Prediction system, prediction program and prediction method, as well as stock trading device and stock trading method

Abstract

To provide more accurate predictions of future levels of interest on the internet compared to conventional technologies. [Solution] A prediction system (S) that predicts the future level of interest in a keyword based on time-series data of frequency information related to the keyword transmitted to a server (3,4), comprising: data acquisition means (Ca2) for acquiring time-series data; data division means (Ca4) for dividing the time-series data into predetermined time intervals; frequency analysis means (Ca5) for analyzing the frequency components of the time-series data for each time interval; and determination means (Ca6) for determining whether or not there will be an increase in future frequency information based on a comparison of the amplitude intensity of low-frequency components in the target time interval with the amplitude intensity of low-frequency components in past time intervals.

Description

[Technical Field]

[0001] The present invention relates to a prediction system, prediction program, and prediction method for predicting the future level of interest in keywords transmitted to a server through searches, posts, etc., and to a stock trading device and stock trading method for buying and selling stocks. [Background technology]

[0002] With the recent development of the information society, individuals and companies are actively disseminating information via the internet. Information dissemination via the internet is not limited to information disseminated by companies through news sites and homepages, but is also spread through information sharing and posting by individuals via SNS (Social Networking Services) and social media. Consequently, information dissemination via the internet is faster and on a larger scale than information dissemination via traditional mass media. Along with this, excessive reactions and excessive interest can occur, including information on social issues, corporate scandals, and information of uncertain veracity. In other words, the activity level of internet users (user dynamics) can become overheated. Excessive reactions can be positive (so-called "going viral"), but they can also be negative (so-called "internet firestorm"), and excessive reactions and excessive interest (overheating of user dynamics) have become a social problem. "Internet firestorm" is sometimes defined as "a state in which a flood of criticism erupts and spreads online in response to actions or statements made both offline and online."

[0003] The following conventional technology 1 is known as a technique for detecting the occurrence of online flame wars. Conventional technology 1 (Non-patent document 1) describes a technology that uses social listening tools or human monitoring to determine if the number of social media posts related to a company exceeds a threshold, and then notifies the user of an alert (warning) regarding the risk of online firestorms. However, this method requires monitoring a vast amount of information, resulting in a high processing load, which may delay the detection of online firestorm risks. Therefore, it is also described that instead of targeting all posts, the target posts should be narrowed down according to the company. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] "Social Risk Countermeasures: Issues in Detecting Online Firestorm Risk in Social Listening (Part 7)," NTT Com Online, December 13, 2017.<URL:https: / / www.nttcoms.com / service / social / column / 20171213 / > [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] (Problems with conventional technology) In conventional technologies, early detection of online firestorms requires appropriately narrowing down the target posts. Methods for narrowing down posts include analyzing the sentiment of the poster as gleaned from the content, and utilizing user metadata (ancillary information). However, sentiment analysis requires human monitoring by experts, as well as the manual selection and extraction of specific keywords likely to be used in negative sentiments. The results are heavily dependent on the experience and skill of the monitors, potentially delaying the detection of online firestorms. Conventional technologies, regardless of the method used, assume that related posts appear after relevant information has spread online; therefore, detection is only possible after relevant information such as company names has become widespread. In addition, in the case of a company dealing with seasonal products, the number of posts may increase or decrease seasonally, and since an increase in the amount of posted related information does not directly lead to an online explosion, it is impossible to detect signs of an online explosion at an early stage with the conventional technology that analyzes the quantity of the number of posts. That is, with the conventional technology, it is possible to detect ongoing online explosions associated with events or scandals that have already become publicly known socially, but it is difficult to detect future online explosions triggered by unpublished information.

[0006] Compared with the prior art, the technical problem of the present invention is to accurately predict the future degree of interest on the Internet and appropriately buy and sell stocks based on the predicted future degree of interest.

Means for Solving the Problem

[0007] To solve the above technical problem, the invention according to claim 1 is A prediction system that predicts the future degree of interest in a keyword based on time-series data of frequency information related to the keyword transmitted to a server, Data acquisition means for acquiring the time-series data, Data classification means for classifying the time-series data into predetermined time intervals, Frequency analysis means for analyzing the frequency components of the time-series data for each time interval, Determination means for determining whether the frequency information will increase in the future compared to the target time interval based on a comparison between the amplitude intensity of a predetermined low-frequency component in the target time interval and the amplitude intensity of the low-frequency component in a time interval prior to the target time interval, A prediction system comprising the above. <000X088>

[0008] The invention according to claim 2 is The determination means determines that the frequency information will increase in the future compared to the target time interval when the amplitude intensity of the low-frequency component in each time interval increases with the passage of time. The prediction system according to claim 1.

[0009] The invention according to claim 3 is wherein the frequency analysis means analyzes frequency components by performing a fast Fourier transform on the time-series data. A prediction system according to claim 1.

[0010] The invention according to claim 4 is wherein the data classification means is configured to be able to change the length of the time interval according to a setting based on a user input. A prediction system according to claim 1.

[0011] The invention according to claim 5 is further comprising analysis means for analyzing a second keyword transmitted to the server together with the keyword to analyze whether an increase in the degree of interest in the keyword is positive or negative. A prediction system according to claim 1.

[0012] The invention according to claim 6 is wherein the determination means determines that the frequency information in the future of the target time interval increases when the difference between the amplitude intensity of a predetermined low-frequency component in the target time interval and the average value of the amplitude intensity of the low-frequency component in a time interval prior to the target time interval is greater than a threshold value determined based on the variance of the amplitude intensity of the low-frequency component in a time interval prior to the target time interval, in the prediction system according to claim 1.

[0013] The invention according to claim 7 is wherein the average value is an exponentially smoothed average and the variance is an exponentially smoothed variance, in the prediction system according to claim 6.

[0014] In order to solve the above technical problem, the invention according to claim 8 is a prediction program for predicting the future degree of interest in a keyword based on time-series data of frequency information related to the keyword transmitted to a server, causing a computer to Data acquisition means for acquiring the aforementioned time-series data, Data segmentation means for dividing the aforementioned time series data into predetermined time intervals, Frequency analysis means for analyzing the frequency components of the time series data for each of the aforementioned time intervals, A determination means for determining whether or not there is an increase in the frequency information in the future beyond the target time period, based on a comparison between the amplitude intensity of a predetermined low-frequency component in a target time period and the amplitude intensity of the low-frequency component in a time period prior to the target time period. This is a prediction program designed to function as such.

[0015] To solve the aforementioned technical problems, the invention described in claim 9 is: The time-series data of frequency information related to the keywords sent to the server is divided into predetermined time intervals. The frequency components of the time series data are analyzed for each of the aforementioned time intervals. Based on a comparison of the amplitude intensity of a predetermined low-frequency component in the target time interval with the amplitude intensity of the low-frequency component in a time interval prior to the target time interval, it is determined whether or not there is an increase in the frequency information in the future beyond the target time interval. This is a prediction method that forecasts the future level of interest in the given keyword.

[0016] To solve the aforementioned technical problems, the invention described in claim 10 is: A frequency information acquisition means for acquiring whether or not there will be an increase in future frequency information of a company name, as predicted by the prediction system according to any one of claims 1 to 7, A short-selling means that, based on whether or not the frequency information of a company name acquired by the frequency information acquisition means will increase in the future, sells shares of a company corresponding to the company name, and buys back the shares based on the share price or the repayment deadline of the shares, This is a stock trading device equipped with [specific features / features].

[0017] To solve the aforementioned technical problems, the invention described in claim 11 is: Computers Using the prediction method described in claim 9, predict whether or not there will be an increase in future frequency information of company names. This is a stock trading method in which, based on whether or not the frequency information of the aforementioned company name will increase in the future, shares of the company corresponding to the aforementioned company name are sold, and the shares are bought back based on the share price of the aforementioned shares or the repayment deadline of the aforementioned shares. [Effects of the Invention]

[0018] According to the inventions described in claims 1, 8, and 9, future predictions of interest levels on the internet can be made with greater accuracy compared to the prior art. According to the invention described in claim 2, since the amplitude intensity of the low-frequency component increases over time, it is possible to determine that future frequency information is increasing. According to the invention described in claim 3, frequency components can be analyzed in a short time using the Fast Fourier Transform. According to the invention described in claim 4, the length of the time interval can be changed according to the accuracy of the analysis required by the user. According to the invention described in claim 5, by including the second keyword in the analysis, it is possible to analyze whether the excessive increase in interest is positive or negative. The inventions described in claims 6 and 7 can accurately predict the increase in interest. The inventions described in claims 10 and 11 enable the trading of stocks with high profit margins. [Brief explanation of the drawing]

[0019] [Figure 1] Figure 1 is an overall diagram illustrating the prediction system of the present invention. [Figure 2] Figure 2 is a functional block diagram of the control unit of the prediction computer in Example 1. [Figure 3] Figure 3 illustrates the influence between users that propagates via links on social networks. [Figure 4] Figure 4 is an explanatory diagram of the Fourier transform of the waveform of the influence transmitted between users. [Figure 5]Figure 5 illustrates the change in the shape of the characteristic polynomial of the Laplacian matrix associated with continuous structural changes in the network. [Figure 6] Figure 6 is an explanatory diagram of the low-frequency modes that appear in vibrational energy. [Figure 7] Figure 7 is an explanatory diagram of the spectral analysis method. [Figure 8] Figure 8 is an explanatory diagram illustrating an example of experimental results, showing the time series of searches for "Fukuoka SoftBank Hawks (registered trademark)" and its frequency spectral distribution. [Figure 9] Figure 9 is an explanatory diagram illustrating an example of experimental results, showing the time series of searches for "job hunting" and its frequency spectral distribution. [Figure 10] Figure 10 is an explanatory diagram illustrating an example of experimental results, showing the time-series data of search volume for "Used Car Sales Company B". [Figure 11] Figure 11 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 10. [Figure 12] Figure 12 is an explanatory diagram illustrating an example of experimental results, and is an explanatory diagram of the time-series data of search volume for "Japan Series". [Figure 13] Figure 13 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 12. [Figure 14] Figure 14 is an explanatory diagram illustrating an example of experimental results, showing the time-series data of search volume for "Ukraine". [Figure 15] Figure 15 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 14. [Figure 16] Figure 16(a) is a graph showing the trend of the lowest frequency component related to the keyword "Used Car Sales Company B" (solid line), the time when a precursor was detected using equation (5) (dotted line), and the date and time when media coverage actually occurred (marked with an "x"). Figure 16(b) is a graph showing the trend of the lowest frequency component related to the keyword "Television Station F" (solid line), the time when a precursor was detected using equation (5) (dotted line), and the date and time when media coverage actually occurred (marked with an "x"). [Figure 17]Figure 17(a) shows a stock trading device and prediction system according to Example 3, and Figure 17(b) is a functional block diagram of the stock trading device. [Figure 18] Figure 18 shows an example of the experimental results for Example 3. [Modes for carrying out the invention]

[0020] Next, specific examples of embodiments of the present invention (hereinafter referred to as "examples") will be described with reference to the drawings, but the present invention is not limited to the following examples. In the following explanation using diagrams, diagrams of components other than those necessary for the explanation have been omitted as appropriate for ease of understanding. [Examples]

[0021] Figure 1 is an overall diagram illustrating the prediction system of the present invention. In Figure 1, the prediction system S of the present invention is an example of an information processing device and includes a prediction computer 1 as an example of a prediction device. The prediction computer 1 is, for example, a general personal computer and includes a computer body 1a, a keyboard 1b and a mouse 1c as examples of input devices, and a display 1d as an example of an output device. Prediction computer 1 is connected to the Internet 2, which is an example of a public network. The Internet 2 also connects to search server 3, SNS server 4, and user computers 6 used by general users. Therefore, prediction computer 1, search server 3, SNS server 4, and user computers 6 can send and receive information from each other via the Internet 2.

[0022] Search server 3 operates a search site that searches for information on the internet and transmits search results corresponding to the search words sent from the user's computer 6 via the internet 2. SNS server 4 operates an SNS site on the internet, receiving SNS posts sent from user computers 6 via the internet 2, and publishing and making public posts via the internet 2.

[0023] (Description of the control unit of the prediction computer in Example 1) Figure 2 is a functional block diagram of the control unit of the prediction computer in Example 1. In Figure 2, the prediction computer 1 of Embodiment 1 is composed of a computer device having an I / O (input / output interface) for inputting and outputting signals to and from the outside and adjusting input / output signal levels, a ROM (read-only memory) that stores programs and data for necessary startup processing, a RAM (random access memory) for temporarily storing necessary data and programs, a CPU (central processing unit) that performs processing according to the startup programs stored in the ROM, etc., and a clock oscillator, etc. Various functions can be realized by executing the programs stored in the ROM and RAM, etc. The control unit Ca of the prediction computer 1 stores basic software that controls basic operations, so-called operating system OS, a prediction program P1 as an example of an application program, and other software not shown.

[0024] (Element connected to the control unit Ca of Example 1) The control unit Ca of the prediction computer 1 receives output signals from signal output elements such as the keyboard 1b and mouse 1c. Furthermore, the control unit Ca of the prediction computer 1 outputs control signals to controlled elements such as the display 1d.

[0025] (Function of the control unit) The prediction program P1 of the control unit Ca of the prediction computer 1 in Example 1 has the following functional means (program modules) Ca1 to Ca8. The keyword memory device Ca1 stores the keywords to be predicted. Examples of keywords to be predicted include the names of companies or individuals whose online controversies are to be predicted, or the names of product names. Data acquisition means Ca2 acquires time-series data of frequency information related to the target keyword transmitted to servers 3 and 4. In Example 1, data acquisition means Ca2 acquires time-series data from search server 3 of the number of times the keyword was transmitted to search server 3 to search for the target keyword (= number of searches, an example of frequency information). Specifically, as time-series data, it acquires the number of searches for the target keyword every hour for the past two years as an example. In addition, data acquisition means Ca2 in Example 1 acquires time-series data from SNS server 4 of the number of times posts containing the target keyword were made (= number of posts, an example of frequency information). Specifically, as time-series data, it acquires the number of posts for the target keyword every hour for the past two years as an example. It should be noted that such time-series data can be obtained from data provided by publicly known services such as Google Trends. Regarding the timing of data acquisition, when the system is started or when predictions begin for new keywords, a large amount of historical data (e.g., two years' worth) is acquired. However, when monitoring the trends of a predicted keyword in real time, after the initial large amount of data (two years' worth) is acquired, data is acquired continuously (e.g., every hour).

[0026] The interval information storage means Ca3 stores time intervals that divide time-series data. The time intervals can be set and changed according to input from the keyboard 1b, mouse 1c, etc., by the user of the prediction program P1. The time intervals can be set by the user according to the accuracy of the analysis, the responsiveness and immediacy of the prediction, etc., and can be set to any value, for example, one week (168 hours) or two weeks (336 hours). The data division means Ca4 divides the time series data into predetermined time intervals. In Example 1, the data division means Ca4 divides the time series data using the time intervals stored in the interval information storage means Ca3. For example, if the time interval is 1024 hours and the time series data covers two years, 168 hours is used as one interval, and the time series data for two years (=2 × 365 days × 24 hours = 17520 hours) is divided into multiple (=17520 / 168 ≈ 104) data points of 168 hours each.

[0027] The frequency analysis means Ca5 analyzes the frequency components of time-series data for each time interval. In Example 1, the frequency analysis means Ca5 treats discrete time-series data at one-hour intervals as wave-shaped data with time on the horizontal axis and frequency information (number of searches + number of posts) on the vertical axis, and performs frequency analysis (frequency spectrum analysis). The frequency analysis can employ any conventionally known frequency analysis method, such as the Fast Fourier Transform or the Discrete Fourier Transform, but in Example 1, the Fast Fourier Transform is used as an example to perform the frequency analysis. Therefore, as an example, the frequency analysis makes it possible to calculate a frequency spectrum distribution with frequency on the horizontal axis and the amplitude intensity of each frequency component on the vertical axis. Regarding amplitude intensity, since the absolute value of the Fourier mode contributes to the vibrational energy representing the intensity of user dynamics related to online flame wars, the phase component is ignored by taking the absolute value, and the amplitude spectrum is calculated as the amplitude intensity. While the calculation of the amplitude spectrum was given as an example of amplitude intensity, it is not limited to this, and any parameter related to user dynamics can be used. For example, the power spectrum, which is the square of the amplitude spectrum, can also be used.

[0028] In the course of a week, topics that tend to become popular on weekends (increasing access numbers), or topics that become popular annually during specific times or seasons, may experience a skewed access count or volatile fluctuations. These phenomena are unrelated to online controversies and therefore need to be excluded. Accordingly, in the frequency analysis means Ca5 of Example 1, the component with frequency 0 is removed from the frequency analysis results, and the remaining frequency spectral distribution is normalized so that the whole is 1. This eliminates the effects of differences in time-series bias (DC component, skewed state) and constant multiples (volatile fluctuations), thereby eliminating the influence of a simple increase in access count (frequency) that is unrelated to online controversies.

[0029] The determination means Ca6 determines whether there is an increase in frequency information in the future beyond the target time interval, based on a comparison of the amplitude intensity of a predetermined low-frequency component in the target time interval with the amplitude intensity of the same low-frequency component in a time interval prior to the target time interval. In the determination means Ca6 of Example 1, the amplitude intensity of the lowest frequency component is used as the predetermined low-frequency component. In Example 1, the method for determining amplitude intensity is to determine that if it increases over time (for example, if it increases for three consecutive time intervals compared to the previous time interval), it indicates that future frequency information will increase, i.e., that user dynamics will increase (it may go viral or become a hot topic). The determination method may also be to determine whether or not there are instances that exceed a predetermined threshold, or if the number of instances that exceed the threshold continues for three or more consecutive time intervals.

[0030] Furthermore, the system initially performs an assessment on the entire period (2 years) of the acquired time-series data. Subsequently, it acquires frequency information every hour, and once a week's worth of this information has been accumulated, it performs frequency analysis on the new time-series data to determine the presence of low-frequency components. It then monitors the increase in amplitude intensity of the low-frequency components, and if an increase in amplitude intensity is detected, it outputs an alert to warn of increased user dynamics. Alternatively, for example, once one day's worth of frequency data has been collected, it is possible to perform frequency analysis (Fourier transform) on the most recent week's data, including the latest day's data, and then perform the analysis in a way that the time intervals overlap by six days while being shifted. Note that the specific time intervals (one day, one week, etc.) are not limited to the example values ​​and can be arbitrarily changed according to the design, specifications, required accuracy, etc. Therefore, it is possible to perform frequency analysis after half a day's worth of data has been collected, or after one week's worth of data has been collected, and the number of overlapping days and the period for frequency analysis can also be changed.

[0031] The judgment result output means Ca7 displays the judgment result from the judgment means Ca6. Therefore, if the judgment means Ca6 outputs an alert, it displays the judgment result (flaming or going viral) that indicates a potential increase in user dynamics.

[0032] The analysis means Ca8 analyzes subkeywords (second keywords) that are sent to servers 3 and 4 along with the main keyword. In Example 1, the analysis means Ca8 analyzes whether there is a favorable / positive interest or an aversion / negative interest in the main keyword based on the subkeywords used together with the main keyword during searches and posts, and their frequency information (number of searches, number of posts). For example, if the main keyword is "Used Car Sales Company A" and subkeywords such as "Good customer service" are present, it is analyzed as positive (buzzworthy), and if subkeywords such as "Persistent sales tactics" are present, it is analyzed as negative (controversial). The analysis can use known analysis methods and tools such as co-occurrence analysis and sentiment analysis. The analysis means Ca8 analyzes that if the analysis result of the most frequently occurring subkeyword is favorable / positive, it is not a controversial topic, and if it is aversion / negative, it is a controversial topic. Furthermore, the analysis means Ca8 performs a positive / negative analysis (estimation) when the determination means Ca6 outputs an alert indicating an increase in user dynamics.

[0033] Furthermore, the judgment result output means Ca7 displays not only the judgment result from the judgment means Ca6 but also the analysis result from the analysis means Ca8. Therefore, in addition to the judgment result of the judgment means Ca6 indicating an increase in user dynamics, the analysis result of either "favorable / positive" or "averse / negative" analyzed by the analysis means Ca8 is also displayed. Thus, users of the prediction system S can make judgments and estimates such as a firestorm or virality based on the judgment result from the judgment means Ca6 and the analysis result from the analysis means Ca8. In Example 1, the output is displayed on display 1d, but this is not the only option; any notification method can be used, such as sending an email to the administrator.

[0034] (Effect of Example 1) In the prediction system S of Embodiment 1, which has the above configuration, frequency spectral analysis is performed on frequency information of search counts and post counts, and when an increase in the intensity of low-frequency components is detected, the possibility of a future online firestorm is predicted. In the case of social issues or corporate scandals, before they become a hot topic, people who know the inside story may search on search engines and social media to investigate whether internal information has been leaked or is not publicly known. As a result of unusual searches being performed by a small number of people, fluctuations are mixed in with the normal search results (waveform) before the topic becomes a hot topic. These fluctuations appear as low-frequency components in frequency spectral analysis. When a scandal comes to light through official announcements, mass media reports, whistleblowing, etc., the number of accesses increases sharply, and when a firestorm occurs, the low-frequency components become even larger. Therefore, the magnitude of the low-frequency components can be used to detect whether a firestorm is occurring or to detect precursors to one. Thus, by detecting when the low-frequency components begin to increase, it becomes possible to predict future firestorms (a state of excessive interest).

[0035] Conventional technologies had the problem that predicting online controversies was affected by keyword selection. Furthermore, the keywords to narrow down varied depending on the nature of the scandal, making it extremely difficult to narrow down keywords before the scandal came to light, resulting in low accuracy of future predictions. In contrast, Example 1 determines the amplitude intensity of low-frequency components from frequency analysis of frequency information, making it possible to predict future online controversies (states of excessive interest) with greater accuracy compared to conventional techniques.

[0036] Furthermore, the prediction system S of Example 1 performs predictions not only when the user starts the prediction program P1, but also by continuously running the prediction program P1, acquiring the latest time-series data every hour and performing analysis with the latest time-series data (i.e., analysis shifted by one hour from the most recent analysis). Therefore, it is possible to monitor and supervise the rise in user dynamics (flaming or buzz) in real time. Consequently, it is possible to predict the early signs of a flaming incident, take measures to calm it down before it becomes large-scale, or, even if a large-scale flaming incident is unavoidable, to respond to and prepare for it.

[0037] (Scientific explanation and explanation of experimental (simulation) results) The following explains that when online user dynamics overheat, the manifestation of low-frequency modes in the time series of user dynamics intensity can be predicted. ◎Theoretical models of online user dynamics There are broadly three types of approaches to the study of online user dynamics. • Data science approach • Phenomenological approach • Fundamental theoretical approach The data science approach aims to understand phenomena based on the observation of real data. The phenomenological approach, on the other hand, aims to understand phenomena by constructing mathematical models that can describe specific observed phenomena. Much research is based on these two approaches, and a superficial understanding of observed phenomena is possible. However, understanding the essential structure underlying the phenomena is difficult. The third, fundamental theoretical approach assumes first principles that are undoubtedly true and aims to understand phenomena through properties logically derived from them. Locality and causality are assumed as first principles. Locality is the property that direct interactions between users occur via links in social networks, and causality is the property that, given a cause and an effect, the cause always occurs first in time.

[0038] Figure 3 illustrates the influence between users that propagates via links on social networks. Figure 4 is an explanatory diagram of the Fourier transform of the waveform of the influence transmitted between users. In this case, the influence between users propagates at a finite speed through the link, as shown in Figure 3. When the waveform of the influence between users is decomposed into trigonometric functions using the Fourier transform, as shown in Figure 4, each decomposed trigonometric function propagates at a finite speed as well. From this, it can be seen that online user dynamics are described by a wave equation on the network. The fundamental theory obtained in this way is called the oscillation model. The intensity of user activity is represented by the oscillation energy derived from the solution of the wave equation, and it is known that this gives an extension of the concepts of order centrality and betweenness centrality. Below, we will briefly explain the oscillation model according to the publicly available reference 1: Masaki Aida, Introduction to Network Dynamics, Morikita Publishing, 2020.

[0039] Consider an online social network represented by a directed graph G(V, E) consisting of n nodes, where V = {1, ..., n} is the set of nodes and E is the set of directed links. Also, for i, j ∈ V, the weights of directed links (i → j) ∈ E are defined as wi j (> 0). The adjacency matrix A = [Ai j]1≦i, j≦n of G(V, E) is defined as shown in the following equation 1.

number

[0040] Furthermore, we define the order matrix as D := diag(d1, ... dn), where di is defined by the following equation 2.

number

[0041] Next, let x(t) :=(x1(t), . . . , xn(t))τ be the state vector of the node at time t. Here, xi(t) ∈ C (i = 1, . . . , n) is the state of node i at time t. In this case, the wave equation on G(V, E) can be written as shown in equation 3 below.

number

[0042] The essential difference between the standard dynamical wave equation and the one in Math III lies in the fact that in Math III, the energy conservation law can be violated due to the effects of directed graphs. In particular, when the eigenvalues ​​of the Laplacian matrix L are not real numbers, the amplitude of the solution x(t) diverges, and the oscillation energy diverges. This phenomenon corresponds to the overheating of user dynamics, such as online flame wars, and the oscillation model can explain the overheating of user dynamics through the network structure (see Public Reference 3: M. Aida, C. Takano and M. Murata, “Oscillation model for describing network dynamics caused byasymmetric node interaction,” IEICE Transactions on Communications, vol. E101-B, no. 1, pp. 123-136, January 2018).

[0043] User dynamics overheating and low-frequency mode Within the framework of the vibration model, excessive activation of user dynamics due to structural changes in social networks occurs because the eigenvalues ​​of the Laplacian matrix L change with the structural changes, and some eigenvalues ​​change from real to imaginary. Theoretically, this process leads to the activation of low-frequency modes in the behavior of vibrational energy. The mechanism is briefly explained below. The characteristic equation of the Laplacian matrix L is given by the following equation 4.

number

[0044] Figure 5 illustrates the change in the shape of the characteristic polynomial of the Laplacian matrix associated with continuous structural changes in the network. The characteristic polynomial on the left-hand side is an n-th degree polynomial in terms of the variable λ, and its shape is illustrated in Figure 5. In Figure 5, the intersections with the horizontal axis are the solutions to the characteristic equation (Equation 4), i.e., the eigenvalues ​​of the Laplacian matrix L. Figure 5 illustrates a situation where, as the network structure changes continuously, the eigenvalues ​​change continuously, shifting from real eigenvalues ​​to imaginary eigenvalues. It can be seen that the intersections of the characteristic polynomial and the horizontal axis approach each other on the horizontal axis before the intersections disappear and imaginary eigenvalues ​​appear. This means that the two eigenvalues ​​take close values.

[0045] In the oscillation model, the frequency of each oscillation mode included in the solution x(t) is given by the square root of the eigenvalue of the Laplacian matrix. Therefore, the proximity of eigenvalues ​​means that oscillation modes with very close frequencies will appear. As a simple example that illustrates the essence of the mechanism, consider the superposition of trigonometric functions with angular frequencies ω1 and ω2. Using the sum-to-product formulas for trigonometric functions, this can be given by the following equation 5.

number

[0046] Figure 6 is an explanatory diagram of the low-frequency modes that appear in vibrational energy. Here, we will examine how low-frequency beats occur in vibrational energy. For simplicity, let's consider the two trigonometric functions as cos(ωt) with angular frequencies ω = 0.10 and ω = 0.11. The left figures in Figures 6(a) and (b) represent these trigonometric functions, and the right figure represents their frequency spectral distribution. Here, the frequency f on the horizontal axis is f := ω / (2π) × 4096, and the peaks in the frequency spectrum are f = 65.2 and f = 71.7, respectively. Figure 6(c) shows the superposition of the two trigonometric functions. Low-frequency beats appear in the left figure, but the frequency spectral distribution in the right figure is a superposition of the original frequencies, and no low frequencies are observed.

[0047] In the vibration model, vibration energy represents the intensity of user dynamics, and since vibration energy is proportional to the square of the waveform amplitude, Figure 6(d) displays the square of the superimposed waveform. Two peaks appear in this frequency spectrum, one at high frequency and one at low frequency. The frequencies of these peaks are f = 137.0 = 65.2 + 71.7 and f = 6.5 = |65.2 - 71.7|. Figure 6(e) shows the frequency spectrum of the superimposed waveform square obtained by taking a moving average with a window width of 64. This operation represents a situation in actual time-series data observation where observation with fine time resolution is not possible, and the data is averaged over short intervals. As a result, the high-frequency peak in the frequency spectrum disappears, and only the low-frequency peak remains. In other words, it can be seen that when the eigenvalues ​​of the Laplacian matrix are close together, there is a possibility that low-frequency vibration modes will appear in the vibration energy. The presence of low-frequency modes in time series representing the intensity of user activity during overheating of user dynamics has been verified using real-world data (see Public Reference 4: M. Aida, K. Nagatani, C. Takano, “Increase of low-frequency modes of user dynamics in online social networks during overheating of discussions,” Nonlinear Theory and Its Applications, IEICE, vol. 13, no. 2, pp. 511-532, 2022).

[0048] ◎ Verification of low-frequency mode increase characteristics during user dynamics overheating Next, to illustrate the characteristics of the frequency spectrum during user dynamics overheating, we will illustrate a data analysis of Google® Trends related to the present invention, based on the results shown in prior art document 4. ○Time-series data to be analyzed and collection methods As an indicator reflecting the intensity of online user dynamics activity, we consider the time series of searches for a specific keyword per unit of time. In this specification, we obtain search volume data for a specific keyword using Google® Trends. The obtained data is the number of searches per hour and covers up to one week, and the data for that period is normalized so that the maximum value is 100. To obtain long-term time series data from this data, we obtain one week's worth of data with some overlapping periods, and then scale and combine the data by utilizing the fact that the overlapping intervals are the same data to obtain long-term time series data.

[0049] Figure 7 is an explanatory diagram of the spectral analysis method. ○ Methods for analyzing frequency spectra The search volume for keywords that are already popular or well-known is high, and it is thought that this includes information unrelated to the activity of online communities. To remove such unnecessary information from time-series data, the frequency spectrum is analyzed using the following method. First, the time series data is divided into fixed time intervals, and a Fast Fourier Transform (FFT) is performed on the time series data for each time interval. Since the absolute value of the Fourier mode contributes to the vibrational energy representing the intensity of user dynamics, the amplitude spectrum is considered by ignoring the phase component (taking the absolute value). Then, the component with frequency 0 is removed, and the remaining frequency spectrum distribution is normalized so that the whole is 1 (see Figure 7). This eliminates the effects of differences in time series bias (DC component) and constant multipliers, and allows for analysis that eliminates the effect of a simple increase in the number of accesses.

[0050] ○Examples of frequency spectrum distribution analysis Figure 8 is an explanatory diagram illustrating an example of experimental results, showing the time series of searches for "Fukuoka SoftBank Hawks (registered trademark)" and its frequency spectral distribution. The upper part of Figure 8 shows the hourly search volume for the professional baseball team "Fukuoka SoftBank Hawks (registered trademark)". The time series data covers the period from January 1, 2019 to May 21, 2019. From this time series, intervals with high and low search volumes (interval length 1024 hours) were selected, and the amplitude spectrum was displayed using the method described above, as shown in the lower part of Figure 8. However, to make the figure easier to read, a moving average with a window width of 50 is used for display. From these results, it can be seen that low-frequency modes are strongly present during periods with high search volumes (during the regular season).

[0051] Figure 9 is an explanatory diagram illustrating an example of experimental results, showing the time series of searches for "job hunting" and its frequency spectral distribution. The upper part of Figure 9 shows the hourly search volume for "job hunting." The time series data covers the period from May 7, 2019 to August 13, 2019. From this time series, three intervals with different search volumes (interval length 512 hours) were selected, and the amplitude spectra were displayed using the method described above, as shown in the lower part of Figure 9. However, to make the figure easier to read, a moving average with a window width of 10 was used for display. From these results, it can be seen that there is no significant difference in the frequency spectral distribution even when the search volume differs.

[0052] From the above, it can be suggested that the sheer volume of searches itself does not affect the frequency spectrum distribution, but rather that the rise in social interest (including online communities) influences the frequency spectrum (through changes in the structure of social networks).

[0053] ◎ Predictive detection of overheating in user dynamics The aforementioned evaluation compared the frequency spectra of time intervals where user dynamics activity was strong / weak, but since it was not an analysis of continuous time intervals, it could not determine the temporal changes in the frequency spectrum. Below, we obtain time-series data of hourly search volume from Google® Trends for several keywords that have become socially popular, divide it into continuous time intervals, and analyze the frequency spectrum of each time interval. This will explain how to detect user dynamics overheating early through frequency spectrum analysis. Insurance fraud case Figure 10 is an explanatory diagram illustrating an example of experimental results, showing the time-series data of search volume for "Used Car Sales Company B". Figure 11 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 10. Figure 10 shows the number of searches for "Used Car Sales Company B" from November 1, 2022 to November 1, 2023 (normalized to a maximum value of 100), obtained from Google® Trends. The president's apology press conference (July 25) is marked with an "x," indicating a sharp increase in search volume immediately afterward. Prior to that, the time when the scandal involving fraudulent insurance claims was reported (July 6) is also marked with an "x." Figure 11 shows the results of spectral analysis performed on this time-series data, which was divided into 1024-hour intervals. In Figure 11, the black line represents the section where the number of searches increased sharply, and it can be seen that the low-frequency mode is more prominent in this section than in other sections. In addition, an increase in the low-frequency mode can be observed from three sections prior to this point. Therefore, by analyzing the low-frequency component, it is possible to determine the signs of an impending online firestorm.

[0054] Japan Series Figure 12 is an explanatory diagram illustrating an example of experimental results, and is an explanatory diagram of the time-series data of search volume for "Japan Series". Figure 12 shows the number of searches for "Japan Series" from January 1, 2022 to January 1, 2023 (normalized to a maximum value of 100), obtained from Google® Trends. Game 7 (10 / 30) is shown with a black "x," and the time when the opposing teams were decided (10 / 21) is shown with a red "x." Next, the time series data was divided into 1024-hour intervals, similar to Figure 11, and the results of spectral analysis are shown in Figure 13.

[0055] Figure 13 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 12. Figure 13 shows the low-frequency portion of the frequency spectrum on the horizontal axis and the normalized amplitude spectrum distribution obtained by the method described above on the vertical axis. However, for ease of comparison, each distribution has been smoothed by taking a moving average with a window size of 40. The black line represents the section in which the number of searches increased sharply, and it can be seen that the low-frequency mode is more prominent in this section than in other sections. In addition, the low-frequency mode is also more prominent in the sections immediately before and after this section compared to other periods. Although the number of searches did not increase significantly in this section, it was a time when the excitement for the Japan Series was building / the afterglow was still lingering.

[0056] Ukraine Figure 14 is an explanatory diagram illustrating an example of experimental results, showing the time-series data of search volume for "Ukraine". Figure 15 is an explanatory diagram of the amplitude spectral distribution for each section in Figure 14. Figure 14 shows the search volume for "Ukraine" (normalized to a maximum value of 100) from August 7, 2021 to January 1, 2023, obtained from Google® Trends. The point at which Russia launched its military invasion (February 24) is indicated by a red "X," and the search volume surged immediately after this. Prior to this, the point at which President Biden hinted at the possibility of a Russian military invasion (January 29) is indicated by a black "X." Figure 15 shows the results of performing the same frequency spectral analysis on the time-series data in Figure 14 as in Figure 13. The black line indicates a period of sharp increase in search volume, where the low-frequency mode is more prominent than in other periods. In addition, the low-frequency mode is similarly prominent in the period immediately preceding this one. While the search volume didn't increase significantly in this period, it coincided with a time when some informed individuals became more interested in Ukraine following news about President Biden. [Examples]

[0057] In the above-described Example 1, the determination means Ca6 determines that there is an increase in future frequency information, that is, there is a possibility that user dynamics increases (goes viral or buzzes), when the amplitude intensity of a predetermined low-frequency component (for example, the lowest-frequency component) increases with the passage of time (for example, increases continuously in three intervals compared to the previous time interval). Also, in the above-described Example 1, the case where the determination means Ca6 makes a determination using a threshold value was described. In contrast, in the present Example 2, it is intended to detect a sign of an increase in future frequency information (a sign of abnormal online user behavior) using a model based on the Z-score method. Note that the configurations of the prediction system S and the prediction computer 1 in Example 2 are the same as those in Example 1 (FIG. 1 and FIG. 2).

[0058] In the present Example 2, the frequency analysis means Ca5 in FIG. 2 analyzes the frequency components of the time-series data for each time interval, as in Example 1. On the other hand, the determination means Ca6 in the present Example 2 sets the lowest-frequency component (amplitude intensity) at time t as x t and sets the n most recent data (lowest-frequency component (amplitude intensity)) at time t as x t-1 , x t-2 , … x t-n to obtain the average (denoted as X t-1 ) and the variance V t-1 of the n most recent data (lowest-frequency component). Further, the determination means Ca6 obtains the unbiased standard deviation σ t-1 from the variance V t-1 .

[0059] Then, the determination means Ca6 determines that there is a sign of abnormal online user behavior at time t when the lowest-frequency component x t at time t satisfies the following equation (1). Note that c is a value (constant) greater than 0. x t - X t-1 > cσ t-1 …(1) By using the above equation (1), the determination means Ca6 compares the low-frequency component (x t ) at time t with the low-frequency component (X t ) in the time interval prior to time tt-1 It can be said that a comparison is made with the results of this comparison, and the determination of a warning sign is made based on these comparison results.

[0060] The mean and variance of the lowest frequency component (amplitude intensity) are thought to vary greatly depending on the type of keyword being analyzed and the observation time period. In this embodiment 2, a model based on the Z-score method using the mean and variance (unbiased standard deviation obtained from the variance) of such lowest frequency component (amplitude intensity) is used to predict abnormal online user behavior, thus enabling accurate prediction.

[0061] Note that in equation (1) above, the mean X t-1 or unbiased standard deviation σ t-1 We decided to use this method, but it is not limited to this; we also use a variance V that prioritizes the most recent data and weights it exponentially. t 'Also introduced, mean X by exponential smoothing t-1 'and standard deviation σ t-1 You may also use '. Average X by exponential smoothing t-1 ' is represented by the following equation (2). X t-1 '=αx t +(1-α)X t-2 …(2) Note X t-2 ' is x t-2 , x t-3 ,...x t-n-1 This is the average of . Also, α (0 < α < 1) is the average X for n. t-1 The data used to calculate ' is selected so that the weighted average value of the data over time matches.

[0062] Furthermore, the standard deviation σ obtained by exponential smoothing t-1 ' is, variance V t-1 Using ', it is expressed by the following equation (3), and the variance V t ' is represented by the following equation (4). σ t-1 '=(V t-1 ') 1 / 2 …(3) V t '=α(x t -X t ')+(1-α)Vt-1 …(4) However, assume that X0' = αx0 and V0' = α(x0 - X0').

[0063] In this case, the determination means Ca6 determines the lowest frequency component x at time t. t If the following equation (5) is satisfied, it is determined that there was a precursor to abnormal online user behavior at time t. x t -X t-1 '>cσ t-1 …(5)

[0064] Used car sales company B The following describes Experiment (1) of Example 2. In Experiment (1), the keyword was "Used Car Sales Company B" (which is actually the official company name), and abnormal online user behavior was predicted using equations (1) or (5) around the time of media coverage of a recent incident (insurance fraud case). For the prediction, n=10000 and c=5 were used for n and c. As a result, when equation (1) was used, the time of prediction was 4:00 AM on April 3, 2023, and when equation (5) was used, the time of prediction was 3:00 AM on April 3, 2023. It should be noted that media coverage of Company B and the public disclosure of the incident occurred on July 6, 2023. Therefore, it was confirmed that when predictive detection using a Z-score model was performed for the keyword "Used Car Sales Company B", abnormal online user behavior could be detected before media coverage, regardless of whether equation (1) or (5) was used. Figure 16(a) is a graph showing the trend of the lowest frequency component related to the keyword "Used Car Sales Company B" (solid line), the time when the precursor was detected using equation (5) (dotted line), and the date and time when the media report actually occurred (marked with an "x"). From Figure 16(a), it can be seen that by using equation (5) above, it is possible to detect the increase in the lowest frequency component that occurred before the media report.

[0065] 〇TV station F company Next, we will describe the experiment (part 2) related to Example 2. In this experiment (part 2), the keyword was "TV station F" (which is actually the official company name), and we performed prediction of abnormal online user behavior using equations (1) or (5) above, around the time of media coverage of a recent incident (a scandal within the TV station) that had attracted attention. For the prediction, n and c were set to n=10000 and c=5. As a result, when equation (1) was used, the time of prediction was 12:00 PM on November 2, 2024, and when equation (5) was used, the time of prediction was 3:00 PM on November 2, 2024. The actual media coverage of F company and the public disclosure of the incident occurred on December 19, 2024. Therefore, it was confirmed that, even with the keyword "TV station F," by performing prediction using a model based on the Z-score method, it was possible to detect signs of abnormal online user behavior before media coverage, regardless of whether equation (1) or (5) was used. Figure 16(b) is a graph showing the trend of the lowest frequency component related to the keyword "TV station F" (solid line), the time when the precursor was detected using equation (5) (dotted line), and the date and time when the media coverage actually occurred (marked with an "x"). From Figure 16(b), it can be seen that by using equation (5) above, it is possible to detect the increase in the lowest frequency component that occurred before the media coverage. [Examples]

[0066] Next, Example 3 will be described. As shown in Figure 17(a), the stock trading device T according to Example 3 is connected to the prediction system S described in Example 1 or Example 2, and is a device that buys and sells stocks based on the predictions of abnormal online user behavior predicted by the prediction system S.

[0067] The stock trading device T is an information processing device such as a PC, and has a CPU, ROM, RAM, storage, etc. The stock trading device T functions as a frequency information acquisition means 30 and a short selling means 32, as shown in Figure 17(b), when the CPU executes a program. Figure 17(b) also shows a stock information DB 40 stored in the storage, etc. The stock information DB 40 is assumed to store information on stocks held by the user (company name, number of shares, repayment deadline, etc.).

[0068] The frequency information acquisition means 30 extracts the company names of the stocks held by the user from the stock information DB 40 and sends them to the prediction system S as keywords. When the prediction system S detects an indication that the frequency information of the keyword (company name) is about to increase, the prediction system S sends information about the indication to the frequency information acquisition means 30. The frequency information acquisition means 30 receives this information and passes it on to the short selling means 32. Generally, even if an indication that the frequency information of a company name is about to increase is detected, it is unclear what impact the social phenomenon suggested by that indication will have on the company's stock price. However, since the frequency information of a company name usually increases when the company is experiencing a backlash on social media, etc., in this embodiment 3, it is assumed that an increase in the frequency information of a company name will have a downward effect on the company's stock price.

[0069] The short-selling tool 32 sells the company's shares when a warning sign is detected. The short-selling tool 32 stores the stock price at the time of sale in the stock information DB 40.

[0070] Furthermore, the short-selling method 32 monitors the stock price after the sale using the internet or other means, and compares it with the stock price at the time of sale stored in the stock information DB 40. When the price exceeds a certain profit margin, the short-selling method 32 buys back the sold shares. Since the repayment period for the shares is six months, even if the profit margin does not exceed a certain level, the short-selling method 32 forcibly buys back the sold shares just before the repayment period, as stored in the stock information DB 40.

[0071] According to the stock trading device of this embodiment 3, by performing the processing described above, it becomes possible to trade stocks with a high profit margin.

[0072] In this embodiment 3, it was assumed that an increase in the frequency information of a company name would have the effect of lowering the company's stock price. However, the embodiment is not limited to this, and it is also possible to estimate whether an increase in the frequency information of a company name would have the effect of lowering or raising the company's stock price based on the sub-keywords (second keyword) analyzed by the analysis means Ca8 described in embodiment 1.

[0073] 〇 Short selling experiment of Meta's stock The experiment for Example 3 is described below. In this experiment, time-series data of web search volume for the keyword "Meta" was obtained using Google Trends, and a precursory indicator was determined using equation (5) from Example 2. Here, c=5 was used as c for the precursory indicator. Furthermore, index smoothing was used to estimate the mean and variance of past data, with a weighted mean of 1080 hours. In addition, when a precursory indicator was detected, only one share of stock was sold, and the stock was to be bought back when the profit margin reached 10% or 6 months had passed since the sale. If the stock price fell after the stock was sold and a buyback was made, no further sales were made even if a precursory indicator was detected for a certain period after the buyback (14 days in this example). Figure 18 shows the results of a simulation conducted on Meta's stock price from January 1, 2021 to January 1, 2025. During this period, Meta's stock price was in the range of approximately $250 to $700. Here, N is the window size (input data size) of the short-time Fourier transform, and the unit is time. The experimental results showed that in the cases of N=512 and N=1024, the profit per short sale was approximately $80, successfully generating a profit. Although the total profit was higher for N=512 than for N=1024, N=1024 required fewer trials (48 trials), allowing for operation with less capital.

[0074] In the above embodiment 3, the short selling means 32 buys back the shares when the profit margin reaches a certain value (a uniform value regardless of the timing), but it is not limited to this. For example, the short selling means 32 may buy back the shares when the profit margin exceeds a threshold that changes over time. In this case, the threshold immediately after the sale can be set to a large value, and the threshold can be set to a smaller value as time passes. Experimental results have shown that this can result in a higher profit margin compared to when the threshold is set to a constant value. The threshold may change linearly, curvilinearly, or in stages over time. Whether the threshold is set to a constant value or a variable value may be set by the user as appropriate, or it may be predetermined.

[0075] (Example of change) Although embodiments of the present invention have been described in detail above, the present invention is not limited to the embodiments described above, and various modifications can be made within the scope of the gist of the present invention as described in the claims. An example of a modification of the present invention (H01) is shown below. (H01) In the above embodiment, an example was given in which centralized processing is performed on the prediction computer 1, but the embodiment is not limited thereto. It is also applicable to an embodiment in which distributed processing is performed on multiple computer devices connected via the Internet or an intranet. [Explanation of symbols]

[0076] 1... Computer, 3,4... Servers, 30...Means for acquiring frequency information, 32...Short selling methods, 40…Stock information database, Ca2...Data acquisition method, Ca4...Data segmentation means, Ca5... Frequency analysis means, Ca6...judgment means, Ca8…analytical means, P1... Prediction program, S... Prediction system T... Stock trading device.

Claims

1. A prediction system that predicts the future level of interest in a keyword based on time-series data of frequency information related to the keyword sent to the server, A data acquisition means for acquiring the aforementioned time-series data, A data segmentation means for dividing the aforementioned time series data into predetermined time intervals, A frequency analysis means for analyzing the frequency components of the time series data for each of the aforementioned time intervals, A determination means for determining whether or not there is an increase in the frequency information in the future beyond the target time period, based on a comparison between the amplitude intensity of a predetermined low-frequency component in a target time period and the amplitude intensity of the low-frequency component in a time period prior to the target time period, A prediction system equipped with the following features.

2. The determination means determines that if the amplitude intensity of the low-frequency component in each time interval increases with the passage of time, then the frequency information for the future beyond the target time interval increases. The prediction system according to claim 1.

3. The frequency analysis means analyzes the frequency components by performing a Fast Fourier Transform on the time series data. The prediction system according to claim 1.

4. The data segmentation means is configured to change the length of the time interval according to settings based on user input. The prediction system according to claim 1.

5. The system further comprises an analysis means that analyzes a second keyword transmitted to the server along with the aforementioned keyword to determine whether the increase in interest in the aforementioned keyword is positive or negative. The prediction system according to claim 1.

6. The prediction system according to claim 1, wherein the determination means determines that the frequency information in the future beyond the target time interval will increase if the difference between the amplitude intensity of a predetermined low-frequency component in the target time interval and the average value of the amplitude intensity of the low-frequency component in a time interval prior to the target time interval is greater than a threshold determined based on the variance of the amplitude intensity of the low-frequency component in a time interval prior to the target time interval.

7. The prediction system according to claim 6, wherein the mean is an exponentially smoothed mean and the variance is an exponentially smoothed variance.

8. A predictive program that predicts the future level of interest in a keyword based on time-series data of frequency information related to the keyword sent to the server, Computers, Data acquisition means for acquiring the aforementioned time-series data, Data segmentation means for dividing the aforementioned time series data into predetermined time intervals, Frequency analysis means for analyzing the frequency components of the time series data for each of the aforementioned time intervals, A determination means for determining whether or not there is an increase in the frequency information in the future beyond the target time period, based on a comparison between the amplitude intensity of a predetermined low-frequency component in a target time period and the amplitude intensity of the low-frequency component in a time period prior to the target time period. A prediction program designed to function as such.

9. The time-series data of frequency information related to the keywords sent to the server is divided into predetermined time intervals. The frequency components of the time series data are analyzed for each of the aforementioned time intervals. Based on a comparison of the amplitude intensity of a predetermined low-frequency component in the target time interval with the amplitude intensity of the low-frequency component in a time interval prior to the target time interval, it is determined whether or not there is an increase in the frequency information in the future beyond the target time interval. A prediction method for forecasting the future level of interest in a given keyword.

10. A frequency information acquisition means for acquiring whether or not there will be an increase in future frequency information of a company name, as predicted by the prediction system according to any one of claims 1 to 7, A short-selling means that, based on whether or not the frequency information of a company name acquired by the frequency information acquisition means will increase in the future, sells shares of a company corresponding to the company name, and buys back the shares based on the share price or the repayment deadline of the shares, A stock trading device equipped with the following features.

11. Computers Using the prediction method described in claim 9, predict whether or not there will be an increase in future frequency information of company names. Based on whether or not the frequency information of the aforementioned company name will increase in the future, the shares of the company corresponding to the aforementioned company name will be sold, and the shares will be bought back based on the share price of the aforementioned shares or the repayment deadline of the aforementioned shares. Stock trading methods.

Images