KR20210063775A - Device for predicting price of asset use of convolution neural network - Google Patents

Abstract

An asset price prediction device using a CNN in accordance with the present invention comprises: a tree module which receives data fitted with a difference by the time series length of a reference date (T) and a predicted operating period (K) between input data for an asset to be predicted and past price data of the asset; and a CNN module which receives an image map generated based on variables of the image map and a comparison value between the predicted price data predicted through the tree module and actual price data corresponding to the predicted operating period (K). The device learns about the predicted price data through the CNN module, predicts an image map prior to a specific time to be predicted through the CNN module, and predicts the final predicted price data for the specific time point to be predicted based on the predicted image map. Therefore, the present invention can overcome the limitations of an artificial intelligence machine learning model.

Description

Asset price prediction device using CNN (Convolution Neural Network) {DEVICE FOR PREDICTING PRICE OF ASSET USE OF CONVOLUTION NEURAL NETWORK}

The present invention relates to an apparatus for predicting the price of an asset using a Convolution Neural Network (CNN).

The bond dealer system refers to a system in which securities companies designated as dealers directly trade with investors with their own funds. Securities companies analyze and forecast future bond yields, and present bid and ask price yields similar to the timing of stock trading. Investors may buy or sell bonds if they determine that the rate of return offered by the securities company is reasonable.

In order to analyze and forecast future bond yields, securities companies make predictions using quantitative indicators to predict prices, such as economic indicators and technical indicators. Ho discloses a financial asset price prediction information providing system using the Internet.

In order to predict the price of an asset, an artificial intelligence algorithm that can learn autocorrelation is mainly used. Algorithms for learning autocorrelation typically include RNN and LSTM, and it is composed of a structure in which previous learning parameters affect next learning.

However, assets are determined by a myriad of quantitative indicators such as employment index, price index, industrial production index, interest rate, and relative strength index. When these innumerable quantitative indicators are input to RNN, LSTM, etc., as the dimension of the input data increases, a curse of dimension occurs, which is not suitable for predicting assets.

A tree module that receives data fitted with a difference as much as the time series length of the reference date (T) and the forecast operating period (K) between the input data for the asset to be predicted and the past price data of the asset, and through the tree module An object of the present invention is to provide an apparatus including a CNN module that receives an image map generated based on a comparison value between the predicted predicted price data and the actual price data corresponding to the predicted operating period (K) and the variables of the image map.

In addition, learning about the predicted price data through the CNN module, predicting an image map prior to a specific time to be predicted through the CNN module, and predicting the specific to be predicted based on the predicted image map It is intended to provide a device for predicting the final predicted price data for a point in time.

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

As a means for achieving the above-described technical problem, an embodiment of the present invention provides as much as the time series length of the reference date (T) and the forecast operation period (K) between the input data for the asset to be predicted and the past price data of the asset. Created based on a comparison value between a tree module that receives data fitted with a difference of , and a comparison value between the predicted price data predicted through the tree module and the actual price data corresponding to the predicted operating period (K) and the variables of the image map It includes a CNN module that receives the image map that has been made, learns about the predicted price data through the CNN module, predicts the image map corresponding to the previous of a specific time to be predicted through the CNN module, and the predicted It is possible to provide an asset price prediction device that predicts the final predicted price data for the specific time point to be predicted based on the image map.

According to another embodiment of the present invention, the tree module may predict the predicted price data corresponding to the period between the reference date and the time between the reference date and the predicted operating period.

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

According to any one of the above-described problem solving means of the present invention, a CNN that can overcome the limitations of the artificial intelligence machine learning model and predict the future asset price more accurately than the existing general price prediction model (eg, LSTM) It is possible to provide an asset price prediction device using

It is possible to provide an asset price prediction device that classifies quantitative indicators of many dimensions using the TREE module and provides the contribution of input data by dimension that affects the price prediction of assets.

By processing quantitative input data into image input data, it solves the curse of dimensionality, different probability distributions between data for predicting the future and data used for AI learning, and overfitting problems that occur during AI learning. It is possible to provide an asset price prediction device that predicts the price of an asset with high accuracy.

As assets can be predicted in the same way for asset classes with different characteristics, such as bonds and stocks, the scope of application is wide, and it can be applied to various portfolio strategies by applying price forecasting, forecasting trends over the forecast period, and price decline and rise accuracy. It is possible to provide an asset price prediction device that allows

1 is a block diagram of a price prediction apparatus according to an embodiment of the present invention.
2A and 2B are exemplary views illustrating simulation conditions for price prediction of assets and historical price data and input data collected according to simulation conditions according to an embodiment of the present invention.
3A to 3C are exemplary diagrams for explaining a process of predicting predicted price data using a first prediction module according to an embodiment of the present invention.
4 is an exemplary view for explaining a process of deriving a comparison value between predicted price data and actual price data according to an embodiment of the present invention.
5 is an exemplary diagram for explaining a process of predicting an image map corresponding to a previous point in time using a generated image map according to an embodiment of the present invention.
6 is an exemplary diagram illustrating predicted final predicted price data according to an embodiment of the present invention.
7 is a flowchart of a method for predicting the price of an asset in the price prediction apparatus according to an embodiment of the present invention.

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

Throughout the specification, when a part is said to be "connected" with another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included, and one or more other features, not excluding other components, unless specifically stated to the contrary. It is to be understood that it does not preclude the presence or addition of any number, step, action, component, part, or combination thereof.

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

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

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a price prediction apparatus according to an embodiment of the present invention. Referring to FIG. 1 , the price prediction apparatus 100 may include a collection unit 110 , a first prediction unit 120 , a derivation unit 130 , a generation unit 140 , and a second prediction unit 150 . have.

The collection unit 110 may collect past price data of an asset to be predicted and input data associated with the asset. For example, the collection unit 110 may collect, as input data, time series data for economic indicators, technical indicators, and global common economic indicators of the market to which the asset to be predicted belongs together with historical price data of the asset to be predicted. . The historical price data of the asset to be predicted and the input data associated with the asset are composed of a dimension (data type), and the number of input data may indicate the length of time series data.

Here, the input data is limited to the Index Price, which recalculates the prices of various assets to be predicted according to a specific methodology, and the types vary greatly depending on the methodology and the types and characteristics of the assets, but the representative index KOSPI (domestic ), S&P 500 (USA), and EMBI (each country). Representative indices are a reference index for many asset management companies and money management institutions, and they are very useful in forecasting because their influence is significant. For example, ETFs, which are index-following products, can be applied to strategies for directly predicting prices and investment strategies for index-related futures options.

For example, when the asset is a bond, the collection unit 110 may collect as input data indicators for understanding the economic situation of the bond issuer and the US economic indicators affecting the global economy. Input data may include, for example, retail sales, industrial production, MoneySupplyM3, CSIICMEA Index, Moody's Seasoned Baa Corporate Bond Yield, Wholesale Trade Sales, Unemployment Rate, Retail Sales Advance, etc. It can contain any data available.

For another example, if the asset is stock, the collection unit 110 may determine the economic situation of the country issuing the stock, and the indicators for US economic indicators and stocks (SMB, HML, RMW, CMA, RF) , investment indicators (PE Ratio, dividend_yield, market_capital ratio), etc. can be collected as input data.

Input data collected in relation to these assets will be described in detail with reference to FIGS. 2A and 2B .

2A and 2B are exemplary views illustrating historical price data and input data collected according to simulation conditions and conditions for price prediction of an asset according to an embodiment of the present invention.

2A is an exemplary diagram illustrating simulation conditions for price prediction according to an embodiment of the present invention. Referring to Figure 2a, the simulation conditions are target country 200, forecast date 201, number of input data 202, reference date 203, forecast operating period 204, MAP SPEC 205, variable 206 , a minimum value 207 , an interval 208 , a number of instances 209 , a maximum value 210 , and the like.

For example, the simulation conditions are: 'South Africa', '2017/06/01~30 (business days: 22 days)', number of input data ( 202): '239', base date (203, expressed as 'T' in the detailed description below): '2017/04/27', forecast operating period (204, expressed as 'K' in the detailed description below): It can be set to '24 days'. Here, the reference date 203 may mean a specific point in time in the past, and the forecast operating period 204 may mean a period to be predicted with respect to the predicted price data of an asset based on the reference date 203 . For example, if the base date 203 is 2017/04/27 and the forecast operating period 204 is set to '24 days', the price of the asset corresponding to the date 24 days have elapsed from 2017/04/27 is can be predicted.

In addition, the simulation condition may be set by the manager for the image map necessary to predict the price of the bond. For example, the simulation conditions are combinations that allow changes to the image map: MAP SPEC (205): '8*40', variables in the image map (206): 'M and N', minimum value (207): 'M-40 days, N-50 days', interval (208): 'M-25 days, N-13 days', number of cases (209): 'M-8, N-40', maximum value (210) ): 'M-215 days, N-557 days' can be set.

Here, 'M' is a time series length to be fitted to a TREE module corresponding to a first prediction module to be used later, and may be a first period value indicating how much past information is to be fitted to the tree. In addition, 'N' is the period of period change rate to express past price data when fitting to the TREE module based on a specific date, and the rate of change from the past date to the time point corresponding to a specific date as a period change rate for some past period may be a second period value indicating whether to express

2B is an exemplary diagram illustrating input data collected according to an embodiment of the present invention. Referring to FIG. 2B , the collection unit 110 collects input data 220 related to assets (eg, bonds) to be predicted from 17/01/27 to 17/03/24, and 17/02 It is possible to collect historical price data 221 of assets (eg bonds) to be predicted from /15 to 17/04/27.

Returning to FIG. 1 again, the first prediction unit 120 may predict the predicted price data through the first prediction module. A process of predicting the predicted price data through the first prediction module will be described in detail with reference to FIGS. 3A to 3C .

3A to 3C are exemplary diagrams for explaining a process of predicting predicted price data using a first prediction module according to an embodiment of the present invention.

3A is an exemplary view for explaining a process in which a first period value and a second period value are set based on a reference date and a predicted operating period according to an embodiment of the present invention. Referring to FIG. 3A , the first prediction unit 120 may receive a reference date (T, 300) and a prediction operation period (K=24) set. Here, the reference date (T, 300) is set to a specific time in the past, and the first period value 302 is based on the time (T-24) obtained by subtracting the predicted operating period (K=24) from the reference date (T, 300). , and the second period value 303 may be set based on the reference date (T, 300).

Referring to FIG. 3A , the first period value 302 is the time point (T-24, 2017/03/24) minus the predicted operating period (K=24) from the reference date (T, 2017/04/27, 300) and From the base date (T, 2017/04/27, 300) when the forecast operating period (K=24) is omitted (T-24, 2017/03/24), the preset minimum value (40 days) is reflected (T-64) , 2017/01/27), and the second period value 303 is a time point (T-50) corresponding to the reference date (T, 2017/04/27, 300) and the preset minimum value (50 days) , 2017/02/16).

3B is an exemplary diagram illustrating a first prediction module according to an embodiment of the present invention. Referring to FIG. 3B , the first prediction module may be a TREE module. The TREE module is a model that can quickly and accurately classify many input data, and has strengths in classifying and fitting past data. Here, when predicting the future using the TREE module, a separate variable adjustment is required.

The reason for using the TREE module as the first prediction module is that the TREE module can quickly classify the dimensions (features of data) of input data of many dimensions, and in classifying it with an arbitrary threshold value of each node of the TREE module. This is because the importance of each dimension can be measured, and through this, explanatory power for the dimensions of the input data can be secured, and accuracy can be increased when appropriate input data and past price data are used.

3A and 3B , the first prediction unit 120 corresponds to the input data corresponding to the first period value 302 among the collected input data and the second period value 303 among the collected past price data. past price data may be input to the first prediction module, and the predicted price data may be predicted through the first prediction module. Also, the first prediction unit 120 may input a period change rate value of the past price data corresponding to the second period value 303 to the first prediction module.

Here, the first prediction unit 120 may fit by placing a difference as much as the time series length of the prediction operating period (K=24) between the input data and the past price data. For example, the first prediction unit 120 may calculate a time point (T-64) at which the value of the first period (302) is reflected from a time point (T-24) obtained by subtracting the prediction operation period (K=24) from the reference date (T, 300). ) and past price data from the time point (T-50) to the reference date (T, 300) at which the second period value 303 is reflected and input into the TREE module, T+1 to T+24 ( 301) can be predicted.

The first prediction unit 120 is a period between the date (T+1) and the reference date (T, 300) after the reference date (T, 300) plus the predicted operating period (K=24) (T+24, 301). The corresponding predicted price data can be predicted. For example, if the forecast operating period (K) is set to 24 days for the reference date (T, 300), up to the point in time (T+24, 301) when the forecast operation period (K=24) is added to the reference date (T, 300) In order to predict the predicted price data corresponding to the index price of , based on the input data up to the reference date (T, 300), the time point (T+24, 301), it is possible to have explanatory power by aiming at this to predict the predicted price data. Here, threshold values of each node for classifying input data with the first prediction module (TREE module) may be set by an administrator.

In addition, the reference date (T, 300) and the predicted operating period (K=24) may be arbitrarily set by the administrator. The reason that the forecast operating period is mainly set to 24 days is that the maximum number of business days excluding weekends is 24 days in one month, and the strategy is often operated by changing the operating period in units of one month.

As another example, the first prediction unit 120 may predict the predicted price change rate value by inputting the price change rate value instead of the past price data into the first prediction module.

Referring to FIG. 3C , the first prediction unit 120 predicts the operating period (K=24) from the 'base date (T, 2017/04/27, 322)' to the 'base date (T, 2017/04/27, 322)' ) plus the time point (T+24, 2017/05/31, 331)', the predicted price data 311 corresponding to the period can be predicted. For example, the first prediction unit 120 inputs the input data 310 corresponding to the time points 'T-24(321) to T(322)' into the TREE module, which is the first prediction module, and 'T+1 It is possible to predict the predicted price data 311 corresponding to the period '2017/04/28 ~ 2017/05/31' (331) corresponding to ~T+24'.

The derivation unit 130 may derive a comparison value between the predicted price data predicted through the first prediction unit 120 and the actual price data. For example, the derivation unit 130 may derive a difference value between the predicted price data and the actual price data corresponding to the predicted operating period, and derive a reciprocal of the derived difference value as a comparison value. Here, the comparison value may be derived as an average value for the predicted operating period. Also, the derivation unit 130 may derive a comparison value between the period change rate value of the predicted price data predicted through the first predictor 120 and the period change rate value of the actual price data. A process of deriving a comparison value will be described in detail with reference to FIG. 4 .

4 is an exemplary view for explaining a process of deriving a comparison value between a period change rate value of predicted price data and a period change rate value of actual price data according to an embodiment of the present invention. Referring to FIG. 4 , the derivation unit 130 derives a difference value between the period change rate value of the predicted price data 410 and the period change rate value of the actual price data 400 corresponding to the predicted operating period, and the derived difference value A reciprocal of ? may be derived as the comparison value 420 . Here, the comparison value 420 may be derived as an average value for the predicted operating period, and may be configured as in Equation 1 below.

[Equation 1]

Referring to Equation 1, K denotes a forecast operating period, T denotes a reference date representing a specific time in the past, and denotes a difference between predicted price data and actual price data for the corresponding period.

Accordingly, the comparison value ZT, 420 at the reference date may be an average value of comparison values in the period T+1 to T+K (prediction operation period). That is, the comparison value (ZT, 420) in the reference date may be derived by comparing the predicted price data 410 and the actual price data 400 in the period T+1 to T+K. Therefore, when the actual price data 400 exists only up to T+K (for example, when T+24 is the current time), it is derived as the average value of the comparison values in the period T+1 to T+K (prediction operation period). The comparison values Z and 420 can be derived up to T (base date).

Here, the comparison value (Z, 420) is the combination of 'M (first period value)*N (second period value)', and the predicted price data 410 and actual price data applied to the first prediction module (TREE module). (400) is the reciprocal of Cost (Z) representing the difference between the values, and the upper comparison value (Z) is the value of the first period value and the second period value [M, N] in which the performance of the first prediction module (TREE module) was good. If the coordinates are indicated and the combination of [M, N] is applied to the first prediction module (TREE module) at a specific time point, it has the advantage of obtaining high-accuracy prediction price data.

The generator 140 may generate an image map based on the first period value, the second period value, and the comparison value. If, in order for the generator 140 to generate the image map corresponding to T+1, it needs to know past price data corresponding to T+25, but when the forecasting period is set to 24 days, the past price up to T+24 Since we know the data, we cannot create an image map corresponding to T+1. Accordingly, the image map may be generated by the reference date T.

The generating unit 140 generates at least one primary image map based on the first period value, the second period value, and the comparison value, and processes the generated primary image maps to be stacked in chronological order to obtain a second second image map. You can create a car image map. For example, the image map may include multidimensional spatial data. Specifically, the first image map may be a two-dimensional image, and the second image map may be a three-dimensional image.

For example, the generator 140 may generate the first image map based on the first period value M, the second period value N, and the comparison value Z. In this case, the generating unit 140 repeatedly performs this to generate a plurality of primary image maps including information on the obtained comparison value Z, and sequentially arranges the plurality of generated primary image maps in chronological order. It is possible to create a second image map by processing it so that it is stacked.

The second prediction unit 150 inputs the generated image map into the second prediction module to learn about the prediction image price data, and predicts the image map corresponding to the previous specific time point to be predicted through the second prediction module. have. A process of generating a map and predicting an image map corresponding to a previous point in time will be described in detail with reference to FIG. 5 .

5 is an exemplary diagram for explaining a process of predicting an image map corresponding to a previous point in time using a generated image map according to an embodiment of the present invention.

The second prediction unit 150 may input the generated image map 503 into the second prediction module to learn about the predicted price data. Here, as the second prediction module, a Convolution Neural Network (CNN) 501, which is mainly used for image learning with high learning accuracy, may be used.

The reason for using the CNN 501 as the second prediction module is that the CNN 501 is a network showing excellent accuracy in image feature extraction and classification problems, and using this network principle, the first period value (M) and the second A high comparison value (Z, feature) among the comparison values (Z), which is a number obtained through the combination of period values (N), can be extracted. Through this, by applying the combination of the first period value (M) and the second period value (N) corresponding to the extracted high comparison value (Z) to the first prediction module using the TREE module, the prediction accuracy of the TREE module is improved. can make it higher.

When using a deep learning model other than CNN (501), to predict assets with numerous dimensions (about 300 dimensions of the input data, which are macroeconomic indicators), an exponential data amount of 300 dimensions is required However, financial time series data has a limitation that it is short (250 days per year -> 250 lengths). In addition, when learning by inputting numerous input data dimensions to each neural network node, it is not easy to find a non-linear relationship.

However, in the case of using the CNN 501, it is advantageous in learning because only as many parameters as the kernel size and the number of filters of the convolution are learned, and the TREE module is the main component of the first prediction module. Because it plays an indirect role to help find the variable, it can be less affected if the prediction fails.

The second prediction unit 150 may predict the image map 502 corresponding to the time point when the prediction operation period has elapsed from the reference date through the second prediction module. For example, when the reference date is 'T' and the prediction operation period is set to '24 days', the second prediction unit 150 uses the secondary image map 503 generated before the reference date (T). To predict after the prediction operation period (K = 24), image learning is performed using the second prediction module (CNN, 501), and T + K time point ( T+24) of the image map 502 can be predicted.

The second prediction unit 150 may learn the predicted price data by using the image map 503 immediately before the price prediction start date in order to predict the price of the asset. The image map 503 may include a coordinate pair of the first period value M and the second period value N, and a comparison value (Z value) corresponding to the coordinate pair. The image map 503 makes it possible to determine which coordinates of the first period value (M) and the second period value (N) have a good comparison value (Z), and based on the price prediction start date (T+24) As such, it can be created only up to the date (T+24-predicted operating period) minus the time series length of the forecast operating period.

At this time, since the image map 502 of the actual (T+24) time point cannot be grasped, the (T+24-) factor is considered through the second prediction module (CNN, 501) in terms of the temporal/spatial (SPATIO-TEMPORAL) After the image map 503 is learned up to the date of the prediction operation period), the image map 502 at the time T+24 can be predicted to be learned.

For example, if today is T+24 (2017.05.31), to predict the price from T+25 to 5+48 (2017.0601 to 2017.06.30), which corresponds to the future period based on 24 days of the operating period, The image map 502 at the time T+24 is required, and the image map 502 at the time point is obtained through the second prediction module (CNN, 501) trained with the image maps 503 generated up to the reference date (T). predictable.

Thereafter, the first prediction unit 120 may predict the final predicted price data for a specific time point to be predicted based on the predicted image map 502 . For example, the first prediction unit 120 extracts at least one comparison value corresponding to a predetermined threshold value or more from the predicted image map 502 , and one first period value corresponding to the extracted comparison value. and a second period value.

In other words, the present invention can extract optimal (M, N, Z) parameters for deriving predicted price data through the first prediction unit and the second prediction unit.

The first prediction unit 120 may input the derived first period value (M) and the second period value (N) into the first prediction module to predict the final predicted price data for a specific time point to be predicted. Alternatively, the first prediction unit 120 inputs the period change rate value of the past price data derived based on the second period value by fitting it into the first prediction module, and the forecast period change rate value for a specific time point to be predicted or the final The rate of change over the forecast period can be predicted The final predicted price data for the predicted specific time will be described in detail with reference to FIG. 6 .

6 is an exemplary diagram illustrating predicted final predicted price data according to an embodiment of the present invention. Referring to FIG. 6 , the first prediction unit 120 extracts a comparison value (Z) corresponding to the top 20% by using the image map predicted through the second prediction module, and adds the extracted comparison value (Z). The corresponding (MN) combination may be fitted back to the first prediction module (TREE module) as quantitative input data, and the ensembled value may be used by hardvoting the finally predicted prices.

For example, after learning the image map generated before the reference date (T) in the second prediction unit 150 on the predicted price data, the image map corresponding to the reference date (T) is input to the reference date (T) It is possible to predict an image map corresponding to the time point (T+24) after the prediction operation period (K=24) has elapsed.

Here, the first prediction unit 120 extracts the comparison value Z corresponding to the top q% (q is designated by the administrator, for example, 5%) of the predicted image map, and the extracted comparison value ( A combination of the first period value M and the second period value N corresponding to Z) may be derived. At this time, according to the extracted combination of the first period value (M) and the second period value (N), the first prediction module (TREE module) is fitted and inputted, and finally the first period value (M) and the second period value (N) are inputted. The final predicted price data for a specific time point to be predicted can be predicted using the average value of the predicted values as many as the number of combinations of period values (N). For example, the average value of the predicted values can be configured as MAP (M = 8, N = 40, the total number of combinations of M and N = 320 (8 * 40), the top 5% = 16 (320 * 0.05)) have.

Referring to FIG. 6 , the first prediction unit 120 predicts the final predicted price data by fitting and inputting a total of 16 combinations of the first period value (M) and the second period value (N) to the first prediction module. can For example, the first prediction unit 120 fits and inputs the first prediction module (TREE module), and predicts the result using the average value (hardvoting) of predictions (= ensemble derived through 16 TREE modules). It is possible to predict the final predicted price data through

The price prediction apparatus 100 may be executed by a computer program stored in a medium including a sequence of instructions for predicting a price. When the computer program is executed by the computing device, it collects historical price data of the asset to be predicted and input data associated with the asset, and the input data corresponding to the first period value among the collected input data and the second of the collected historical price data. Input the past price data corresponding to the 2 period values into the first prediction module, predict the predicted price data through the first prediction module, derive a comparison value between the predicted predicted price data and the actual price data, and the first period An image map is generated based on the value, the second period value and the comparison value, the generated image map is input to the second prediction module to learn about the predicted price data, and a specific point in time to be predicted through the second prediction module It is possible to provide a computer program stored in a medium including a sequence of instructions for predicting an image map corresponding to the previous of , and predicting the final predicted price data for a specific point in time to be predicted based on the predicted image map.

7 is a flowchart of a method for predicting the price of an asset in the price prediction apparatus according to an embodiment of the present invention. The method of predicting the price of an asset in the price prediction apparatus 100 illustrated in FIG. 7 includes steps of time-series processing according to the embodiments illustrated in FIGS. 1 to 6 . Therefore, even if omitted below, it is also applied to the method of predicting the price of an asset in the price prediction apparatus 100 according to the embodiment shown in FIGS. 1 to 6 .

In step S710 , the price prediction apparatus 100 may collect past price data of an asset to be predicted and input data associated with the asset.

In step S720, the price prediction apparatus 100 may input the input data corresponding to the first period value among the collected input data and the past price data corresponding to the second period value among the collected past price data to the first prediction module. have.

In step S730, the price prediction apparatus 100 may predict the predicted price data through the first prediction module.

In step S740, the price prediction apparatus 100 may derive a comparison value between the predicted predicted price data and the actual price data.

In operation S750, the price prediction apparatus 100 may generate an image map based on the first period value, the second period value, and the comparison value.

In step S760, the price prediction apparatus 100 may input the generated image map into the second prediction module to learn about the predicted price data.

In step S770 , the price prediction apparatus 100 may predict an image map corresponding to a previous specific time point to be predicted through the second prediction module.

In step S780, the price prediction apparatus 100 may predict the final predicted price data for a specific time point to be predicted based on the predicted image map.

In the above description, steps S710 to S780 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, and the order between steps may be switched.

The method of predicting the price of an asset in the price prediction apparatus described through FIGS. 1 to 7 may be implemented in the form of a computer program stored in a medium executed by a computer or a recording medium including instructions executable by a computer. have. In addition, the method of predicting the price of an asset in the price prediction apparatus described with reference to FIGS. 1 to 7 may be implemented in the form of a computer program stored in a medium executed by a computer.

Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include a computer storage medium. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

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

100: price prediction device
110: collection unit
120: first prediction unit
130: derivation part
140: generator
150: second prediction unit

Claims (7)

In the asset price prediction device using C&N,
a tree module that receives data fitted with a difference as much as the time series length of the reference date (T) and the forecast operating period (K) between the input data of the asset to be predicted and the past price data of the asset;
A CNN module that receives an image map generated based on the comparison value between the predicted price data predicted through the tree module and the actual price data corresponding to the prediction operating period (K) and the variables of the image map;
Learning about the predicted price data through the CNN module, predicting an image map corresponding to a previous point of time to be predicted through the CNN module, and based on the predicted image map at a specific time point to be predicted To predict the final forecast price data for
Asset price prediction device using CNN. According to claim 1, wherein the tree module,
Predicting the forecast price data corresponding to the period between the reference date and the time point of adding the forecast operating period to the reference date,
Asset price prediction device using CNN. The method of claim 1,
The input data is characterized in that it is an index price that recalculates the prices of the assets to be predicted,
The index includes one or more of KOSPI (domestic), S&P 500 (USA), EMBI (each country),
Asset price prediction device using CNN. According to claim 1, wherein the comparison value,
Derived from the average value for the forecast operating period,
Asset price prediction device using CNN. The method of claim 1, wherein the parameters of the image map are:
A first period value set on the basis of the time when the predicted operating period (K) is subtracted from the reference date (T);
including a second period value set based on the time point of subtracting the predicted operating period (K) from the reference date (T),
Asset price prediction device using CNN. The method of claim 5, wherein the image map,
At least one first image map generated based on the first period value, the second period value, and the comparison value, and a second generated by processing the generated first image maps to be stacked in chronological order containing the car image map,
Asset price prediction device using CNN. The method of claim 6,
The first image map is a two-dimensional image, and the second image map is a three-dimensional image,
Asset price prediction device using CNN.

Images