Estimation system, estimation method, and program

Abstract

Improve the accuracy of estimation of future estimation targets. [Solution] An estimation target information acquisition unit (201) of an estimation system (1) acquires estimation target information related to an estimation target for which estimation in the future is performed. A macro environmental variable acquisition unit (202) acquires macro environmental variables related to a macro environment. An intermediate result information acquisition unit (203) acquires intermediate result information related to estimation results estimated by each of a plurality of intermediate estimation models specialized for a specific macro environment, which are intermediate estimation models that perform intermediate estimation of an estimation target in the future, based on the estimation target information. A weight acquisition unit (204) acquires weights estimated by a weight estimation model that estimates the weights of each of the plurality of intermediate estimation models, based on the macro environmental variables. A final result information acquisition unit (205) acquires final result information related to a final estimation result of the estimation target in the future, based on the intermediate result information of each of the plurality of intermediate estimation models and the weights of each of the plurality of intermediate estimation models.

Description

[Technical field]

[0001] The present disclosure relates to an estimation system, an estimation method, and a program. [Background technology]

[0002] Conventionally, in various fields such as marketing or finance, there are known methods for estimating an estimation target such as a management index or a stock price based on a macro environment. For example, Patent Document 1 describes a technology that estimates the future price of an investment product based on the price of the investment product at a certain point in time, the fluctuation of the price of the investment product in the past, and events that may affect the price of the investment product, and presents a user with questions about the macro environment related to the price of the investment product in the form of a quiz. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2023-157901 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, with the technology of Patent Document 1, the user can obtain knowledge of the macro environment related to the price of the investment product, but cannot estimate the future price of the investment product based on the macro environment. For this reason, the technology of Patent Document 1 does not provide sufficient accuracy in estimating the future price of the investment product. This point is not limited to the estimation of the price of the investment product as in Patent Document 1. The same applies to the case where an estimation target other than the price of the investment product (for example, an estimation of the future management index of a certain service) is estimated.

[0005] One of the objectives of the present disclosure is to improve the estimation accuracy of future estimation targets. [Means for solving the problem]

[0006] The estimation system of the present disclosure includes an estimation target information acquisition unit that acquires estimation target information regarding an estimation target for which estimation is to be performed in the future, a macro environmental variable acquisition unit that acquires macro environmental variables related to a macro environment, an intermediate estimation model that performs an intermediate estimation of the estimation target in the future based on the estimation target information, and an intermediate result information acquisition unit that acquires intermediate result information regarding estimation results estimated by each of a plurality of intermediate estimation models specialized for a specific macro environment, a weight acquisition unit that acquires weights estimated by a weight estimation model that estimates weights for each of the plurality of intermediate estimation models based on the macro environmental variables, and a final result information acquisition unit that acquires final result information regarding a final estimation result of the estimation target in the future based on the intermediate result information of each of the plurality of intermediate estimation models and the weights of each of the plurality of intermediate estimation models. Effect of the Invention

[0007] The present disclosure can improve the estimation accuracy of future estimation targets. [Brief description of the drawings]

[0008] [Figure 1] FIG. 2 is a diagram illustrating an example of a hardware configuration of the estimation system. [Diagram 2] FIG. 2 is a diagram illustrating an example of functions realized by the estimation system. [Diagram 3] FIG. 13 is a diagram illustrating an example of an intermediate estimation model and a weight estimation model. [Figure 4] FIG. 2 is a diagram illustrating an example of a training database. [Diagram 5] FIG. 11 is a diagram illustrating an example of an estimation target information database. [Figure 6] FIG. 11 is a diagram illustrating an example of a macro environmental variable database. [Figure 7] FIG. 2 is a diagram illustrating an example of a process executed by the estimation system. [Figure 8] FIG. 13 is a diagram showing an example of an estimation result screen. [Figure 9] FIG. 13 is a diagram illustrating an example of a function realized in a modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] [1. Hardware configuration of the estimation system] An example of an embodiment of an estimation system, an estimation method, and a program according to the present disclosure will be described. FIG. 1 is a diagram showing an example of a hardware configuration of an estimation system. For example, the estimation system 1 includes a learning terminal 10, a server 20, and a user terminal 30. Each of the learning terminal 10, the server 20, and the user terminal 30 is connected to a network N such as the Internet, a LAN, or a public telephone line. Although FIG. 1 shows one each of the learning terminal 10, the server 20, and the user terminal 30, there may be a plurality of at least one of the learning terminal 10, the server 20, and the user terminal 30.

[0010] The learning terminal 10 is a computer that performs the learning described below. For example, the learning terminal 10 is a personal computer, a server computer, or a tablet. The learning terminal 10 includes a control unit 11, a memory unit 12, a communication unit 13, an operation unit 14, and a display unit 15. The control unit 11 includes at least one processor. The memory unit 12 includes at least one of a volatile memory such as a RAM and a non-volatile memory such as a flash memory. The communication unit 13 includes at least one of a communication interface for wired communication and a communication interface for wireless communication. The operation unit 14 is an input device such as a touch panel or a mouse. The display unit 15 is a display such as a liquid crystal or organic EL.

[0011] The server 20 is a server computer. For example, the server 20 includes a control unit 21, a storage unit 22, and a communication unit 23. The hardware configurations of the control unit 21, the storage unit 22, and the communication unit 23 may be similar to those of the control unit 11, the storage unit 12, and the communication unit 13, respectively.

[0012] The user terminal 30 is a user's computer. For example, the user terminal 30 is a personal computer, a server computer, a tablet, a smartphone, or a wearable terminal. For example, the user terminal 30 includes a control unit 31, a storage unit 32, a communication unit 33, an operation unit 34, and a display unit 35. The hardware configurations of the control unit 31, the storage unit 32, the communication unit 33, the operation unit 34, and the display unit 35 may be similar to those of the control unit 11, the storage unit 12, the communication unit 13, the operation unit 14, and the display unit 15, respectively.

[0013] The programs stored in the storage units 12, 22, 32 may be supplied to the learning terminal 10, the server 20, or the user terminal 30 via the network N. At least one of a reading unit (e.g., a memory card slot) that reads a computer-readable information storage medium and an input / output unit (e.g., a USB port) for inputting and outputting data to and from an external device may be included in the learning terminal 10, the server 20, or the user terminal 30. For example, a program stored in an information storage medium may be supplied to the learning terminal 10, the server 20, or the user terminal 30 via at least one of the reading unit and the input / output unit.

[0014] Furthermore, the estimation system 1 may include at least one computer. The computers included in the estimation system 1 are not limited to the example of FIG. 1. For example, the estimation system 1 may include only the learning terminal 10 and the server 20. In this case, the user terminal 30 exists outside the estimation system 1. The estimation system 1 may include only the server 20. In this case, the learning terminal 10 and the user terminal 30 exist outside the estimation system 1. For example, the estimation system 1 may include the server 20 and another computer not shown in FIG. 1.

[0015] [2. Overview of the estimation system] In this embodiment, the estimation system 1 performs estimation regarding the estimation target based on a learning model, which is a machine learning model. Various known methods can be used as the machine learning method. The machine learning method may be any of supervised learning, semi-supervised learning, and unsupervised learning. There are various theories about the meaning of the term machine learning, but the machine learning of this embodiment includes various known meanings. For example, the machine learning of this embodiment includes AI (Artificial Intelligence) and deep learning. The learning model is data including a program and parameters created by the machine learning method. An example of the learning model will be described later.

[0016] The estimation target is a target for which estimation is performed by the learning model. In other words, the estimation target is information output by the learning model. The estimation target may be any target. In this embodiment, an example is given in which the estimation target is a management indicator in an organization. Therefore, the section explaining the management indicator can be read as the estimation target. The management indicator is an indicator indicating the management status. For example, the management indicator may be GMS (Gross Merchandise Sales), GMV (Gross Merchandise Value), a KPI (Key Performance Indicator) not classified as GMS or GMV, or another indicator.

[0017] The organization indicated by the management indicators may be any organization such as a company. The estimation target may be a management indicator of a specific service in an organization that handles multiple services. The estimation target may be an object other than the management indicators. The estimation target may be an object that may be affected by the macro environment described below. For example, the estimation target may be the stock price of a corporation, the price of an investment trust, the price of gold or platinum, the fair price of a product or service, customer needs, or other objects. In this embodiment, the description of management indicators can be replaced with these other estimation targets.

[0018] For example, the estimation system 1 estimates the estimation target based on macro environmental variables related to the macro environment. The macro environment is the external environment surrounding the estimation target. In this embodiment, since the management indicators of the organization correspond to the estimation target, the macro environment is the environment surrounding the organization. For example, the macro environment is domestic or foreign economic, political, and social factors, technological progress, weather, environment, legal changes, or other environments. The macro environment is basically an environment that cannot be controlled by the organization. The opposite environment to the macro environment is the micro environment. The micro environment is basically an environment that can be controlled by the organization. For example, the micro environment is the organization's customers, business partners, employees, partner companies, competitors, or other environments.

[0019] A macroenvironmental variable is a variable related to the macroenvironment. The macroenvironmental variable is expressed by a numerical value, a character, a symbol, or a combination thereof. In this embodiment, the macroenvironmental variable is a variable related to the macroenvironment surrounding the organization. For example, in the case of an economic macroenvironment, variables indicating minimum wage, average annual income, salary increase, price increase, or other economic factors correspond to macroenvironmental variables. In the case of a political macroenvironment, variables indicating cabinet approval rating, voter turnout, policy content, political stability, international situation, or other political factors correspond to macroenvironmental variables. In the case where the macroenvironment indicates factors other than the economy and politics, the macroenvironmental variable may be a variable indicating the other factors. There are also microenvironmental variables in the microenvironment. The macroenvironment and the microenvironment may have the well-known meanings used in fields such as economics or marketing.

[0020] For example, macro environmental variables are variables that do not directly affect the object of estimation, whereas micro environmental variables are variables that directly affect the object of estimation. For this reason, if the estimation system 1 estimates the object of estimation by treating macro environmental variables and micro environmental variables in the same way, the estimation accuracy of the object of estimation cannot be sufficiently improved. Furthermore, rather than directly contributing to the object of estimation, such as a management indicator, macro environmental variables may be factors that change the mechanism of the relationship between micro variables and the object of estimation. In this respect, too, if macro environmental variables and micro environmental variables are treated in the same way, the estimation accuracy of the object of estimation cannot be sufficiently improved.

[0021] Therefore, in this embodiment, instead of treating the macro environmental variables and the micro environmental variables in the same way, a learning model specialized for the macro environment of a specific macro environmental variable is prepared. The estimation system 1 estimates the estimation target based on the learning model specialized for the macro environment of a specific macro environmental variable. This allows the estimation system 1 to perform estimation according to the macro environmental variables, thereby improving the estimation accuracy of the estimation target. Details of the estimation system 1 will be described below.

[0022] [3. Functions realized by the estimation system] FIG. 2 is a diagram illustrating an example of functions realized by the estimation system 1. As shown in FIG.

[0023] [3-1. Functions realized on the learning device] For example, the learning terminal 10 includes a data storage unit 100 and a learning unit 101. The data storage unit 100 is realized by the storage unit 12. The learning unit 101 is realized by the control unit 11.

[0024] [Data storage section] The data storage unit 100 stores data necessary for learning the learning model. In this embodiment, a plurality of intermediate estimation models IM1, IM2 and a weight estimation model WM are prepared as the learning model. Hereinafter, when the intermediate estimation model IM1 and the intermediate estimation model IM2 are not distinguished from each other, the numbers at the end of the symbols are omitted and the model is referred to as the intermediate estimation model IM. In addition, the whole including the plurality of intermediate estimation models IM and the weight estimation model WM is referred to as the overall model. When the plurality of intermediate estimation models IM and the weight estimation model WM are not distinguished from each other, they are referred to as the learning model.

[0025] For example, the data storage unit 100 stores an intermediate estimation model IM before learning and a weight estimation model WM before learning. When learning is performed by the learning unit 101 described later, the data storage unit 100 stores the trained intermediate estimation model IM and the trained weight estimation model WM. Each of the intermediate estimation model IM and the weight estimation model WM includes a program indicating a process such as calculation of an embedded representation, and a parameter adjusted by learning. These programs and parameters may be programs and parameters used in known machine learning methods. For example, the program may include an input layer that accepts input, an intermediate layer that performs intermediate processing, and an output layer that performs output. For example, the parameters may include weights and biases.

[0026] FIG. 3 is a diagram showing an example of an intermediate estimation model IM and a weight estimation model WM. As shown in FIG. 3, in this embodiment, an example is given in which the number of intermediate estimation models IM is two. The number of intermediate estimation models IM may be three or more. The number of intermediate estimation models IM may be any number. The overall model may include only the intermediate estimation model IM and the weight estimation model WM, but in this embodiment, the overall model also includes a program that performs calculations based on the output of the intermediate estimation model IM and the output of the weight estimation model WM. The program corresponds to the output layer of the overall model. For example, the program includes program codes for the multiplication and addition processes in FIG. 3.

[0027] The intermediate estimation model IM is a model that performs intermediate estimation of a future estimation target. The intermediate estimation model IM can also be said to be a model that estimates an intermediate result of an estimation target. Any machine learning method may be adopted for the intermediate estimation model IM. In this embodiment, the intermediate estimation model IM is exemplified as a neural network in the Mixture of Experts method. The Mixture of Experts method is a machine learning method that combines estimation results by each of a plurality of experts. The intermediate estimation model IM is an expert in the Mixture of Experts method. An expert is sometimes called an expert function. An expert is a learning model specialized for a specific task.

[0028] In this embodiment, the intermediate estimation model IM, which is an expert, is a model specialized for a macro environment of a certain macro environmental variable. For example, when the macro environmental variable indicates the minimum wage and the macro environment indicates an economic climate corresponding to the minimum wage, the first expert is a model specialized for a macro environment of a favorable economic climate, and the second expert is a model specialized for a macro environment of a poor economic climate. When the macro environmental variable indicates the temperature in winter and the macro environment indicates consumption corresponding to the temperature in winter, the first expert is a model specialized for a macro environmental disaster of a warm winter, and the second expert is a model specialized for a macro environmental disaster of a cold winter.

[0029] The intermediate estimation model IM may be a learning model of any machine learning method. For example, when a method other than a neural network is used in the Mixture of Experts, the intermediate estimation model IM may be a learning model of the other method. The intermediate estimation model IM may be a learning model of a method other than the Mixture of Experts method. For example, the intermediate estimation model IM may be a support vector machine, a linear regression model, a decision tree model, a large-scale language model, or another learning model. The machine learning method adopted as the intermediate estimation model IM is not limited to the example of this embodiment.

[0030] In this embodiment, estimation target information regarding the estimation target, which is an example of a management index, is input to the intermediate estimation model IM. For example, the estimation target information indicates the estimation target at least in the present and / or past. The estimation target in the future is estimated from the estimation target at least in the present and / or past. The estimation target information can also be said to be the result or performance of the estimation target at least in the present and / or past. For example, when the GMS in a certain organization corresponds to the estimation target, the estimation target information indicates the GMS in the organization at least in the present and / or past.

[0031] For example, the estimation target information may indicate a temporal change regarding the estimation target. The temporal change regarding the estimation target is a time-series change of the estimation target. The temporal change regarding the estimation target may be a change in any period. For example, the temporal change regarding the estimation target may be a change of the estimation target in the last year, a change of the estimation target in the last 10 years, or a change of the estimation target in another period. The temporal change regarding the estimation target may be a change in a period from a point in the past to another point in the past, rather than the most recent period.

[0032] For example, the estimation target information may indicate a spatiotemporal change regarding the estimation target. The spatiotemporal change regarding the estimation target is a temporal and spatial change of the estimation target. The space may also be referred to as a location. The spatial change of the estimation target may be a change in any space. For example, the spatial change of the estimation target may be a country, a region, a prefecture, a city, a town, a village, or other space. For example, if the estimation target is the GMS of a card company that issues credit cards, the spatiotemporal change regarding the estimation target may be a change in the GMS for each month of each prefecture. The combination of time and space in the spatiotemporal change regarding the estimation target may be any combination.

[0033] The output from the intermediate estimation model IM is intermediate result information on an intermediate estimation result according to the estimation target information input to the intermediate estimation model IM. The intermediate result information is used for calculating the final estimation result. In other words, the intermediate result information is substituted into a calculation formula for calculating the final estimation result. For example, if the estimation target is the GMS of a certain organization, the intermediate result information indicates an intermediate estimation result of the GMS of the organization in the future. Since the parameters of one intermediate estimation model IM and another intermediate estimation model IM are different from each other, different intermediate result information is output even if the same estimation target information is input.

[0034] For example, suppose that intermediate estimation model IM1 is a learning model specialized for a macro environment of economic boom. Furthermore, suppose that intermediate estimation model IM2 is a learning model specialized for a macro environment of economic recession. In this case, intermediate estimation model IM1 has learned trends in economic boom. The intermediate estimation model IM1 outputs intermediate result information indicating the estimation result in economic boom based on the estimation target information. The intermediate estimation model IM2 has learned trends in economic recession. The intermediate estimation model IM2 outputs intermediate result information indicating the estimation result in economic recession based on the estimation target information.

[0035] For example, suppose that the intermediate estimation model IM1 is a learning model specialized for a warm winter macro environment. Furthermore, suppose that the intermediate estimation model IM2 is a learning model specialized for a cold winter macro environment. In this case, the intermediate estimation model IM1 has learned trends in warm winters. The intermediate estimation model IM1 outputs intermediate result information indicating the estimation results in warm winters based on the estimation target information. The intermediate estimation model IM2 has learned trends in cold winters. The intermediate estimation model IM2 outputs intermediate result information indicating the estimation results in cold winters based on the estimation target information.

[0036] The weight estimation model WM is a model that estimates the weights of each of the multiple intermediate estimation models IM. The weights can also be called coefficients or importance. The weights are used in the calculation of the final estimation result. In the example of FIG. 3, the weight estimation model WM outputs a weight w1 of the intermediate estimation model IM1 and a weight w2 of the intermediate estimation model IM2. Hereinafter, when the weights w1 and w2 are not distinguished from each other, the weight w is written with the numeral at the end of the code omitted. In this embodiment, the weight estimation model WM is exemplified as a neural network in the Mixture of Experts method. The weight estimation model WM corresponds to a gate in the Mixture of Experts method. The gate is sometimes called a gate function. The gate estimates the weights of the experts. For example, an activation function such as a softmax function is used as the gate.

[0037] The weight estimation model WM may be a learning model of any machine learning method. For example, when a method other than a neural network is used in the Mixture of Experts, the weight estimation model WM may be a learning model of another method. The weight estimation model WM may be a learning model of a method other than the Mixture of Experts method. For example, the weight estimation model WM may be a support vector machine, a linear regression model, a decision tree model, a large-scale language model, or another learning model. The machine learning method adopted as the weight estimation model WM is not limited to the example of this embodiment.

[0038] In this embodiment, a macro environmental variable is input to the weight estimation model WM. A plurality of macro environmental variables may be input to the weight estimation model WM. For example, if the macro environment is the world economy, the macro environmental variable indicates the minimum wage, which is one indicator of the economy. At least one of the current and past macro environmental variables are input to the weight estimation model WM. The weight w of the intermediate estimation model IM is estimated from at least one of the current and past macro environmental variables. The macro environmental variables can also be said to be variables that indicate at least one of the current and past macro environment.

[0039] For example, the weight estimation model WM may be input with a temporal change in the macro environmental variables. The temporal change in the macro environmental variables is a time-series change in the macro environmental variables. The temporal change in the macro environmental variables may be a change in any period. For example, the temporal change in the macro environmental variables may be a change in the macro environmental variables in the last year, the last 10 years, or a change in the macro environmental variables in another period. The temporal change in the macro environmental variables may be a change in a period from a certain point in the past to another point in the past, not just in the recent past. In this embodiment, a case is exemplified in which the time in the spatiotemporal change in the macro environmental variables and the time in the spatiotemporal change related to the estimation target are the same, but they may be different from each other.

[0040] For example, the weight estimation model WM may be input with spatiotemporal changes in the macro environmental variables. The spatiotemporal changes in the macro environmental variables are temporal and spatial changes in the macro environmental variables. The meaning of space is as described for the spatiotemporal changes in the estimation target. In this embodiment, the time and space in the spatiotemporal changes in the macro environmental variables and the time and space in the spatiotemporal changes in the estimation target are the same as each other, but they may be different from each other. For example, if the macro environmental variables are the minimum wage in Japan, the spatiotemporal changes in the macro environmental variables indicate the changes in the minimum wage for each month in each prefecture. The combination of time and space in the spatiotemporal changes in the macro environmental variables may be arbitrary.

[0041] For example, the overall model calculates a final estimation result by adding together values ​​obtained by multiplying the output of each of the intermediate estimation models IM by the weight of the intermediate estimation model IM. In the example of Fig. 3, the overall model outputs final result information, which is a final estimation result, by adding together a value obtained by multiplying the output of the intermediate estimation model IM1 by the weight w1 of the intermediate estimation model IM1 and a value obtained by multiplying the output of the intermediate estimation model IM2 by the weight w2 of the intermediate estimation model IM2. The data storage unit 100 stores a training database DB1 necessary for learning the overall model including the multiple intermediate estimation models IM and the weight estimation model WM.

[0042] FIG. 4 is a diagram showing an example of the training database DB1. For example, the training database DB1 stores a plurality of training data to be learned by the overall model. The training data includes an input portion input to the overall model during learning, and an output portion that becomes a correct answer during learning. The input portion of the training data has basically the same format as the input data input to the overall model during estimation. The output portion of the training data has basically the same format as the output data output from the overall model during estimation. Note that the input portion of the training data may have a slightly different format from the input data input to the overall model during estimation. Similarly, the output portion of the training data may have a slightly different format from the output data output from the overall model during estimation.

[0043] In this embodiment, the input portion of the training data includes training macro environmental variables and training estimation target information. For example, the input portion of the training data may include temporal changes in the training macro environmental variables and estimation target information indicating temporal changes related to the training estimation target. The input portion of the training data may include spatiotemporal changes in the training macro environmental variables and estimation target information indicating spatiotemporal changes related to the training estimation target. The output portion of the training data is training final result information. In this embodiment, since the estimation target is a management indicator of a certain organization, the training final result information indicates the management indicator that is the correct answer during learning.

[0044] The training data may be created by the administrator of the estimation system 1, or may be created by a known tool. The training data is not limited to the example of FIG. 4. In this embodiment, an example is given of a case where the learning of the overall model is performed all at once, but as in a modified example described later, when the learning of each of the intermediate estimation model IM and the weight estimation model WM is performed separately, the training database DB1 may store training data of the intermediate estimation model IM and training data of the weight estimation model WM. The training data may be actual management indicators and macro environmental variables, or may be fictitious management indicators and macro environmental variables prepared by the administrator of the estimation system 1.

[0045] [Learning Department] The learning unit 101 learns the learning model. In this embodiment, the learning unit 101 learns the overall model based on the training data stored in the training database DB1. The learning unit 101 learns the overall model such that an output portion of the training data is output when an input portion of the training data is input. The learning unit 101 learns the overall model by adjusting parameters of the overall model based on the training data. This allows the learning unit 101 to adjust the parameters of each of the intermediate estimation model IM and the weight estimation model WM included in the overall model. When the learning is completed, the learning unit 101 records the learned overall model in the data storage unit 100. The learning unit 101 uploads the learned overall model to the server 20.

[0046] The algorithm for learning the overall model may be a known algorithm used in the field of machine learning. In this embodiment, since each of the intermediate estimation model IM and the weight estimation model WM is a learning model in the Mixture of Experts method, the learning unit 101 may perform learning based on the algorithm in the Mixture of Experts method. For example, the learning unit 101 may cause the overall model to learn training data based on a deep learning algorithm. The loss function used during learning may also be a known loss function. The learning unit 101 calculates a loss, which is an error between the output part of the training data and the output from the overall model during learning, based on the loss function. The learning unit 101 completes learning when the loss becomes small to a certain extent. The learning unit 101 may cause the overall model to learn training data based on other algorithms such as gradient descent or backpropagation.

[0047] [3-2. Functions realized by the server] For example, the server 20 includes a data storage unit 200, an estimation target information acquisition unit 201, a macro environmental variable acquisition unit 202, an intermediate result information acquisition unit 203, a weight acquisition unit 204, and a final result information acquisition unit 205. The data storage unit 200 is realized by the storage unit 22. Each of the estimation target information acquisition unit 201, the macro environmental variable acquisition unit 202, the intermediate result information acquisition unit 203, the weight acquisition unit 204, and the final result information acquisition unit 205 is realized by the control unit 21.

[0048] [Data storage section] The data storage unit 100 stores data necessary for estimating an estimation target. For example, the data storage unit 100 stores a trained overall model. That is, the data storage unit 100 stores a plurality of trained intermediate estimation models IM and a trained weight estimation model WM. In this embodiment, the data storage unit 100 also stores information to be input to the overall model. That is, the data storage unit 100 stores information to be input to each of the plurality of intermediate estimation models IM and information to be input to the weight estimation model WM. For example, the data storage unit 100 stores an estimation target information database DB2 and a macro environmental variable database DB3.

[0049] FIG. 5 is a diagram showing an example of an estimation target information database DB2. The estimation target information database DB2 is a database in which estimation target information input to the intermediate estimation model IM when estimating an estimation target is stored. The estimation target information stored in the estimation target database is estimation target information for estimation. For this reason, the estimation target information for estimation has the same data format as the estimation target information for training, but the specific values ​​are different. The estimation target information for estimation may include newer information than the estimation target information for training. For example, if the estimation target information for estimation is information from January to December 2023, the estimation target information for training may be information from January 2010 to December 2022.

[0050] For example, the estimation target information database DB2 may store a temporal change in the estimation target information. The estimation target information database DB2 may store a spatiotemporal change in the estimation target information. The estimation target information database DB2 may store a plurality of estimation target information. In this embodiment, the estimation target information stored in the estimation target information database DB2 indicates the change in GMS for each month of each prefecture. In the example of FIG. 5, the estimation target information indicates the change in GMS for each month from January to December 2023 for each prefecture. The combination of time and space in the spatiotemporal change regarding the estimation target indicated by the estimation target information stored in the estimation target information database DB2 may be any combination. Note that the estimation target information stored in the estimation target database may indicate only either the change in the estimation target in the past period or the latest estimation target at the present time.

[0051] FIG. 6 is a diagram showing an example of the macro environmental variable database DB3. The macro environmental variable database DB3 is a database in which macro environmental variables input to the weight estimation model WM when estimating an estimation target are stored. The macro environmental variables stored in the estimation target database are macro environmental variables for estimation. For this reason, the macro environmental variables for estimation have the same data format as the macro environmental variables for training, but have different specific values. The macro environmental variables for estimation may include newer information than the macro environmental variables for training. For example, if the macro environmental variables for estimation are information from January to December 2023, the macro environmental variables for training may be information from January to December 2010.

[0052] For example, the macro environmental variable database DB3 may store a temporal change in the macro environmental variable. The macro environmental variable database DB3 may store a spatiotemporal change in the macro environmental variable. The macro environmental variable database DB3 may store a plurality of macro environmental variables. In this embodiment, the macro environmental variables stored in the macro environmental variable database DB3 indicate a change in the minimum wage for each month in each prefecture. In the example of FIG. 6, the macro environmental variables indicate a change in the minimum wage for each month from January to December 2023 for each prefecture. If the minimum wage does not change from month to month, the macro environmental variable database DB3 may indicate the minimum wage for each year. The unit of time indicated by the estimation target information stored in the estimation target information database DB2 and the unit of time of the macro environmental variable stored in the macro environmental variable database DB3 may be different from each other.

[0053] The combination of time and space in the spatiotemporal change related to the estimation target indicated by the macro environmental variables stored in the macro environmental variable database DB3 may be any combination. The macro environmental variables stored in the estimation target database may indicate only either the change in the macro environmental variables in the past period or the latest macro environmental variables at the present time. The macro environmental variable database DB3 may also store other macro environmental variables (e.g., macro environmental variables such as weather or politics) other than the minimum wage. The data stored in the data storage unit 200 is not limited to the above examples. The data storage unit 200 may store data necessary for estimating the estimation target. For example, the data storage unit 200 may store data for displaying the estimation result of the overall model on the user terminal 30.

[0054] [Estimation target information acquisition section] The estimation target information acquisition unit 201 acquires estimation target information on an estimation target for which estimation in the future is performed. In this embodiment, the estimation target information is information on management indicators. The estimation target information may indicate the estimation target itself in at least one of the present and the past, but the estimation target information may be other information other than the estimation target. The estimation target information may be information used in estimating the estimation target. In other words, the estimation target information may be information that is correlated with the estimation target, and may not be the estimation target itself in at least one of the present and the past. For example, if the estimation target is a GMS of a certain organization, the estimation target information may indicate a micro environmental variable in this organization, rather than the GMS itself in at least one of the present and the past. When the GMS corresponds to a micro environmental variable, the estimation target information may indicate other micro environmental variables other than the GMS (for example, the number of affiliated stores, the number of campaigns, the number of customers, the number of payments, or other variables).

[0055] In this embodiment, the estimation target information acquisition unit 201 acquires the estimation target information from the estimation target information database DB2. The estimation target information acquisition unit 201 may acquire the estimation target information from a database other than the estimation target information database DB2. The estimation target information acquisition unit 201 may acquire the estimation target information from a computer other than the server 20 or an external information storage medium. For example, the estimation target information acquisition unit 201 may acquire the estimation target information related to a temporal change related to the estimation target. The estimation target information acquisition unit 201 may acquire the estimation target information related to a spatiotemporal change related to the estimation target. The temporal change and the spatiotemporal change are as described above.

[0056] [Macro environment variable acquisition section] The macro environment variable acquisition unit 202 acquires macro environment variables related to the macro environment. In this embodiment, the macro environment variable acquisition unit 202 acquires the macro environment variables from the macro environment variable database DB3. The macro environment variable acquisition unit 202 may acquire the macro environment variables from a database other than the macro environment variable database DB3. The macro environment variable acquisition unit 202 may acquire the macro environment variables from a computer other than the server 20 or an external information storage medium. For example, the macro environment variable acquisition unit 202 may acquire a macro environment variable related to a temporal change related to the estimation target. The macro environment variable acquisition unit 202 may acquire a macro environment variable related to a spatiotemporal change related to the estimation target. The temporal change and the spatiotemporal change are as described above.

[0057] [Interim result information acquisition section] The interim result information acquisition unit 203 acquires interim result information on the estimation results estimated by each of a plurality of intermediate estimation models IM specialized for a specific macro environment, which are intermediate estimation models IM that perform intermediate estimation of an estimation target in the future based on the estimation target information. In this embodiment, the interim result information is information on the intermediate estimation result of a management indicator in the future. For example, the interim result information acquisition unit 203 may acquire the interim result information of each of the plurality of intermediate estimation models IM based on the temporal change indicated by the estimation target information. The interim result information acquisition unit 203 may acquire the interim result information of each of the plurality of intermediate estimation models IM based on the spatiotemporal change indicated by the estimation target information.

[0058] In this embodiment, the intermediate result information acquisition unit 203 acquires intermediate result information estimated by each of a plurality of intermediate estimation models IM, which are a plurality of neural networks in the Mixture of Experts method. For example, the intermediate result information acquisition unit 203 inputs estimation target information to each of a plurality of intermediate estimation models IM. Although a case where the contents of the estimation target information input to each intermediate estimation model IM are the same is taken as an example, the estimation target information input to each intermediate estimation model IM may differ from each other in at least part of its contents. In the example of the estimation target information in FIG. 5, the prefecture and month indicated by the estimation target information input to the intermediate estimation model IM1 may be different from the prefecture and month indicated by the estimation target information input to the intermediate estimation model IM2.

[0059] For example, the intermediate estimation model IM calculates an embedded representation of the estimation target information based on the estimation target information input thereto and the parameters adjusted by learning. The embedded representation is information indicating the characteristics of the estimation target information. The embedded representation may also be called a feature amount. In this embodiment, the embedded representation is expressed in a vector format, but the embedded representation may be expressed in a format other than the vector format. For example, the embedded representation may be expressed in an array, a matrix, a collection of multiple numerical values, a single numerical value, or in another format. The intermediate estimation model IM outputs intermediate result information corresponding to the embedded representation. In this embodiment, since the GMS in a certain organization is finally estimated, the intermediate estimation model IM outputs the GMS corresponding to the macro environment specialized by itself as the intermediate result information. The intermediate result information acquisition unit 203 acquires the intermediate result information output from the intermediate estimation model IM.

[0060] In the example of FIG. 3, the intermediate result information acquisition unit 203 inputs estimation target information to each of the intermediate estimation models IM1 and IM2. Each of the intermediate estimation models IM1 and IM2 calculates an embedded expression of the estimation target information based on its own parameters, and outputs intermediate result information according to the embedded expression. Since the parameters of the intermediate estimation models IM1 and IM2 are different from each other, the intermediate result information output from each of the intermediate estimation models IM1 and IM2 is different from each other. The intermediate result information acquisition unit 203 acquires the intermediate result information output from each of the intermediate estimation models IM1 and IM2. Similarly, when there are three or more intermediate estimation models IM, the intermediate result information acquisition unit 203 inputs estimation target information to each of the three or more intermediate estimation models IM, and acquires intermediate result information from each of the three or more intermediate estimation models IM.

[0061] [Weight acquisition section] The weight acquisition unit 204 acquires the weights w estimated by the weight estimation model WM, which estimates the weights w of each of the intermediate estimation models IM based on the macro environmental variables. For example, the weight acquisition unit 204 may acquire the weights w estimated by the weight estimation model WM based on a temporal change in the macro environmental variables. The weight acquisition unit 204 may acquire the weights w estimated by the weight estimation model WM based on a spatiotemporal change in the macro environmental variables.

[0062] In this embodiment, the weight acquisition unit 204 acquires the weight w estimated by the weight estimation model WM, which is a neural network in the Mixture of Experts method. For example, the weight acquisition unit 204 inputs macro environmental variables to the weight estimation model WM. The weight acquisition unit 204 may input temporal changes in the macro environmental variables (i.e., multiple macro environmental variables arranged in chronological order) or may input spatiotemporal changes in the macro environmental variables (i.e., multiple macro environmental variables arranged in chronological order for each space) to the weight estimation model WM. The weight acquisition unit 204 may input each of the multiple macro environmental variables to the weight estimation model WM. In this case, it is assumed that each of the multiple macro environmental variables for training has been learned in the weight estimation model WM.

[0063] For example, the weight estimation model WM calculates an embedded expression of the macro environmental variable based on the macro environmental variable input thereto and the parameter adjusted by learning. The embedded expression is information indicating the characteristics of the macro environmental variable. The meaning of the embedded expression is as described in the processing of the intermediate estimation model IM. The weight estimation model WM outputs the weight w of each intermediate estimation model IM according to the embedded expression. The weight acquisition unit 204 acquires the weight w of each intermediate estimation model IM output from the weight estimation model WM. In this embodiment, the case where the total value of the weights w is 1 is taken as an example, but the total value of the weights w may not be 1. In the example of FIG. 3, the weight acquisition unit 204 acquires the weight w1 of the intermediate estimation model IM1 and the weight w2 of the intermediate estimation model IM2. Similarly, when three or more intermediate estimation models IM exist, the weight acquisition unit 204 may acquire the weight w of each of the three or more intermediate estimation models IM.

[0064] [Final result information acquisition section] The final result information acquisition unit 205 acquires final result information regarding a final estimation result of the estimation target in the future based on the interim result information of each of the multiple intermediate estimation models IM and the weights w of each of the multiple intermediate estimation models IM. In this embodiment, the final result information is information regarding the final estimation result of the management indicator. The final result information acquisition unit 205 records the final result information in the data storage unit 100.

[0065] The final result information acquisition unit 205 calculates a final estimation result by adding together a value obtained by multiplying the output of each of the multiple intermediate estimation models IM by the weight w of the intermediate estimation model IM. In the example of Fig. 3, the final result information acquisition unit 205 acquires final result information, which is a final estimation result, by adding together a value obtained by multiplying the output of the intermediate estimation model IM1 by the weight w1 of the intermediate estimation model IM1 and a value obtained by multiplying the output of the intermediate estimation model IM2 by the weight w2 of the intermediate estimation model IM2. The final result information acquisition unit 205 executes this series of processes based on a program included in the overall model.

[0066] [3-3. Functions realized on user terminals] For example, the user terminal 30 includes a data storage unit 300, an operation reception unit 301, and a display control unit 302. The data storage unit 300 is realized by the storage unit 32. The operation reception unit 301 and the display control unit 302 are realized by the control unit 31.

[0067] [Data storage section] The data storage unit 300 stores data necessary for a user to use the estimation system 1. For example, the data storage unit 300 stores an application dedicated to the estimation system 1. When a user uses the estimation system 1 from a browser instead of the application, the data storage unit 300 stores the browser.

[0068] [Operation reception section] The operation receiving unit 301 receives various operations from the user. For example, the operation receiving unit 301 receives operations for an application or a browser dedicated to the estimation system 1. The operation receiving unit 301 transmits data indicating the content of the user's operation to the server 20.

[0069] [Display control section] The display control unit 302 causes various screens to be displayed on the display unit 35. For example, the display control unit 302 causes an estimation result screen SC, which will be described later, to be displayed on the display unit 35. The display control unit 302 communicates with the server 20, receives data necessary for displaying the estimation result screen SC, and causes the display unit 35 to display the estimation result screen SC.

[0070] [4. Processing performed by the estimation system] Fig. 7 is a diagram showing an example of processing executed by the estimation system 1. The control units 11, 21, and 31 execute programs stored in the storage units 12, 22, and 32, respectively, to thereby execute the processing in Fig. 7.

[0071] As shown in Fig. 7, the learning terminal 10 learns the overall model based on the training database DB1 (S1). The learning terminal 10 uploads the learned overall model to the server 20 (S2). The server 20 receives the learned overall model and records it in the storage unit 22 (S3). The user terminal 30 executes a process between the server 20 and the user to log in to the estimation system 1, and requests the server 20 to estimate future management indicators (S4). When the server 20 receives the request from the user terminal 30 (S5), it acquires estimation target information from the estimation target information database DB2 (S6).

[0072] The server 20 inputs estimation target information to each of the multiple intermediate estimation models IM (S7). The server 20 acquires intermediate result information output from each of the multiple intermediate estimation models IM (S8). The server 20 acquires macro environmental variables from a macro environmental variable database DB3 (S9). The server 20 inputs macro environmental variables to the weight estimation model WM (S10). The server 20 acquires weights w of each of the multiple intermediate estimation models IM from the weight estimation model WM (S11).

[0073] The server 20 acquires final result information based on the intermediate result information output from each of the multiple intermediate estimation models IM and the weights w of each of the multiple intermediate estimation models IM (S12). The server 20 generates display data for an estimation result screen based on the final result information and transmits it to the user terminal 30 (S13). Upon receiving the display data for the estimation result screen, the user terminal 30 displays the estimation result screen on the display unit 35 (S14), and this process ends.

[0074] FIG. 8 is a diagram showing an example of an estimation result screen. For example, the user terminal 30 displays the estimation result screen SC on the display unit 35 based on the display data received from the server 20. The user can grasp the future management index of the organization to which the user belongs by checking the estimation result screen SC. In the example of FIG. 8, a case is shown in which the final result information shows the estimation results of each of a plurality of time points in the future. In this case, the output part of the training data of the intermediate estimation model IM shows the estimation results that are the correct answers for each of a plurality of time points in the future relative to the input part of the training data. The intermediate estimation model IM can perform estimation at each of a plurality of time points in the future by learning such training data. The final result information may show the estimation result of one time point in the future. In this case, the output part of the training data of the intermediate estimation model IM shows the estimation result that is the correct answer for one time point in the future relative to the input part of the training data. The intermediate estimation model IM can perform estimation at one time point in the future by learning such training data.

[0075] [5. Summary of the embodiment] The estimation system 1 of the present embodiment acquires intermediate result information on the estimation results estimated by each of a plurality of intermediate estimation models IM specialized in a specific macro environment based on the estimation target information. The estimation system 1 acquires the weights w estimated by the weight estimation model WM based on the macro environmental variables. The estimation system 1 acquires final result information based on the intermediate result information of each of the plurality of intermediate estimation models IM and the weights w of each of the plurality of intermediate estimation models IM. The estimation system 1 does not treat the macro environmental variables and the micro environmental variables in the same way, but acquires final result information based on a plurality of intermediate estimation models IM specialized under the macro environment of a specific macro environmental variable and the weights w of each of the plurality of intermediate estimation models IM, thereby improving the estimation accuracy of the future estimation target. That is, the estimation system 1 can perform estimation specific to the macro environmental variables, thereby improving the estimation accuracy of the future estimation target.

[0076] Furthermore, the estimation system 1 acquires estimation target information regarding changes over time regarding the estimation target. The estimation system 1 acquires intermediate result information of each of the multiple intermediate estimation models IM based on the changes over time indicated by the estimation target information. This allows the estimation system 1 to acquire final result information taking into account the changes over time indicated by the estimation target information. The estimation system 1 can further improve the estimation accuracy of the future estimation target by taking into account the correlation between the changes over time and the final result information.

[0077] Furthermore, the estimation system 1 acquires estimation target information related to spatiotemporal changes related to the estimation target. The estimation system 1 acquires interim result information of each of the multiple intermediate estimation models IM based on the spatiotemporal changes indicated by the estimation target information. This allows the estimation system 1 to acquire interim result information that takes into account the spatiotemporal changes indicated by the estimation target information. The estimation system 1 can further improve the estimation accuracy of the future estimation target by taking into account the correlation between the spatiotemporal changes and the final result information.

[0078] Furthermore, the estimation system 1 acquires the weights w estimated by the weight estimation model WM based on the temporal changes in the macroenvironmental variables. This allows the estimation system 1 to acquire final result information based on the weights w that take into account the temporal changes in the macroenvironmental variables. The estimation system 1 can further improve the estimation accuracy of future estimation targets by considering the correlation between the temporal changes in the macroenvironmental variables and the weights w.

[0079] Furthermore, the estimation system 1 obtains the weights w estimated by the weight estimation model WM based on the spatiotemporal changes in the macroenvironmental variables. This allows the estimation system 1 to obtain final result information based on the weights w that take into account the spatiotemporal changes in the macroenvironmental variables. The estimation system 1 can further improve the estimation accuracy of future estimation targets by considering the correlation between the spatiotemporal changes in the macroenvironmental variables and the weights w.

[0080] Furthermore, the estimation system 1 acquires intermediate result information estimated by each of a plurality of intermediate estimation models IM, which are a plurality of neural networks in the Mixture of Experts method. The estimation system 1 acquires weights w estimated by a weight estimation model WM, which is a neural network in the Mixture of Experts method. This allows the estimation system 1 to increase the estimation accuracy of the intermediate result information by having the intermediate estimation models IM, which are experts, output the intermediate result information. Furthermore, the estimation system 1 can acquire weights w of the estimation results of each expert. As a result, the estimation system 1 can increase the estimation accuracy of the final result information.

[0081] Furthermore, the estimation target is a management indicator in the organization. The estimation target information is information related to the management indicator. The macro environmental variables are variables related to the macro environment surrounding the organization. The interim result information is information related to an interim estimation result of the future management indicator. The final result information is information related to the final estimation result of the management indicator. The estimation system 1 can improve the estimation accuracy of future management indicators.

[0082] [6. Modifications] The present disclosure is not limited to the above-described embodiments. The present disclosure can be modified as appropriate without departing from the spirit and scope of the present disclosure.

[0083] 9 is a diagram showing an example of functions realized in the modified example. For example, the learning terminal 10 includes a clustering execution unit 102, a first generation unit 103, a second generation unit 104, and an identification unit 105. Each of the clustering execution unit 102, the first generation unit 103, the second generation unit 104, and the identification unit 105 is realized by the control unit 11. The learning unit 101 includes a first learning unit 101A, a second learning unit 101B, a third learning unit 101C, and a fourth learning unit 101D.

[0084] [6-1. Variation 1] For example, in the embodiment, the entire model is learned all at once based on the training data stored in the training database DB1, but the weight estimation model WM may be learned individually. In the first modification, training data for learning the weight estimation model WM is generated. The weight estimation model WM is learned based on the training data for learning the weight estimation model WM.

[0085] The estimation system 1 of the first modification includes a clustering execution unit 102, a first generation unit 103, and a first learning unit 101A. The clustering execution unit 102 executes clustering based on the training macro environmental variables. Clustering is a process of identifying information with similar characteristics from a plurality of pieces of information. A known method may be used for the clustering. For example, the clustering execution unit 102 may execute clustering of the training macro environmental variables based on K-means clustering, hierarchical clustering, DBSCAN, a Mean-Shift method, a GMM (Gaussian Mixture Models) method, or another method.

[0086] For example, the training macro environmental variables indicate the change in the minimum wage for each month in each prefecture. In this case, the clustering execution unit 102 identifies prefectures with similar trends in the change in GMS for each month by clustering the training macro environmental variables. The number of clusters may be any number, and the number of clusters may be specified by the operator of the learning terminal 10. In the first modification, for the sake of simplicity of explanation, the number of clusters is set to two. The clustering execution unit 102 classifies the prefectures into those belonging to a cluster showing a favorable economic trend and those belonging to a cluster showing a poor economic trend. The number of clusters is set to the same as the number of intermediate estimation models IM.

[0087] The clusters classified by clustering are not limited to economic conditions. The clusters may be any clusters according to the macro environmental variables. For example, if the macro environmental variables indicate winter temperatures, the prefectures are classified into prefectures belonging to a cluster indicating a tendency for warm winters and prefectures belonging to a cluster indicating a tendency for cold winters. If the macro environmental variables indicate whether or not there will be an election, the prefectures may be classified into prefectures belonging to a cluster with many elections and prefectures belonging to a cluster with few elections. Furthermore, the clustering execution unit 102 may cluster each month instead of clustering each prefecture. The clustering execution unit 102 may cluster pairs of prefectures and months.

[0088] The first generation unit 103 generates training data for the weight estimation model WM based on the execution result of the clustering. The input and output parts of the training data for the weight estimation model WM have the same format as the input and output of the weight estimation model WM at the time of estimation. For example, the input part of the training data for the weight estimation model WM is a macro environmental variable for training. The output part of the training data for the weight estimation model WM is each weight w of the multiple intermediate estimation models IM. The output part of the training data for the weight estimation model WM may be manually input by the person operating the learning terminal 10, or may be automatically determined by a training data creation tool.

[0089] In the first modification, the first generating unit 103 generates training data for each cluster, the training data including, as an input part, training macro environmental variables belonging to the cluster, and the weight w according to the cluster, as an output part. For example, the number of clusters is assumed to be n (n is an integer of 2 or more). Furthermore, the number of n is assumed to be the same as the number of intermediate estimation models IM. In this case, the first generating unit 103 generates training data including, as an input part, training macro environmental variables belonging to the k-th cluster (k is an integer of 1 to n), and the weight w according to the macro environment indicated by the k-th cluster is higher than the weight w of the other intermediate estimation models IM, as an output part.

[0090] For example, in the case of the two clusters of the booming economy and the recessionary economy described above, the first generation unit 103 generates training data including, as an input part, the training macro environmental variables belonging to the booming economy cluster, and including, as an output part, a weight w according to the booming economy cluster. When there are two intermediate estimation models IM as shown in Fig. 3 and the intermediate estimation model IM1 is a learning model specialized for the booming economy, the first generation unit 103 generates training data for the weight estimation model WM so that the weight w1 is greater than the weight w2 (for example, so that the weight w1 is greater than 0.5).

[0091] For example, the first generation unit 103 generates training data including a training macro environmental variable belonging to a recession cluster as an input part, and including a weight w according to the recession cluster as an output part. When there are two intermediate estimation models IM as shown in Fig. 3 and the intermediate estimation model IM2 is a learning model specialized for recessions, the first generation unit 103 generates training data for the weight estimation model WM so that the weight w2 is larger than the weight w1 (for example, larger than 0.5). Similarly, when there are three or more clusters and intermediate estimation models IM, the first generation unit 103 generates training data for the weight estimation model WM.

[0092] The first learning unit 101A makes the weight estimation model WM learn the training data of the weight estimation model WM. The first learning unit 101A learns the weight estimation model WM so that when an input portion of the training data of the weight estimation model WM is input, an output portion of the training data of the weight estimation model WM is output. The first learning unit 101A learns the weight estimation model WM by adjusting the parameters of the weight estimation model WM based on the training data of the weight estimation model WM. When the learning is completed, the first learning unit 101A records the learned weight estimation model WM in the data storage unit 100. The first learning unit 101A uploads the learned weight estimation model WM to the server 20. For example, the first learning unit 101A uploads an entire model including the learned weight estimation model WM.

[0093] The algorithm for learning the weight estimation model WM may be a known algorithm used in the field of machine learning. In this embodiment, since the weight estimation model WM is a learning model in the Mixture of Experts method, the first learning unit 101A may perform learning based on the algorithm in the Mixture of Experts method. For example, the first learning unit 101A may cause the weight estimation model WM to learn training data based on a deep learning algorithm. The loss function used during learning may also be a known loss function. The first learning unit 101A calculates a loss, which is an error between the output part of the training data and the output from the weight estimation model WM during learning, based on the loss function. The first learning unit 101A completes learning when the loss becomes small to a certain extent. The first learning unit 101A may cause the weight estimation model WM to learn training data based on other algorithms such as gradient descent or backpropagation.

[0094] In the first modification, after the first learning unit 101A has learned the weight estimation model WM, the learning unit 101 learns the overall model in the same manner as in the embodiment. When combining the later-described second modification with the first modification, the learning of the overall model may not be executed, and the learning of the weight estimation model WM and the intermediate estimation model IM may be performed separately. The weight acquisition unit 204 acquires the weight w estimated by the weight estimation model WM learned by the first learning unit 101A. Although the learning method of the weight estimation model WM is different from the embodiment, the method in which the weight acquisition unit 204 acquires the weight w may be the same as in the embodiment. Other processes during estimation may also be the same as in the embodiment.

[0095] The estimation system 1 of the first modification executes clustering based on the training macro environmental variables. The estimation system 1 generates training data for the weight estimation model WM based on the execution result of the clustering. The estimation system 1 makes the weight estimation model WM learn the training data for the weight estimation model WM. The estimation system 1 acquires the weight w estimated by the weight estimation model WM learned by the first learning unit 101A. This allows the estimation system 1 to learn the weight estimation model WM based on training data specific to the weight estimation model WM, thereby improving the estimation accuracy of the weight w. The estimation system 1 acquires final estimation information based on the accurate weight w, thereby further improving the estimation accuracy of a future estimation target.

[0096] [6-2. Variation 2] For example, training data for learning the intermediate estimation model IM may be generated based on the execution result of the clustering in Modification 1. In Modification 2, the intermediate estimation model IM is learned based on the training data for learning the intermediate estimation model IM. That is, the intermediate estimation model IM is learned individually.

[0097] The estimation system 1 of the first modification includes a second generation unit 104 and a second learning unit 101B. The second generation unit 104 generates training data for each of the multiple intermediate estimation models IM based on the estimation target information for training and the execution result of clustering. The input part and the output part of the training data of the intermediate estimation model IM have the same format as the input and the output of the intermediate estimation model IM at the time of estimation. For example, the input part of the training data of the intermediate estimation model IM is the estimation target information for training. The output part of the training data of the intermediate estimation model IM is the intermediate result information that is the correct answer at the time of learning. The output part of the training data of the intermediate estimation model IM may be manually input by the person operating the learning terminal 10, or may be automatically determined by a training data creation tool.

[0098] In the second modification, the second generation unit 104 generates training data for each cluster, which includes, as an input portion, training estimation target information belonging to the cluster, and includes, as an output portion, intermediate result information corresponding to the cluster. For example, the number of clusters is assumed to be n (n is an integer of 2 or more). Furthermore, the number of n is assumed to be the same as the number of intermediate estimation models IM. In this case, the second generation unit 104 generates training data which includes, as an input portion, training estimation target information belonging to the k-th cluster (k is an integer of 1 to n), and includes, as an output portion, intermediate result information such that the intermediate result information of the intermediate estimation model IM corresponding to the macro environment indicated by the k-th cluster is higher than the intermediate result information of the other intermediate estimation models IM.

[0099] For example, consider a case where each prefecture belongs to one of the two clusters of economic boom and economic downturn described above. The second generation unit 104 generates training data that includes, as an input part, estimation target information of prefectures that belong to the economic boom cluster among the estimation target information for training, and that includes, as an output part, intermediate result information corresponding to the economic boom cluster. When there are two intermediate estimation models IM as shown in FIG. 3 and the intermediate estimation model IM1 is a learning model specialized for economic boom, the second generation unit 104 generates training data for the weight estimation model WM so that the intermediate result information shows a higher value than when the economy is down.

[0100] For example, the second generation unit 104 generates training data including, among the estimation target information for training, estimation target information of prefectures belonging to the recession cluster as an input part, and including intermediate result information corresponding to the recession cluster as an output part. When there are two intermediate estimation models IM as shown in Fig. 3 and the intermediate estimation model IM2 is a learning model specialized for the recession, the second generation unit 104 generates training data for the weight estimation model WM so that the intermediate result information shows a lower value than in the booming economy.

[0101] The second learning unit 101B makes the intermediate estimation model IM learn the training data of each of the multiple intermediate estimation models IM. The second learning unit 101B learns the intermediate estimation model IM so that, when an input portion of the training data of the intermediate estimation model IM is input, an output portion of the training data of the intermediate estimation model IM is output. The second learning unit 101B learns the intermediate estimation model IM by adjusting the parameters of the intermediate estimation model IM based on the training data of the intermediate estimation model IM. When the learning is completed, the second learning unit 101B records the learned intermediate estimation model IM in the data storage unit 100. The second learning unit 101B uploads the learned intermediate estimation model IM to the server 20. For example, the second learning unit 101B uploads an entire model including the learned intermediate estimation model IM.

[0102] The algorithm for learning the intermediate estimation model IM may be a known algorithm used in the field of machine learning. In this embodiment, since the intermediate estimation model IM is a learning model in the Mixture of Experts method, the second learning unit 101B may perform learning based on the algorithm in the Mixture of Experts method. For example, the second learning unit 101B may cause the intermediate estimation model IM to learn training data based on a deep learning algorithm. The loss function used during learning may also be a known loss function. The second learning unit 101B calculates a loss, which is an error between the output part of the training data and the output from the intermediate estimation model IM during learning, based on the loss function. The second learning unit 101B completes learning when the loss becomes small to a certain extent. The second learning unit 101B may cause the intermediate estimation model IM to learn training data based on other algorithms such as gradient descent or backpropagation.

[0103] In the second modification, after the second learning unit 101B has learned the intermediate estimation model IM, the learning unit 101 may learn an overall model in the same manner as in the embodiment, or may not particularly learn an overall model. The intermediate result information acquisition unit 203 in the second modification acquires intermediate result information estimated by each of the multiple intermediate estimation models IM learned by the second learning unit 101B. Although the method of learning the intermediate estimation model IM differs from that of the embodiment, the method of the intermediate result information acquisition unit 203 acquiring the intermediate result information may be the same as in the embodiment. Other processes during estimation may also be the same as in the embodiment.

[0104] The estimation system 1 of the second modification generates training data for each of the multiple intermediate estimation models IM based on the estimation target information for training and the execution result of clustering. The estimation system 1 causes the intermediate estimation model IM to learn the training data for each of the multiple intermediate estimation models IM. The estimation system 1 acquires intermediate result information estimated by each of the multiple intermediate estimation models IM learned by the second learning unit 101B. This allows the estimation system 1 to learn the intermediate estimation model IM based on training data specific to the intermediate estimation model IM, thereby improving the estimation accuracy of the intermediate result information. The estimation system 1 can further improve the estimation accuracy of the future estimation target by acquiring final estimation information based on highly accurate intermediate result information.

[0105] [6-3. Variation 3] For example, as described somewhat in Modification 2, the weight estimation model WM and the intermediate estimation model IM may be individually trained before the overall model is trained. The estimation system 1 of Modification 3 includes a first learning unit 101A, a second learning unit 101B, and a third learning unit 101C.

[0106] The first learning unit 101A of the third modification performs pre-learning of the weight estimation model WM. The pre-learning of the weight estimation model WM is individual learning of the weight estimation model WM performed before learning of the overall model. The first learning unit 101A may perform learning of the weight estimation model WM in the same manner as in the first modification, or may perform learning of the weight estimation model WM based on training data that is not based on clustering as in the first modification. For example, the training database DB1 may store training data of the weight estimation model WM that is created without based on clustering. The input part and the output part of the training data of the weight estimation model WM are as described in the first modification. For example, the training data of the weight estimation model WM may be created manually by a person who operates the learning terminal 10, or may be created automatically by a training data creation tool.

[0107] The second learning unit 101B performs pre-learning of each of the multiple intermediate estimation models IM. The pre-learning of the intermediate estimation model IM is individual learning of the intermediate estimation model IM performed before learning of the overall model. The second learning unit 101B may perform learning of the intermediate estimation model IM in the same manner as in the modified example 2, or may perform learning of the intermediate estimation model IM based on training data that is not based on clustering as in the modified example 2. For example, the training database DB1 may store training data of the intermediate estimation model IM that is created without based on clustering. The input part and the output part of the training data of the intermediate estimation model IM are as described in the modified example 2. For example, the training data of the intermediate estimation model IM may be manually created by a person who operates the learning terminal 10, or may be automatically created by a training data creation tool.

[0108] After completing the pre-learning of the weight estimation model WM and the pre-learning of each of the multiple intermediate estimation models IM, the third learning unit 101C performs overall learning including the weight estimation model WM and the multiple intermediate estimation models IM. That is, the third learning unit 101C performs learning of the overall model. The third learning unit 101C performs learning of the overall model including the pre-learned weight estimation model WM and each of the multiple pre-learned intermediate estimation models IM. The learning performed by the third learning unit 101C may be the same as the learning by the learning unit 101 described in the embodiment. The training data of the overall model is also the same as in the embodiment.

[0109] The intermediate result information acquisition unit 203 of the third modification acquires intermediate result information estimated by each of the multiple intermediate estimation models IM that have been entirely trained. Although the training method of the intermediate estimation model IM is different from that of the embodiment, the method by which the intermediate result information acquisition unit 203 acquires the intermediate result information may be the same as that of the embodiment. The weight acquisition unit 204 acquires the weight w estimated by the weight estimation model WM that has been entirely trained. Although the training method of the weight estimation model WM is different from that of the embodiment, the method by which the weight acquisition unit 204 acquires the weight w may be the same as that of the embodiment. Other processes during estimation may also be the same as those of the embodiment.

[0110] The estimation system 1 of the third modified example performs pre-learning of the weight estimation model WM. The estimation system 1 performs pre-learning of each of the multiple intermediate estimation models IM. After completing the pre-learning of the weight estimation model WM and the pre-learning of each of the multiple intermediate estimation models IM, the estimation system 1 performs overall learning including the weight estimation model WM and the multiple intermediate estimation models IM. The estimation system 1 acquires intermediate result information estimated by each of the multiple intermediate estimation models IM for which overall learning has been performed. The estimation system 1 acquires the weight w estimated by the weight estimation model WM for which overall learning has been performed. Thereby, the estimation system 1 can improve the estimation accuracy of the weight w by the pre-learning specific to the weight estimation model WM. The estimation system 1 can improve the estimation accuracy of the intermediate result information by the pre-learning specific to the intermediate estimation model IM. The estimation system 1 can further improve the estimation accuracy of the future estimation target by acquiring final estimation information based on the highly accurate weight w and the highly accurate intermediate result information.

[0111] [6-4. Variation 4] For example, when there are many macro environmental variables, some of them contribute to the estimation of the estimation target, while others do not contribute much to the estimation of the estimation target. For this reason, the estimation system 1 may identify macro environmental variables that contribute to the estimation of the estimation target from among the many macro environmental variables. The estimation system 1 may learn at least one of a plurality of intermediate estimation models IM and a weight estimation model WM based on the macro environmental variables that contribute to the estimation of the estimation target.

[0112] The estimation system 1 of the fourth modification includes an identification unit 105 and a fourth learning unit 101D. The identification unit 105 identifies at least one macro environmental variable that contributes to the estimation of the estimation target in the future from among the multiple macro environmental variables based on the weight coefficient model. The identification unit 105 may identify at least one macro environmental variable that contributes to the estimation of the estimation target by using a feature selection or feature reduction method in a machine learning method. For example, the identification unit 105 may calculate the importance of the macro environmental variables as features, and identify a macro environmental variable with a relatively high importance as at least one macro environmental variable that contributes to the estimation of the estimation target. A known index such as information gain or Gini importance may be used as the importance of the feature. The identification unit 105 may select at least one of the multiple macro environmental variables based on a known feature selection method that selects a feature that contributes to the estimation of the learning model. For example, the identification unit 105 may select a macro environmental variable whose importance calculated based on the feature selection method is equal to or greater than a threshold from among the multiple macro environmental variables, or may select a predetermined number of macro environmental variables in order of increasing importance.

[0113] The fourth learning unit 101D learns at least one of the multiple intermediate estimation models IM and the weight estimation model WM based on at least one macro environmental variable identified by the identification unit 105. In the fourth modification, an example is given in which the fourth learning unit 101D learns both the multiple intermediate estimation models IM and the weight estimation model WM, but the fourth learning unit 101D may learn only one of the multiple intermediate estimation models IM or the weight estimation model WM. The fourth learning unit 101D learns at least one of the multiple intermediate estimation models IM and the weight estimation model WM based on training data that includes at least one macro environmental variable identified by the identification unit 105 as a macro environmental variable for training.

[0114] For example, in the case where the overall model is learned all at once as in the embodiment, the fourth learning unit 101D generates training data for the overall model based on at least one macro environmental variable identified by the identification unit 105. The fourth learning unit 101D acquires only training data including at least one macro environmental variable identified by the identification unit 105 from among the training data stored in the training database DB1 described in the embodiment. The fourth learning unit 101D causes the overall model to learn the acquired training data. The fourth learning unit 101D may re-learn the overall model based on the acquired training data, or may re-learn the overall model that has already been learned.

[0115] For example, as in the first modification, when the weight estimation model WM is individually learned, the fourth learning unit 101D generates training data for the weight estimation model WM based on at least one macro environmental variable identified by the identification unit 105. The fourth learning unit 101D performs clustering as in the first modification based on at least one macro environmental variable identified by the identification unit 105, and generates training data for the weight estimation model WM. The fourth learning unit 101D makes the weight estimation model WM learn the training data for the weight estimation model WM. The method of making the weight estimation model WM learn the training data for the weight estimation model WM may be the same as that of the first modification. Note that the fourth learning unit 101D may generate training data including at least one macro environmental variable identified by the identification unit 105 as a macro environmental variable for training without using clustering as in the first modification, and make the weight estimation model WM learn the training data.

[0116] For example, as in the second modification, when the learning of the intermediate estimation model IM is performed individually, the fourth learning unit 101D generates training data for the intermediate estimation model IM based on the result of clustering based on at least one macro environmental variable identified by the identification unit 105. The fourth learning unit 101D performs clustering as in the second modification based on at least one macro environmental variable identified by the identification unit 105 to generate training data for the intermediate estimation model IM. The fourth learning unit 101D makes the intermediate estimation model IM learn the training data for the intermediate estimation model IM. The method of making the intermediate estimation model IM learn the training data for the intermediate estimation model IM may be the same as that of the second modification. Note that the fourth learning unit 101D may generate training data including at least one macro environmental variable identified by the identification unit 105 as a macro environmental variable for training without using clustering as in the second modification, and make the intermediate estimation model IM learn the training data.

[0117] The estimation system 1 of the fourth modification example identifies, based on a weighting coefficient model, at least one macro environmental variable that contributes to estimation of a future estimation target among a plurality of macro environmental variables. The estimation system 1 learns at least one of a plurality of intermediate estimation models IM and a weighting estimation model WM based on the at least one macro environmental variable. The estimation system 1 can improve the accuracy of at least one of the plurality of intermediate estimation models IM and the weighting estimation model WM. The estimation system 1 can further improve the estimation accuracy of a future estimation target.

[0118] [6-5. Other Modifications] For example, the above modifications 1 to 4 may be combined.

[0119] For example, the functions described as being realized by the learning terminal 10 may be realized by the server 20, the user terminal 30, or another computer. The processes described as being realized by the server 20 may be realized by the learning terminal 10, the user terminal 30, or another computer. The functions of the estimation system 1 may be shared among multiple computers.

[0120] [7. Notes] For example, the estimation system may be configured as follows. (1) an estimation target information acquisition unit that acquires estimation target information regarding an estimation target for which estimation in the future is to be performed; a macro environment variable acquisition unit for acquiring a macro environment variable related to the macro environment; an intermediate estimation model for performing intermediate estimation of the estimation target in the future based on the estimation target information, and an intermediate result information acquisition unit for acquiring intermediate result information regarding estimation results estimated by each of a plurality of the intermediate estimation models specialized for a specific macro environment; a weight acquisition unit that acquires weights estimated by a weight estimation model that estimates weights of the plurality of intermediate estimation models based on the macro environmental variables; a final result information acquisition unit that acquires final result information regarding a final estimation result of the estimation target in the future based on the intermediate result information of each of the plurality of intermediate estimation models and the weight of each of the plurality of intermediate estimation models; An estimation system comprising: (2) The estimation target information acquisition unit acquires the estimation target information related to a time change related to the estimation target, the intermediate result information acquisition unit acquires the intermediate result information of each of the plurality of intermediate estimation models based on the temporal change indicated by the estimation target information. The estimation system according to (1). (3) the estimation target information acquisition unit acquires the estimation target information relating to a spatiotemporal change relating to the estimation target; the intermediate result information acquisition unit acquires the intermediate result information of each of the plurality of intermediate estimation models based on the spatiotemporal change indicated by the estimation target information. (2) An estimation system according to the present invention. (4) The weight acquisition unit acquires the weights estimated by the weight estimation model based on a temporal change in the macro environmental variables. The estimation system according to any one of (1) to (3). (5) The weight acquisition unit acquires the weights estimated by the weight estimation model based on spatiotemporal changes in the macro environmental variables. The estimation system according to any one of (1) to (4). (6) The estimation system comprises: A clustering execution unit that executes clustering based on the macro environmental variables for training; a first generation unit that generates training data for the weight estimation model based on a result of the clustering; a first learning unit that causes the weight estimation model to learn the training data of the weight estimation model; Further comprising: The weight acquisition unit acquires the weights estimated by the weight estimation model learned by the first learning unit. The estimation system according to any one of (1) to (5). (7) The estimation system comprises: a second generation unit that generates training data for each of the plurality of intermediate estimation models based on the estimation target information for training and a result of the clustering; a second learning unit that learns the training data of each of the plurality of intermediate estimation models into the intermediate estimation model; Including, the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the plurality of intermediate estimation models trained by the second learning unit. (6) An estimation system according to the present invention. (8) The estimation system comprises: A first learning unit that performs pre-learning of the weight estimation model; A second learning unit that performs pre-learning on each of the plurality of intermediate estimation models; a third learning unit that performs learning of the entire system including the weight estimation model and the intermediate estimation models after the pre-learning of the weight estimation model and the pre-learning of each of the intermediate estimation models are completed; Including, the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the plurality of intermediate estimation models that have been subjected to the overall learning, The weight acquisition unit acquires the weights estimated by the weight estimation model that has been entirely trained. An estimation system according to any one of (1) to (7). (9) The estimation system comprises: An identification unit that identifies at least one of the plurality of macro environmental variables that will contribute to the estimation of the estimation target in the future based on the weighting coefficient model; a fourth learning unit that learns at least one of the intermediate estimation models and the weight estimation model based on the at least one macro environmental variable; The estimation system according to any one of claims 1 to 8, further comprising: (10) the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the intermediate estimation models, which are a plurality of neural networks in a Mixture of Experts method; The weight acquisition unit acquires the weights estimated by the weight estimation model, which is a neural network in the Mixture of Experts method. An estimation system according to any one of (1) to (9). (11) The estimation target is a management indicator in an organization, The estimation target information is information regarding the management index, The macro environmental variables are variables related to the macro environment surrounding the organization, The interim result information is information regarding an interim estimated result of the management indicator in the future, The final result information is information regarding a final estimated result of the management indicator. The estimation system according to any one of (1) to (10). [Explanation of symbols]

[0121] 1 Estimation system, 10 Learning terminal, 11, 21, 31 Control unit, 12, 22, 32 Memory unit, 13, 23, 33 Communication unit, 14, 34 Operation unit, 15, 35 Display unit, 20 Server, 30 User terminal, N Network, IM, IM1, IM2 Intermediate estimation model, SC Estimation result screen, WM Weight estimation model, 100, 200, 300 Data memory unit, 101 Learning unit, 101A First learning unit, 101B Second learning unit, 101C Third learning unit, 101D Fourth learning unit, 102 Clustering execution unit, 103 First generation unit, 104 Second generation unit, 105 Identification unit, 201 Estimation target information acquisition unit, 202 Macro environmental variable acquisition unit, 203 Intermediate result information acquisition unit, 204 Weight acquisition unit, 205 Final result information acquisition unit, 301 An operation reception unit, 302 a display control unit, DB1 a training database, DB2 an estimation target information database, and DB3 a macro environmental variable database.

Claims

1. an estimation target information acquisition unit that acquires estimation target information regarding an estimation target for which estimation in the future is to be performed; a macro environment variable acquisition unit for acquiring a macro environment variable related to the macro environment; an intermediate result information acquisition unit that acquires intermediate result information regarding estimation results estimated by each of a plurality of intermediate estimation models specialized for a specific macro environment, the intermediate estimation models being intermediate estimation models that perform intermediate estimation of the estimation target in the future based on the estimation target information; a weight acquisition unit that acquires weights estimated by a weight estimation model that estimates weights of the plurality of intermediate estimation models based on the macro environmental variables; a final result information acquisition unit that acquires final result information regarding a final estimation result of the estimation target in the future based on the intermediate result information of each of the plurality of intermediate estimation models and the weight of each of the plurality of intermediate estimation models; An estimation system comprising:

2. The estimation target information acquisition unit acquires the estimation target information related to a time change related to the estimation target, the intermediate result information acquisition unit acquires the intermediate result information of each of the plurality of intermediate estimation models based on the temporal change indicated by the estimation target information. The estimation system according to claim 1 .

3. the estimation target information acquisition unit acquires the estimation target information relating to a spatiotemporal change relating to the estimation target; the intermediate result information acquisition unit acquires the intermediate result information of each of the plurality of intermediate estimation models based on the spatiotemporal change indicated by the estimation target information. The estimation system according to claim 2 .

4. The weight acquisition unit acquires the weights estimated by the weight estimation model based on a temporal change in the macro environmental variables. The estimation system according to any one of claims 1 to 3.

5. The weight acquisition unit acquires the weights estimated by the weight estimation model based on spatiotemporal changes in the macro environmental variables. The estimation system according to any one of claims 1 to 3.

6. The estimation system comprises: A clustering execution unit that executes clustering based on the macro environmental variables for training; a first generation unit that generates training data for the weight estimation model based on a result of the clustering; a first learning unit that causes the weight estimation model to learn the training data of the weight estimation model; Further comprising: The weight acquisition unit acquires the weights estimated by the weight estimation model learned by the first learning unit. The estimation system according to any one of claims 1 to 3.

7. The estimation system comprises: a second generation unit that generates training data for each of the plurality of intermediate estimation models based on the estimation target information for training and a result of the clustering; a second learning unit that learns training data for each of the plurality of intermediate estimation models into the intermediate estimation model; Including, the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the plurality of intermediate estimation models trained by the second learning unit. The estimation system according to claim 6.

8. The estimation system comprises: A first learning unit that performs pre-learning of the weight estimation model; a second learning unit that performs pre-learning on each of the plurality of intermediate estimation models; a third learning unit that performs learning of the entire system including the weight estimation model and the plurality of intermediate estimation models after pre-learning of the weight estimation model and pre-learning of each of the plurality of intermediate estimation models are completed; Including, the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the plurality of intermediate estimation models that have been subjected to the overall learning, The weight acquisition unit acquires the weights estimated by the weight estimation model that has been entirely trained. The estimation system according to any one of claims 1 to 3.

9. The estimation system comprises: an identification unit that identifies, based on the weighting coefficient model, at least one of the plurality of macro environmental variables that will contribute to the estimation of the estimation target in the future; a fourth learning unit that learns at least one of the intermediate estimation models and the weight estimation model based on the at least one macro environmental variable; The estimation system according to any one of claims 1 to 3, further comprising:

10. the intermediate result information acquisition unit acquires the intermediate result information estimated by each of the intermediate estimation models, which are a plurality of neural networks in a Mixture of Experts method; The weight acquisition unit acquires the weights estimated by the weight estimation model, which is a neural network in the Mixture of Experts method. The estimation system according to any one of claims 1 to 3.

11. The estimation target is a management indicator in an organization, The estimation target information is information regarding the management index, The macro environmental variables are variables related to the macro environment surrounding the organization, The interim result information is information regarding an interim estimated result of the management indicator in the future, The final result information is information regarding a final estimated result of the management indicator. The estimation system according to any one of claims 1 to 3.

12. A computer comprising: An estimation target information acquisition step of acquiring estimation target information regarding an estimation target for which estimation in the future is performed; a macro environment variable acquisition step for acquiring a macro environment variable related to the macro environment; an intermediate result information acquisition step of acquiring intermediate result information on estimation results estimated by each of a plurality of intermediate estimation models specialized for a specific macro environment, the intermediate estimation models being intermediate estimation models for the estimation object in the future based on the estimation object information; a weight acquisition step of acquiring weights estimated by a weight estimation model that estimates weights of each of the plurality of intermediate estimation models based on the macro environmental variables; a final result information acquisition step of acquiring final result information regarding a final estimation result of the estimation target in the future based on the intermediate result information of each of the plurality of intermediate estimation models and the weight of each of the plurality of intermediate estimation models; The estimation method to perform.

13. an estimation target information acquisition unit that acquires estimation target information regarding an estimation target for which estimation in the future is to be performed; a macro environment variable acquisition unit for acquiring macro environment variables related to the macro environment; an intermediate result information acquisition unit that acquires intermediate result information regarding estimation results estimated by each of a plurality of intermediate estimation models specialized for a specific macro environment, the intermediate estimation models being intermediate estimation models for performing intermediate estimation of the estimation target in the future based on the estimation target information; a weight acquisition unit that acquires weights estimated by a weight estimation model that estimates weights for each of the plurality of intermediate estimation models based on the macro environmental variables; a final result information acquisition unit that acquires final result information regarding a final estimation result of the estimation target in the future based on the intermediate result information of each of the plurality of intermediate estimation models and the weight of each of the plurality of intermediate estimation models; A program that makes a computer function as a

Images