KR102591935B1 - Method of simulating time series data by combination of base time series through cascaded feature and computer device performing the same - Google Patents

Abstract

The present invention relates to a method for simulating time series data. The method for simulating time series data performed in a computing device according to the present invention includes the steps of receiving target time series data to be simulated; calculating a proportion of each underlying time series that minimizes the error between simulated time series data determined by the proportions of a plurality of underlying time series and the target time series data; and outputting the calculated specific gravity to have the minimized error. According to the present invention, a diversified time series data set can be constructed by calculating the optimized proportion of base time series that can simulate target time series data.

Description

Method of simulating step-by-step characteristics of time series data through combination of base time series and computer device performing the same {Method of simulating time series data by combination of base time series through cascaded feature and computer device performing the same}

The present invention relates to a method for simulating time series data. Preferably, it relates to a method for simulating combined time series data using underlying time series with different time-varying characteristics.

Time series data is a plurality of data with various displacement values over time, and many algorithms are being developed to predict changes in time series data.

With the development of artificial intelligence technology, neural network-based models are being designed to learn causal relationships between the characteristics of past sections and the features of the current section.

In general, recurrent neural networks have been used for natural language processing and other signal analysis by storing past information in a hidden state and performing calculations on sequences through the RNN (Recurrent neural network) structure.

However, RNN is limited in its ability to learn longer-term dependencies, and therefore, for longer-term data prediction, LSTM ((Long Shot Term Memory network) models are widely used in the field of time series prediction.

However, these neural network models are optimized for extracting the characteristics of changes in single time series data, and it is impossible to comprehensively predict changes in high-level data generated through the combination of multiple underlying time series, and are verified by focusing on representative characteristics and learning. Overfitting problems occur in the process.

Even if the composition of the lower base time series that make up the upper data is known, the change patterns may vary depending on the individual proportions, so a more specific prediction method is needed.

In particular, this problem of predicting high-level time series data can be applied to various financial products that reflect the characteristics of complex changes in the real world.

For example, in the case of an active fund that is managed to raise profits that exceed the market rate of return, it can be managed by flexibly allocating assets and selecting stocks according to the market outlook. In order to predict returns through changes in the fund, a fund is formed. It is necessary to understand the time series nature of each stock.

The purpose of the present invention is to propose a method for optimizing base time series that can simulate target time series data.

The purpose of the present invention is to propose a method for determining the proportion of underlying time series through embedding of micro and macro characteristics of target time series data.

The purpose of the present invention is to propose a method to solve the overfitting problem that focuses on representative characteristics by going through a regularization process using various significant macroscopic characteristics that are not linked to the observation time of time series data.

The purpose of the present invention is to propose a method for optimizing the proportion of the entire underlying time series through optimization of each segment of the target time series data.

In addition, the purpose of the present invention is to increase liquidity in the combination of base time series by allowing the composition and proportion of the base time series to be freely set according to the goal using upper time series data.

A method for simulating time series data performed in a computing device according to the present invention to solve the above technical problem includes receiving target time series data to be simulated; calculating a proportion of each underlying time series that minimizes the error between simulated time series data determined by the proportions of a plurality of underlying time series and the target time series data; and outputting the calculated specific gravity to have the minimized error.

The target time series data is determined according to a random ratio between the first base time series constituting the first base time series set, and the simulated time series data is determined according to the ratio between the second base time series constituting the second base time series set. desirable.

The target time series data is a first investment fund composed of the first base time series of each product in the first investment portfolio defined as a first base time series set, and the simulated time series data is a random product for the first investment fund that is not investable. It is preferable that it is a second investment fund determined according to the proportion between the second underlying time series.

The composition of the products in the second investment portfolio, which is composed of a combination of the arbitrary products, varies depending on the user, and the step of calculating the proportion is preferably to recalculate the changed proportion of the product.

In the step of calculating the specific gravity, it is preferable to calculate the specific gravity that minimizes the difference between the unit change rate of the simulated time series data and the target time series data.

The step of calculating the ratio includes calculating the ratio that minimizes the distance in a common embedding space between the first embedding vector of the target time series data extracted from the learned time series model and the second embedding vector of the simulated time series data, The time series model consists of an encoder that extracts the features of each overlapped section of time series data as an embedding vector and a decoder that generates synthetic time series data through the extracted embedding vector, while minimizing the error between the time series data and the virtual time series data. It is desirable to be taught to do so.

Calculating the specific gravity includes dimensionally expanding the target time series data and the simulated time series data; extracting topological feature vectors each defining periodic characteristics within the extended dimensional space; and calculating the specific gravity that minimizes the difference between the extracted topological feature vectors.

Calculating the specific gravity includes calculating statistical values of the target time series data and the simulated time series data; and calculating the specific gravity that minimizes the difference between the statistical values.

In the step of calculating the specific gravity, it is preferable to calculate the specific gravity that minimizes the weighted sum of the terms to be minimized in the target function through a predetermined tuning parameter.

The step of calculating the specific gravity is preferably performed repeatedly so that the weighted sum satisfies the termination condition while calculating the specific gravity that independently minimizes the term to be minimized.

A computing device according to the present invention for solving the above technical problem includes a processor; and a memory that communicates with the processor, wherein the memory stores instructions that cause the processor to perform operations, wherein the operations include receiving target time series data to be simulated, the ratio of a plurality of base time series. An operation of calculating a proportion of each base time series that minimizes the error between the determined simulated time series data and the target time series data, and an operation of outputting the proportion calculated to have the minimized error.

According to the present invention, a diversified time series data set can be constructed by calculating the optimized proportion of base time series that can simulate target time series data.

In addition, by simulating changes in real objects with time series information using a combination of underlying time series, independent decisions can be made for each underlying time series. In addition, portability to other environments is increased for general stock products such as funds. By creating similar combinations of underlying stocks in a new environment, various derivative investment methods can be provided.

In addition, by combining offshore financial products that cannot be directly invested in Korea with products that can be invested in Korea, it is possible to create the effect of indirectly investing in the same product at a relatively low cost.

In addition, investment freedom can be increased by increasing the liquidity of fund products and replicating the characteristics of the target fund according to individual decision-making.

Figure 1 is an exemplary diagram showing the configuration of target time series data according to an embodiment of the present invention.
Figure 2 is an exemplary diagram showing the configuration of simulated time series data according to an embodiment of the present invention.
Figure 3 is a flowchart showing a time series data simulation method according to an embodiment of the present invention.
Figure 4 is an exemplary diagram showing a recursive simulation method of time series data according to an embodiment of the present invention.
Figure 5 is an exemplary diagram showing a learning pipeline for simulation of time series data according to an embodiment of the present invention.
Figure 6 is an exemplary diagram showing a simulation method using unit rate of change according to an embodiment of the present invention.
Figures 7 and 8 are exemplary diagrams showing a simulation method using a time series model according to an embodiment of the present invention.
Figure 9 is a flowchart showing a simulation method considering topological characteristics according to an embodiment of the present invention.
Figures 10 and 11 are exemplary diagrams showing a simulation method considering topological characteristics according to an embodiment of the present invention.
Figure 12 is a flowchart showing a simulation method considering statistical characteristics according to an embodiment of the present invention.
Figure 13 is an exemplary diagram showing an implementation in the form of a computing device that performs a simulation method according to embodiments of the present invention.

The following merely illustrates the principles of the invention. Therefore, a person skilled in the art can invent various devices that embody the principles of the invention and are included in the concept and scope of the invention, although not clearly described or shown herein. In addition, all conditional terms and embodiments listed in this specification are, in principle, clearly intended only for the purpose of ensuring that the inventive concept is understood, and should be understood as not limiting to the specifically listed embodiments and states. .

The above-mentioned purpose, features and advantages will become clearer through the following detailed description in relation to the attached drawings, and accordingly, those skilled in the art in the technical field to which the invention pertains will be able to easily implement the technical idea of the invention. .

Additionally, when describing the invention, if it is determined that a detailed description of known technology related to the invention may unnecessarily obscure the gist of the invention, the detailed description will be omitted. Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

Figure 1 is a diagram illustrating the configuration of target time series data to be simulated according to an embodiment of the present invention.

Referring to Figure 1, the target time series data is a plurality of base time series ( ) set of It can be composed of and has integrated time-varying characteristics according to the determined proportion between each underlying time series.

These target time series data may, for example, be the temporal changes of a comprehensive fund composed of a combination of various investment items in a fluctuating stock market, and the overall time-varying data may vary depending on the know-how of the fund manager or various decisions in response to external environmental variables. Characteristics can be determined.

For example, the proportion of each stock that makes up the fund can be set according to various external environmental factors that affect the target time series data, and the changing pattern of the target time series data can be continuously updated to meet the investment purpose. You can.

At this time, investment products such as offshore funds based in the United States and Europe charge relatively high fees, and there is a problem in that it is difficult to determine the composition of the specific investment portfolio and the proportions of the components from the outside in a timely manner. Alternatively, depending on the policies of the local investment market, there may be various restrictions, such as restrictions on investment in Korea or high time and financial costs.

Therefore, the simulation method according to this embodiment defines an investment fund consisting of the underlying time series of each product in the investment portfolio as the target time series data, and simulates the investment fund with the underlying time series of any product through an investment impossible or relatively low fee. make it possible

Furthermore, when replicating a fund, the composition of products in the investment portfolio can be varied depending on the user, taking into account the individual's investment strategy and taste, but it is also possible to grant autonomy by replicating the variable product composition.

In other words, the time series data simulation method according to this embodiment can best simulate the target time series data with a set of the same or different base time series without knowing the composition of the base time series or the respective proportions of the target time series data. The purpose is to propose a method for generating simulated time series data that can be simulated.

When described with reference to FIG. 2, FIG. 2 is a diagram illustrating the configuration of time series data that simulates target time series data.

Simulated time series data is a set of underlying time series composed of various underlying time series. It can be created with

At this time, the base time series set may be partially or completely different from the composition of the set constituting the target time series data. Alternatively, if they have the same configuration, the time series data simulation method according to the present invention can be performed in the form of predicting the proportions without knowing the proportions of each underlying time series constituting the target time series data.

The basis time series set U ₀ = {x ₁ (t), ..., x _m (t)} constituting the given target time series data (S ^* (t) ≡ Signal ^* (t)) according to this embodiment is m. Assuming that it is composed of a certain unknown linear combination of different underlying time series, each x _i (t) can be generated from some unknown stochastic process.

At this time, it can be assumed that the composition and proportions w ^* ₁ , ..., w ^* _m of the underlying time series set U ₀ are generally unknown.

Therefore, the time series data simulation method according to this embodiment can be defined as the problem of configuring a set of base time series with the most similar pattern to the target time series data and finding the optimized proportion of each base time series.

That is, in this embodiment, in the time series data simulation method, the dynamic characteristics of the target time series data signal are obtained from the base time series set U = {y ₁ (t), ..., y _n (t)} that constitutes the simulated time series data. Best simulated time series data S(t) ≡ The purpose is to synthesize (t).

Referring to FIG. 3, the time series data simulation method according to this embodiment is performed repeatedly to minimize the error with the target time series data by updating the proportion of the base time series set constituting the simulated time series data through a recursive method. You can.

Specifically, in this embodiment, the time series data simulation method is based on an unsupervised learning model and topological analysis to simulate the macroscopic dynamic characteristics of a given time series as closely as possible, thereby simulating existing simple statistical-based time series data. A more accurate simulation can be performed using this method.

Hereinafter, a method for simulating time series data performed in a computing device according to an embodiment of the present invention will be described in more detail with reference to FIG. 4.

Referring to Figure 4, first, target time series data to be simulated can be input (S100).

As described above, the target time series data (S(t) ≡ Signal ^* (t)) can be determined by a set U ₀ composed of base time series x _m (t) and the proportion w ^* _m of each base time series.

Next, the time series data simulation method according to this embodiment calculates the optimal ratio by repeatedly updating the ratio that minimizes the error between the simulated time series data and the target time series data determined by a plurality of base time series and each ratio for simulation. (S200).

This process of repeatedly updating the ratio can be performed by dividing the characteristics of the target time series data according to size from relatively small range of microscopic characteristics to overall macroscopic characteristics and optimizing each divided part.

Specifically, referring to FIG. 5, the learning pipeline for updating the proportion is divided into a process of extracting each embedding vector divided into target time series data and the second time series set constituting the simulated time series data as input.

In this embodiment, the learning pipeline can sequentially segment and simulate the time-varying characteristics of the target time series data based on four types of similarity measurement functions.

For example, unit change characteristics can be simulated through the minimum unit change rate of the first size, which is the smallest of the subdivided sizes, and time-varying characteristics of a larger section of the second size can be simulated using a learned neural network model.

Next, by defining the periodic change of the section as a topological characteristic at the third size, time-varying characteristics can be simulated, and at the fourth size, simulation can be performed to reduce the difference in statistical values due to the global characteristics of both time series data. .

As the size of dividing time series data to extract characteristics increases, the proportion of macroscopic similarities between both time series data may increase. At this time, the macroscopic similarity proportion can be determined by the tuning parameters of the learning pipeline itself and can be adjusted during the learning process.

Hereinafter, each learning pipeline will be described in more detail.

First, the unit change measurement module 134-1 in the learning pipeline extracts the unit change rate of the section divided based on the measurement unit (first size) of the time series data in order to simulate the most minimized characteristics and extracts the unit change rate vector of the unit change rate vector. It includes the process of calculating the difference.

Referring to FIG. 6, the times of the target time series data and the simulated time series data can be synchronized, and the difference between the change rates for each unit divided based on the measurement unit of the target time series data is calculated.

The unit change measurement module 134-1 measures the difference in unit change rate ((t)=δS(t)-δ (t)) can be calculated for each section.

Subsequently, the unit change measurement module 134-1 is the sum of the unit change rate differences for each unit. The proportion that can be minimized Calculate .

Furthermore, in order to simulate the characteristics of a larger range of target time series data, changes for each section according to the second size can be embedded.

The section change measurement module 134-2 embeds the change characteristics for each section with respect to the determined section T and minimizes the difference between the embedded section embedding vectors. Optimize.

At this time, sections are allowed to overlap with each other so that characteristics that affect changes between sections can be further emphasized.

In this embodiment, a pre-trained model can be used for section embedding, and the model can be a time series model that extracts recursive characteristics, preferably an LSTM model.

At this time, the LSTM model can be trained in advance to well extract the implicit features of the target time series characteristics as an embedding vector, and a unique second learning pipeline can be additionally configured for learning.

Referring to FIG. 7, the second learning pipeline may be composed of an encoder that extracts the characteristics of each section of the original time series data into an embedding vector of a determined dimension and a decoder that generates synthetic time series data through the extracted embedding vector. Continuous LSTM model learning can be performed by minimizing the error between the input time series data and the virtual time series data output from the decoder by configuring it as a loss function.

In this embodiment, the loss function for learning the LSTM model can be composed of the mean squared error (MSE) between the original time series data and the generated virtual time series data and the error between the feature vectors in the embedding dimension. .

At this time, the error between feature vectors can be calculated through the distance between embedding vectors within the determined embedding dimension.

Specifically, in this embodiment, a false negative neighbor (FNN) can be used to estimate the dimension of the embedding vector, which is prerequisite for extracting the embedding vector.

False nearest neighbors calculate the number of neighbors of a point as the embedding dimension increases. If the embedding dimension is too low, many neighbors are false, but above an appropriate embedding dimension, many neighbors are true. This is the dimension that best represents the characteristics of the data. decide.

In other words, the most appropriate embedding dimension for reducing time series data in an LSTM model can be determined by determining how the number of false neighbors changes as a function of dimension as the dimension increases.

At this time, the LSTM model used in this embodiment can be learned using a learning data set collected from a common domain of target time series data, for example, a similar fund of a specific fund in the stock market or a fund within the same country. Robust performance can be achieved by using the fluctuations in the index as a learning data set.

Through the learned LSTM model above, the embedding vectors of the target time series data and the embedding vectors of the simulated time series data are calculated and the proportions are optimized to minimize the error between each vector.

Specifically, the error between embedding vectors (L _FNN ) is optimized by calculating the distance of the defined embedding vectors in the determined dimension.

Referring to FIG. 8, the ^distance between the feature vectors (x _{i , ...} d _j ) and can be calculated as the sum of each distance ) is ultimately minimized. optimize

Furthermore, the time series data simulation method according to this embodiment extracts a larger range of features and performs simulation using the features.

Preferably, the phase measurement module 134-3 extracts topological features for a section larger than the section used in the above-described LSTM model.

In this embodiment, the phase measurement module 134-3 has very robust characteristics compared to other algebraic characteristics by selecting and statistically quantifying phase characteristics that appear consistently regardless of the observation period.

Referring to FIG. 9, first, the phase measurement module can dimensionally expand the target time series data and the simulated time series data (S210).

For example, when mapping time series data defined as displacement values on one dimension to a higher dimension (K dimension), the repetitive characteristics of the time series data may be further emphasized.

Specifically, by performing time delay embedding of time series data in K dimension, features overlapping with delayed time are expanded and structured in a high dimension. Therefore, in an expanded dimension, the interdependent characteristics of different time periods can be better emphasized.

For example, the topological characteristic of time series data extracted according to this embodiment may be periodicity or quasi-periodicity. The phase measurement module 134-3 extracts (quasi-) periodic characteristics of time series data in the K dimension, at least part of which is repeated at a delay time determined through time delay embedding, as a topological feature vector (S220).

Referring to FIG. 10, the characteristics of time series data (a) with periodicity may appear with more emphasis on repeated patterns in the delayed time of the K dimension, while the characteristics of time series data (b) with weak periodicity may appear scattered within the K dimension. A pattern can be seen.

Therefore, the phase measurement module 134-3 extracts only statistically significant homology characteristics through the characteristics represented by the patterns within the above K dimensions, thereby enabling the determination of more global characteristics of the time series data.

Specifically, referring to FIG. 11, in order to calculate the difference in topological feature vectors, the K-dimensional phase vector can be compressed back to a lower dimension with respect to the homology space related to periodicity (e.g., H_1, H_2), for example For example, by one-dimensionalizing the compressed phase vector, a persistence landscapes vector representing the characteristics of data persistence can be generated.

In the landscape vector, which represents the topological characteristics of such time series data, the denser the distribution, the stronger the specific periodicity, and the time series data itself can be seen as being relatively less influenced by external factors and having a greater tendency to maintain patterns.

Therefore, the phase measurement module 134-3 can prevent errors due to the environment from occurring by well simulating the topological characteristics of the target time series data so that the simulated time series data well reflects the influence of the actual environment or maintains periodicity. You can.

Next, the phase measurement module 134-3 calculates the specific gravity to minimize the difference between the extracted topological feature vectors (S230). The phase measurement module 134-3 calculates the error between the landscape vectors, and minimizes the sum of the errors between each part (H ₁ , H ₂ ) landscape vectors. Optimize,

Furthermore, in this embodiment, the learning pipeline is also capable of determining the proportion of internal underlying time series of the simulated time series data using statistical characteristics.

The statistical module 134-4 calculates statistical values using the most global characteristics of time series data and compares the differences between statistical values. can be optimized.

Referring to FIG. 12, the statistical module 134-4 calculates statistical values of the target time series data and the simulated time series data (S210-1).

In this embodiment, the statistical module 134-4 can use various statistical indicators as a similarity comparison measure of time series data. For example, mean, variance, correlation coefficient (Pearson correlation coefficient), and DTW (Dynamic time warping) can be comprehensively used.

When using multiple statistical indicators comprehensively, the statistical module 134-4 sets the weight for each indicator in the direction of minimizing the difference from the target time series data. The specific gravity is calculated by optimizing .

At this time, the error proportion between each statistical indicator can be set separately (for example, 2:1:1:1)) and the proportion is calculated so that the total error sum is minimized (S210-2).

Referring again to FIG. 5, the optimization process of each module in the learning pipeline may require additional, complementary processes by concentrating on characteristics of different sizes and simulating target time series data.

Therefore, the learning pipeline according to this embodiment can perform an iterative update process according to predetermined constraints and termination conditions.

Each weight as a constraint You can set the minimum and maximum values, or the historical proportion in the process of continuous updating of simulated time series data. Within a certain range based on ( ) only the specific gravity It is also possible to set it so that can be determined.

In addition, in this embodiment, the learning pipeline is a problem of minimizing the overall error by optimizing the proportion according to each size, and the objective function (QP (Qudratic Program) Formulate) of the learning pipeline is defined as the equation below: And the specific gravity Calculate .

[Equation]

(Here, Rw is the unit change rate of the simulated time series data, r _tgt is the unit change rate of the target time series data, Ew is the embedding vector extracted as the time series model for the simulated time series data, and e _tgt is the time series model extracted as the target time series data. Embedding vector, Tw is the landscape vector for the simulated time series data, t _tgt is the landscape vector for the simulated time series data, Sw, S _tgt are the statistical values of the simulated time series data and target time series data, respectively)

The optimization process performs optimization within the range that satisfies the conditions defined in the set C that defines the constraints, but the proportion of each underlying time series is Set the sum to be 1.

When performing time series simulation, practicality dictates that only a subset of the entire available set of underlying time series be used. For this purpose, linear constraints can be included to ensure that the number of underlying time series with positive weight does not exceed a maximum of k, and in this case, the optimization model is mixed-integer quadratic programming.

Additionally, the weights λ _e , λ _t , and λ _s for the terms in the objective function can be preset as tuning parameters, and the tuning parameters can be calculated by alternately performing learning and verification within the learning data set.

In this embodiment, the learning pipeline sequentially optimizes each term in the function for optimization and calculates the proportion by reducing the overall error.

In other words, more efficient optimization is performed by defining the optimization problem of the goal function as QP (Quadratic Program). In this embodiment, the learning pipeline is performed through iterative optimization of each part that constitutes QP, and in particular, it is an ADMM (Alternating direction method of multipliers) solver for the most advanced form of QP in big data optimization. Optimization of the components that make up each problem is performed at high speed through OSQP (Operator Splitting QP).

In particular, in this embodiment, the macroscopic similarity-related indicators can be updated at a relatively slow cycle, so the proposed module can be performed in virtually real time on the collected time series data.

Hereinafter, a detailed hardware implementation of the computing device 300 that performs a method for simulating time series data according to an embodiment of the present invention will be described.

Referring to FIG. 11, in some embodiments of the present invention, the computing device 300 may be implemented in the form of a server. One or more of each module constituting the computing device 300 is implemented on a general-purpose computing processor, and thus includes a processor 308, input/output I/O 302, memory 340, and interface. It may include 306 and bus 314. The processor 308, input/output device 302, memory 340, and/or interface 306 may be coupled to each other through a bus 314. The bus 314 corresponds to a path through which data moves.

Specifically, the processor 308 includes a Central Processing Unit (CPU), Micro Processor Unit (MPU), Micro Controller Unit (MCU), Graphic Processing Unit (GPU), microprocessor, digital signal processor, microcontroller, and application processor (AP). , application processor) and logic elements capable of performing similar functions.

The input/output device 302 may include at least one of a keypad, a keyboard, a touch screen, and a display device. The memory device 340 may store data and/or programs.

The interface 306 may perform the function of transmitting data to or receiving data from a communication network. Interface 306 may be wired or wireless. For example, the interface 306 may include an antenna or a wired or wireless transceiver. The memory 340 is a volatile operating memory that improves the operation of the processor 308 and protects personal information, and may further include high-speed DRAM and/or SRAM.

Additionally, memory 340 stores programming and data configurations that provide the functionality of some or all of the modules described herein. For example, it may include logic to perform selected aspects of the method for simulating time series data described above.

A program or application is loaded with a set of instructions including each step of performing the above-described time series data simulation method stored in the memory 340 and allows the processor to perform each step.

Furthermore, various embodiments described herein may be implemented in a recording medium readable by a computer or similar device, for example, using software, hardware, or a combination thereof.

According to hardware implementation, the embodiments described herein include application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), and field programmable gate arrays (FPGAs). It may be implemented using at least one of processors, controllers, micro-controllers, microprocessors, and other electrical units for performing functions. In some cases, as described herein, The described embodiments may be implemented as a control module itself.

According to software implementation, embodiments such as procedures and functions described in this specification may be implemented as separate software modules. Each of the software modules may perform one or more functions and operations described herein. Software code can be implemented as a software application written in an appropriate programming language. The software code may be stored in a memory module and executed by a control module.

The above description is merely an illustrative explanation of the technical idea of the present invention, and various modifications, changes, and substitutions can be made by those skilled in the art without departing from the essential characteristics of the present invention. will be.

Accordingly, the embodiments disclosed in the present invention and the accompanying drawings are not intended to limit the technical idea of the present invention, but are for illustrative purposes, and the scope of the technical idea of the present invention is not limited by these embodiments and the attached drawings. . The scope of protection of the present invention should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be construed as being included in the scope of rights of the present invention.

Claims (19)

In a method for simulating time series data performed on a computing device,
A step of receiving target time series data to be simulated;
calculating a proportion of each underlying time series that minimizes the error between simulated time series data determined by the proportions of a plurality of underlying time series and the target time series data; and
Including the step of outputting the calculated specific gravity to have the minimized error,
The step of calculating the specific gravity is,
The distance within the common embedding space between the first embedding vector of the target time series data and the second embedding vector of the simulated time series data for each first section extracted from the learned time series model, and
Calculate the proportion that minimizes the difference between the topological feature vectors of the target time series data and the simulated time series data for each second section extracted from the phase measurement module,
The time series model consists of an encoder that extracts the features of each overlapping first section of the time series data as an embedding vector and a decoder that generates synthetic time series data through the extracted embedding vector, and between the time series data and the virtual time series data. Learned to minimize errors,
The phase measurement module is a method for simulating time series data, characterized in that the topological feature vector of the second section is larger than the first section. According to claim 1,
The target time series data is determined according to a random ratio between the first base time series constituting the first base time series set,
A method for simulating time series data, wherein the simulated time series data is determined according to the proportion between the second base time series constituting the second base time series set. According to claim 2,
The target time series data is a first investment fund consisting of a first base time series of each product in the first investment portfolio defined as a first base time series set,
A method of simulating time series data, characterized in that the simulated time series data is a second investment fund determined according to the ratio between the second base time series of an arbitrary product with respect to the first investment fund that cannot be invested. According to claim 3,
The composition and optimization form of the product in the second investment portfolio, which is composed of a combination of the above arbitrary products, is implemented in a variable manner according to the user's purpose and inclination,
The step of calculating the specific gravity is a method of simulating time series data, characterized in that the specific gravity of the changed product is recalculated when necessary. According to claim 2,
The step of calculating the specific gravity is,
A simulation method of time series data, characterized in that calculating the specific gravity that minimizes the difference between the unit change rate of the simulated time series data and the target time series data. delete According to claim 2,
The step of calculating the specific gravity is,
dimensionally expanding the target time series data and the simulated time series data;
extracting topological feature vectors each defining periodic characteristics within the extended dimensional space; and
A method for simulating time series data, comprising calculating the proportion that minimizes the difference between the extracted topological feature vectors. According to claim 2,
The step of calculating the specific gravity is,
calculating specific statistical values describing characteristics of the target time series data and the simulated time series data; and
A method for simulating time series data, comprising calculating the specific gravity that minimizes the difference between the statistical values. According to claim 2,
The step of calculating the specific gravity is,
A simulation method of time series data characterized by calculating a proportion that minimizes the weighted sum of the terms to be minimized within the objective function through predetermined tuning parameters. According to clause 9,
The step of calculating the specific gravity is,
Calculate the proportion that independently minimizes the term to be minimized,
A method for simulating time series data, characterized in that the weighted sum is repeatedly performed to satisfy a termination condition. processor; and
comprising a memory in communication with the processor,
The memory stores instructions that cause the processor to perform operations,
The above operations are,
The operation of receiving target time series data to be copied,
An operation of calculating the proportion of each underlying time series that minimizes the error between simulated time series data determined by the proportions of a plurality of underlying time series and the target time series data, and
Including an operation of outputting the calculated specific gravity to have the minimized error,
The operation of calculating the specific gravity is,
The distance within the common embedding space between the first embedding vector of the target time series data and the second embedding vector of the simulated time series data for each first section extracted from the learned time series model, and
Calculate the proportion that minimizes the difference between the topological feature vectors of the target time series data and the simulated time series data for each second section extracted from the phase measurement module,
The time series model consists of an encoder that extracts the features of each overlapping first section of the time series data as an embedding vector and a decoder that generates synthetic time series data through the extracted embedding vector, and between the time series data and the virtual time series data. Learned to minimize errors,
The phase measurement module is a computing device characterized in that it extracts the topological feature vector of the second section that is larger than the first section. According to claim 11,
The target time series data is determined according to a random ratio between the first base time series constituting the first base time series set,
A computing device wherein the simulated time series data is determined according to the proportion between the second basis time series constituting the second basis time series set. According to claim 12,
The operation of calculating the specific gravity is,
Computing device for calculating the specific gravity that minimizes the difference between the unit rate of change of the simulated time series data and the target time series data. delete According to claim 12,
The operation of calculating the specific gravity is,
An operation of dimensionally expanding the target time series data and the simulated time series data,
An operation of extracting topological feature vectors defining periodic characteristics within the extended dimensional space, respectively, and
Computing device comprising calculating the specific gravity that minimizes the difference between the extracted topological feature vectors. According to claim 12,
The operation of calculating the specific gravity is,
An operation of calculating statistical values of the target time series data and the simulated time series data, and
Computing device comprising calculating the specific gravity that minimizes the difference between the statistical values. According to claim 12,
The operation of calculating the specific gravity is,
A computing device characterized in that it calculates a proportion that minimizes the weighted sum of terms to be minimized within a goal function through predetermined tuning parameters. According to claim 17,
The operation of calculating the specific gravity is,
Calculate the proportion that independently minimizes the term to be minimized,
A computing device characterized in that the weighted sum is repeatedly performed to satisfy a termination condition. A program stored in a computer-readable recording medium that performs a method of simulating time series data performed by the computing device according to any one of claims 1 to 5 and 7 to 10.

Images