction system 100 can generate analytical data from correlation between stock market index data and economic index data, together with the emotion related evaluation data selected by predetermined conditions among the accumulated emotional evaluation data.

The stock price fluctuation prediction system 100 according to an embodiment of the present invention may store emotion-related evaluation data and stock index data in the database 120 according to the above-described method, and may use the data as input data.

When the stock market related data is input as input data, the input data can be assigned to the artificial neural network model. The input data assigned to the artificial neural network model can be calculated as a single value by appropriately connecting strengths by weight and bias between nodes.

The calculated single value is input to the activation function as an input value, and the output unit 150 outputs a high value or a low value as an output for a single value.

The neural network model described above may include a hidden node layer. Such a hidden node layer can be configured in a multi-layer structure. Hereinafter, an artificial neural network model having a hidden node layer that can be configured as a multi-layer will be described in detail.

4 is a diagram for explaining a multi-layered artificial neural network model according to an embodiment of the present invention.

Referring to FIG. 4, the input node layer 131 may include n input nodes. That is, n input nodes are composed of X ₁ to X _n . It is preferable that the input node layer 131 is formed of a single layer. In addition, it is desirable that the number of nodes constituting the input node layer 131 is designed so as to correspond one-by-one to the number of input data.

The hidden node layer 132 may be composed of p hidden nodes. Although the hidden node layer 132 shown in FIG. 4 is configured as a single layer, it is preferable that the hidden node layer 132 includes at least one or more multi-layers. The hidden node layer 132 may include a plurality of hidden nodes in one layer. The plurality of hidden nodes may be composed of Z ₁ to Z _p . In this case, the number of hidden nodes is plural, but it is not necessarily configured to correspond to the number of nodes of the input node layer in a one-to-one correspondence.

The hidden node layer 132 may be composed of a plurality of layers, and the individual layers may be composed of a plurality of hidden nodes. As the number of hidden nodes and hidden nodes in the hidden node layer 132 increases, the mechanical learning power is improved. However, since the structure is complicated, the predictive power is not improved. Therefore, it is necessary to design the number of hidden nodes and hidden nodes Do. Preferably, the number of hidden layers in the hidden node layer 132 may be designed to be three to five.

The output node layer 133 may be designed as one output node or may be designed as a plurality of output nodes. However, it is preferable that the output node layer 133 is designed as a single layer like the input node layer 131.

As shown in FIG. 4, connection strength can be defined between the nodes between the input node layer 131 and the hidden node layer 132. Also, connection strengths may be defined between the nodes between the hidden node layer 132 and the output node layer 133. [

This connection strength can be updated through mechanical learning of the neural network model.

According to an embodiment of the present invention, the prediction accuracy of the stock price fluctuation prediction system using the artificial neural network model can be improved by updating the connection strength of the nodes constituting the artificial neural network model through mechanical learning.

As shown in FIG. 4, the relationship between data modeling results by an artificial neural network model for a plurality of input data in a multi-layer neural network structure having one or more hidden layers will be described in detail below.

The following Equation (7) describes the relationship between the data which the input data (X ₁ to X _n) allocated to the hidden layer node.

Here, X _i denotes data allocated to the input node, V _ij denotes a weight between the input node and the hidden node, V _0j denotes the bias of the hidden node, and Z _j denotes data allocated to the hidden node .

Referring to Equation (7), the data allocated to the hidden node may be determined by weighting the weighted data of the input node, performing an arithmetic operation, and adding the bias defined at the hidden node. The data allocated to the determined hidden node is again applied to the output node by applying the weight.

Equation (8) below describes the relationship between data that allocates data assigned to the hidden node layer to the output node layer.

Wherein, Z _i is a mean data associated with a hidden node, and, W _ij denotes a weight between the hidden nodes and the output node, and, W _0k refers to the bias of the output node and, y _k are assigned to the output node data .

If y _k is expressed by integrating Equation (7) and Equation (8), it can be expressed as Equation (9).

The value calculated by the above-described Equation (9) can be used as data for inputting to another node through the activation function () or as a result value for input data.

This activation function can be basically defined as a unit step function as shown in Equation (10) below.

Where u is the data value assigned to the output node, [theta] is the predefined threshold, and [phi] is the activation function.

In addition, the activation function can be defined by various kinds of functions as shown in FIG. 5. It can be calculated as a single value. At this time, various kinds of activation functions can be utilized as shown in FIG.

 FIG. 5 is a diagram illustrating exemplary activation functions of various kinds according to an exemplary embodiment of the present invention. Referring to FIG.

5, the activation function may be implemented as a step function of outputting a low value when the single value U is smaller than the threshold value and outputting a high value when the single value U is larger than the threshold value (refer to FIG. 5 (a) ). Alternatively, the activation function may be implemented as a signum function that outputs -1 as a low value when the single value U is smaller than the threshold value and outputs 1 as a high value when the single value U is larger than the threshold value (see FIG. . Alternatively, the activation function may be implemented as a linear function that linearly increases or decreases according to the single value U (see Fig. 5 (c)). Alternatively, the activation function may be a unipolar sigmoid function (see Fig. 5 (d)) that converges to 1 or 0 according to the single value U, or a bipolar sigmoid function converging to 1 or -1 according to the single value U (See (e) of FIG. 5).

FIG. 6 is a flowchart for explaining a stock price fluctuation prediction method using an artificial neural network model according to another embodiment of the present invention.

Referring to FIG. 6, the stock fluctuation prediction system 100 receives input of a plurality of stock-related data from a database (S610). The database can be configured separately from the stock price fluctuation prediction system 100, and the database stores a plurality of stock-related data.

The stock price fluctuation prediction system 100 allocates a plurality of stock market related data inputted to the first plurality of nodes (S620). That is, the stock price fluctuation prediction system 100 generates the first plurality of nodes by the number of the plurality of stock-related data, and allocates the data inputted to each of the first plurality of nodes.

The stock change estimating system 100 applies the first weight to the data allocated to the first plurality of nodes (S630). That is, the stock price fluctuation prediction system 100 generates a second plurality of nodes individually connected to the first plurality of nodes, and applies a weight to each of the nodes connecting the first plurality of nodes and the second plurality of nodes to the data .

The stock change estimating system 100 allocates the weighted data to the second plurality of nodes (S640). In other words, the stock price fluctuation prediction system 100 weights the weighted data and allocates them by applying a bias defined individually to the second plurality of nodes. After the step S640 of allocating to the second plurality of nodes, the stock advance estimation system 100 applies a third weight for each data allocated to the second plurality of nodes and assigns the third weight to the third plurality of nodes to form a multi-layer structure can do.

The stock price fluctuation prediction system 100 applies the second weight to the data allocated to the second plurality of nodes, and calculates the weighted sum by summing the data to which the second weight is applied (S650). In other words, the stock move prediction system 100 applies the second weight to data allocated to the second plurality of nodes, and obtains an arithmetic sum of the weighted data. The stock price fluctuation prediction system 100 applies the bias of the output node to the arithmetic sum result.

The stock fluctuation prediction system 100 compares the calculated single value with a threshold (S670). If the calculated single value is smaller than the threshold value (S670-N), the upstream / downstream prediction system 100 calculates a low value as a result value (S680). If the single value is larger than the threshold value (S670-Y) , The stock price fluctuation prediction system 100 calculates a high value as a result value (S690).

The stock fluctuation prediction system 100 may further execute a step of calculating an error value by comparing an actual measurement value with a resultant value. Then, the stock price fluctuation prediction system 100 uses the error value to adjust the first weight, the second weight, and the third weight applied to the data allocated to the first plurality of nodes, the second plurality of nodes, and the third plurality of nodes Back Propagation algorithm can be performed.

The above-described methods according to various embodiments of the present invention may be stored as a code in a computer-readable storage medium. The code for performing the above-described methods according to various embodiments of the present invention may be stored in a memory such as a random access memory (RAM), a flash memory, a ROM (Read Only Memory), an EPROM (Erasable Programmable ROM), an Electrically Erasable and Programmable ROM ), A register, a hard disk, a removable disk, a memory card, a USB memory, a CD-ROM, and the like.

Although illustrative embodiments and applications of the present invention have been shown and described, many changes and modifications may be made without departing from the scope of the present invention, and such modifications may be made by one of ordinary skill in the art to which the present invention pertains Can be clearly understood. Accordingly, the described embodiments are illustrative and not restrictive, and the invention is not limited by the accompanying detailed description, but is capable of modifications within the scope of the claims.

110: input unit 120: database
130: a modeling unit 150:
170: output unit 190:

Claims (10)

In a stock price fluctuation prediction system using an artificial neural network model,
An input unit for receiving a plurality of stock-related data as input data;
A modeling unit for modeling an artificial neural network model by assigning the input plurality of stock-related data to a first plurality of nodes, applying a first weight to data assigned to the first plurality of nodes, and assigning the first weight to a second plurality of nodes;
Calculating a weighted sum by applying a second weight to data allocated to the second plurality of nodes, inputting the weighted sum into an activation function, and calculating the weighted sum as a single value; And
An output unit for comparing the single value with a predetermined threshold value and outputting a high value as an output value when the single value is greater than the threshold value and outputting a low value as the output value when the single value is less than the threshold value; A Stock Price Prediction System Using an Artificial Neural Network Model. The method according to claim 1,
The modeling unit,
And applying the third weight to each of the plurality of data allocated to the second plurality of nodes to assign the third weight to the third plurality of nodes to model the ANN model in a multi-layer structure. 3. The method of claim 2,
Characterized in that the first plurality of nodes is an input node layer, the second plurality of nodes is a hidden node layer, and the third plurality of nodes constitute a multi-layer structure as an output node layer Model of stock price fluctuation prediction system. The method of claim 3,
Wherein the hidden node layer is composed of a plurality of layers, and the multi-layer has two to six layers. 3. The method of claim 2,
Wherein the stock price fluctuation prediction system comprises:
A first weight, a second weight, and a third weight applied to the data allocated to the first plurality of nodes and the second plurality of nodes using the error value, wherein the error value is calculated by comparing the single value and the actual measured value, And an adjustment unit for performing a Back Propagation algorithm to adjust the weights. In a stock price fluctuation prediction method using an artificial neural network model,
Inputting a plurality of stock market related data as input data;
Assigning the input plurality of stock-related data to a first plurality of nodes;
Applying a first weight to data assigned to the first plurality of nodes;
Assigning the first weight and the first bias applied data to a second plurality of nodes;
Applying a second weight to data allocated to the second plurality of nodes to compute a weighted sum;
Calculating the weighted sum as a single value by inputting the calculated weighted sum into an activation function; And
Comparing the single value with a predefined threshold and outputting a high value if the single value is greater than the threshold value and outputting a low value if the single value is less than the threshold value. A method of predicting stock price fluctuation. The method according to claim 6,
After the step of assigning to the second plurality of nodes,
And applying a third weight to data allocated to the second plurality of nodes and allocating the third weight to a third plurality of nodes to form a multi-layer structure. 8. The method of claim 7,
The step of forming the multi-layer structure may include: the first plurality of nodes is a plurality of input nodes, the second plurality of nodes is a plurality of concealed nodes, and the third plurality of nodes is a multi- A method for predicting stock price fluctuation using an artificial neural network model. 9. The method of claim 8,
Wherein the plurality of hidden nodes are composed of a plurality of hidden nodes of a multi-layer, and the multi-layers are two to six. 8. The method of claim 7,
The method includes:
After the outputting step,
Calculating an error value by comparing the single value with an actual measured value; And
A second weight, and a third weight applied to the data allocated to the first plurality of nodes, the second plurality of nodes, and the third plurality of nodes using the error value, the backward propagation is adjusted to adjust the first weight, the second weight, Wherein the step of estimating the stock price is performed using an artificial neural network model.

Images