Prediction device, prediction method, and prediction program

Abstract

Enables highly accurate time series prediction. [Solution] The performance acquisition unit acquires performance data, which is time-series data for a first period, including dates and values ​​of a dependent variable on those dates. The base estimation unit uses a base model to calculate a base estimate, which is an estimate of the dependent variable for a second period following the first period, from the performance data. The base model includes a trend term representing fluctuations in the dependent variable over the period related to the time-series data and a periodic term representing periodic fluctuations in the dependent variable in the time-series data. The residual estimation unit uses a residual model to calculate an estimate of a residual for the second period. The residual model is a machine learning model trained to output an estimate of a residual using the residual between the value of the dependent variable in the performance data and the base estimate. The prediction unit calculates the sum of the base estimate and the residual estimate for the second period as a predicted value of the dependent variable for the second period.

Description

[Technical Field]

[0001] The present invention relates to a prediction device, a prediction method, and a prediction program. [Background technology]

[0002] There are known employment matching services that match employers with job seekers. Some of these jobs require the results of health tests. For example, Patent Document 1 discloses a technology that manages stool test results and other personal information of job seekers in a system that matches employers with job seekers. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-001994 Summary of the Invention [Problem to be solved by the invention]

[0004] To have job seekers undergo health tests, test kits must be provided to them. Therefore, it is necessary to manage inventory of test kits so that test kits are available when job seekers apply for jobs. Therefore, highly accurate demand forecasting is required. An object of the present invention is to provide a prediction device, a prediction method, and a prediction program that enable highly accurate time series prediction. [Means for solving the problem]

[0005] One aspect of the present disclosure provides a prediction device including: an actual result acquisition unit that acquires actual result data, which is time-series data relating to a first period, including dates and values ​​of a dependent variable on the dates; a base estimation unit that calculates, from the actual data, base estimation data, which is time-series data of a dependent variable relating to a second period following the first period, using a base model that is a trained machine learning model including a trend term that represents a fluctuation in the dependent variable over the period relating to the time-series data and a periodic term that represents a periodic fluctuation of the dependent variable in the time-series data; a residual estimation unit that calculates residual estimation data, which is an estimate of a residual relating to the second period, using a residual model that is a machine learning model trained to output an estimate of a residual using residual data between the value of the dependent variable in the actual data and the base estimation data; and a prediction unit that calculates the sum of the base estimation data and the residual estimation data as a predicted value of the dependent variable relating to the second period.

[0006] One aspect of the present disclosure provides a prediction method including the steps of: acquiring performance data, which is time-series data relating to a first period, the data including dates and values ​​of a dependent variable on the dates; calculating, from the performance data, base estimation data, which is time-series data of a dependent variable relating to a second period following the first period, using a base model, which is a trained machine learning model, the base model including a trend term representing a fluctuation in the dependent variable during the period relating to the time-series data and a periodic term representing a periodic fluctuation of the dependent variable in the time-series data; calculating residual estimation data, which is an estimate of the residual relating to the second period, using a residual model, which is a machine learning model trained to output an estimate of the residual, using residual data between the dependent variable values ​​of the performance data and the base estimation data; and calculating the sum of the base estimation data and the residual estimation data as a predicted value of the dependent variable relating to the second period.

[0007] One aspect of the present disclosure provides a prediction program that causes a computer to function as: an actual performance acquisition unit that acquires actual performance data, which is time-series data relating to a first period, including dates and values ​​of a dependent variable on the dates; a base estimation unit that calculates, from the actual performance data, base estimated data, which is time-series data of a dependent variable relating to a second period following the first period, using a base model that is a trained machine learning model including a trend term that represents a fluctuation in the dependent variable over the period relating to the time-series data and a periodic term that represents a periodic fluctuation of the dependent variable in the time-series data; a residual estimation unit that calculates residual estimated data, which is an estimate of a residual relating to the second period, using a residual model that is a machine learning model trained to output an estimate of a residual, using residual data between the value of the dependent variable in the actual performance data and the base estimated data; and a prediction unit that calculates the sum of the base estimated data and the residual estimated data as a predicted value of the dependent variable relating to the second period. [Effects of the Invention]

[0008] According to at least one of the above aspects, highly accurate time series prediction can be performed for the dependent variable. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a schematic diagram showing the configuration of a recruitment system 1 according to a first embodiment. [Figure 2] 1 is a flowchart showing a method for predicting demand for test kits by the recruitment system 1 according to the first embodiment. [Figure 3] FIG. 1 is a schematic block diagram illustrating the configuration of a computer according to at least one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] First Embodiment Hereinafter, the embodiments will be described in detail with reference to the drawings. FIG. 1 is a schematic diagram showing the configuration of a recruitment system 1 according to a first embodiment. The recruitment system 1 provides a matching service for spot work. Depending on the content of the spot work, health tests using test kits may be required. Examples of health tests include stool tests, urine tests, blood tests, and mucosal tests. In addition to providing a matching service, the recruitment system according to the first embodiment forecasts demand for test kits.

[0011] The recruitment system 1 comprises a server 10. The server 10 is communicably connected to a business operator terminal T1, a store terminal T2, a testing company terminal T3, and a worker terminal T4 via a network such as the Internet. The server 10 functions as a web server. The business operator terminal T1 is operated by a person in charge of the business that employs the worker. The store terminal T2 is operated by a person in charge of a store owned by the business. The testing company terminal T3 is operated by a person in charge of a testing company that provides test kits. The worker terminal T4 is operated by a worker. The worker is a user of the recruitment system 1.

[0012] The server 10 includes a database 101 , a management unit 102 , a service providing unit 103 , a performance acquisition unit 104 , a base learning unit 105 , a base estimation unit 106 , a residual calculation unit 107 , a residual learning unit 108 , a residual estimation unit 109 , and a prediction unit 110 .

[0013] The database 101 stores a store table 1011 , a worker table 1012 , a test request table 1013 , a test kit management table 1014 , a job offer management table 1015 , and an application management table 1016 .

[0014] The store table 1011 is a table that stores master data related to stores that employ workers. The store table 1011 stores store IDs, authentication information, store names, addresses, contact information, and the like in association with each other. The store ID is an ID that uniquely identifies a store. The authentication information is information used by store personnel to log in to the server 10. The authentication information may be, for example, a combination of an account name and a password.

[0015] The worker table 1012 is a table that stores master data related to workers. The worker table 1012 stores, in association with each other, worker IDs, authentication information, names, addresses, contact information, and expiration dates of health test certificates. The worker ID is an ID for uniquely identifying a worker. The expiration date of the health test certificate is the expiration date of the most recent test certificate (negative certificate) issued by the testing company. If the job that a worker desires requires a negative certificate, the worker cannot accept the job unless the expiration date is later than the working date. The expiration date of the health test certificate may be a value obtained by joining the test request table 1013 and the test kit management table 1014 using the worker ID as a key.

[0016] The test request table 1013 is a table that stores information about test requests made by workers. The test request table 1013 stores the test request ID, worker ID, request date, test status, and kit ID in association with each other. The test request ID is an ID that uniquely identifies the test request made by a worker. The worker ID indicates the worker who requested the test. Values ​​for the test status include, for example, requested, kit registered, and test results registered. The kit ID is an ID that uniquely identifies the test kit provided to the worker by the testing provider. The kit ID is entered when the testing provider assigns a test kit to a worker in response to a test request.

[0017] The test kit management table is a table that manages information about test kits. The test kit management table stores kit IDs, test dates, test results, and expiration dates in association with each other. The test date is the date on which the testing provider conducted a test using the test kit returned by the worker and obtained the results.

[0018] The recruitment management table 1015 is a table that manages recruitment information provided to workers by stores. The recruitment management table 1015 stores, in association with each other, a recruitment ID, store ID, creation date, description, work dates, work hours, recruitment deadline, number of people being recruited, test certificate required flag, and recruitment status. The recruitment ID is an ID that uniquely identifies the recruitment information. The store ID indicates the store that created the recruitment information. The creation date is the date the recruitment information was created. The recruitment information is provided to workers from the creation date until the recruitment deadline. The test certificate required flag is a flag that indicates whether a negative health test certificate is required. Values ​​for the recruitment status include, for example, recruiting, recruitment completed, working, and work completed.

[0019] The application management table 1016 is a table for managing applications for job postings by workers. The application management table 1016 stores job posting IDs, worker IDs, and cancellation flags in association with each other. The application management table 1016 uses the decryption key for the job posting ID and worker ID as the primary key. The job posting ID represents the job posting that the worker has applied for. The worker ID represents the worker who has applied for the job posting. The cancellation flag indicates whether the application for the job posting has been canceled. Cancellation may occur due to the worker's or the store's circumstances, or may occur due to a positive test result.

[0020] The management unit 102 performs processes related to the management functions of the job recruitment service. For example, the management unit 102 performs login processing for users via terminals and manages the master data stored in the database 101. The service providing unit 103 performs processing related to the provision of recruitment services. For example, the service providing unit 103 provides recruitment information to workers, registers recruitment information by businesses or stores, and allows workers to apply for recruitment.

[0021] The result acquisition unit 104, the base learning unit 105, the base estimation unit 106, the residual calculation unit 107, the residual learning unit 108, the residual estimation unit 109, and the prediction unit 110 perform demand prediction for the test kits. In other words, the server 10 is an example of a prediction device.

[0022] The performance acquisition unit 104 acquires performance data, which is time-series data regarding demand for test kits, from the database 101. The performance data includes date information regarding the date (target date), the number of applications for test kits on the target date, and external variable information related to the applications for the test kits. The performance acquisition unit 104 acquires performance data related to the period up to the most recent date recorded in the database 101 (first period). The number of applications for test kits represents the number of test kits in demand.

[0023] The date information may include the year, month, day of the week, week, date, and event flag. In date information, months may be represented by trigonometric functions with one cycle per year. Specifically, months may be represented as sin(2π / 12π*M) or cos(2π / 12*M), where M is an integer between 1 and 12, inclusive, representing the month. In date information, days of the week may be represented by trigonometric functions with one cycle per week. Specifically, days of the week may be represented as sin(2π / 7*WD) or cos(2π / 7*WD), where WD is an integer between 1 and 7, inclusive, representing the day of the week. In date information, weeks may be represented by trigonometric functions with one cycle per year. Specifically, weeks may be represented as sin(2π / 52*W) or cos(2π / 52*W), where W is an integer between 1 and 52, inclusive, representing the week. In date information, days may be represented by trigonometric functions with one cycle per year and trigonometric functions with one cycle per month. Specifically, a day may be expressed as sin(2π / 365*D), cos(2π / 365*D), where D is an integer between 1 and 365, inclusive, representing the day. A day may be expressed as sin(2π / 31*MD), cos(2π / 31*MD), where MD is an integer between 1 and 31, inclusive, representing the day of the month. An event flag may be a Boolean value that indicates whether or not a day is a public holiday, whether or not it is a weekend, whether or not it is the day before a holiday, whether or not it is the day after a holiday, etc. An event flag may also indicate whether or not there is an event.

[0024] The external variable information may include the most recent number of applications for test kits, the number of workers, the number of stores hiring, the number of job openings, the number of people being recruited, etc. The most recent number of test kit applications may be, for example, the total number of applications up to 7 days prior to the date related to the performance data, the total number of applications up to 14 days prior, or the total number of applications up to 28 days prior. The most recent number of test kit applications affects the number of test kit applications on the target date. The number of workers may be the total number of workers registered in database 101 on the target date, the number of workers who have applied for a test kit by the target date, the number of workers who have experience with testing using a test kit within a specified period (30 days, 60 days, 90 days, 180 days, 360 days, etc.) immediately preceding the target date, the number of workers who have applied for a job posting, the number of workers who have applied for a job posting that requires testing, the number of workers who have applied for a test kit and are applying for a job posting within a specified period immediately preceding the target date, etc. The number of workers is related to the number of applications for test kits. This is because the greater the number of workers, the greater the likelihood of applications for test kits. The number of stores includes the number of stores with job postings whose target dates coincide with work days, the number of stores with job postings valid on the target dates, the number of stores with job postings valid on the target dates that require testing, etc. The number of stores is related to the number of test kit applications, as the greater the number of stores, the greater the likelihood of test kit applications. The number of job openings includes the number of job openings whose target date coincides with a work day, the number of job openings active on the target date, and the number of job openings active on the target date that require testing. The number of job openings is related to the number of test kit applications, as the higher the number of job openings, the greater the likelihood of test kit applications. The number of job openings can be the total number of job openings for which the target date coincides with a work day, the total number of job openings for which the target date is valid, the total number of job openings for which the target date is valid and which require testing, etc. The number of job openings is related to the number of requests for test kits, as the more job openings there are, the more likely test kit requests there will be.

[0025] The base learning unit 105 performs learning processing on a base model, which is a machine learning model for predicting time-series data, based on performance data. The base model may be, for example, a Prophet model. The Prophet model is expressed as the sum of a trend term representing fluctuations in the objective variable over a period related to the time-series data, a periodic term representing periodic fluctuations in the objective variable in the time-series data, and an event term representing the impact of non-periodic events on the objective variable. The trend term of the Prophet model is a function that approximates long-term fluctuations in the objective variable over the entire period of the training dataset using a combination of lines or curves. The periodic term of the Prophet model is a function representing short-term or medium-term fluctuations in the objective variable over a period such as a year, month, or week. The periodic term may be expressed as a Fourier series. The event term represents the impact of events such as holidays, public holidays, and event dates on the objective variable. The event term may be a function that outputs 0 if the target date is not the day of the event and outputs the learned parameter value if the target date is the day of the event. The trend term, periodic term, and event term all use the date as an explanatory variable. The base learning unit 105 calculates the coefficients of the trend term, periodic term, and event term so as to minimize the error with the time series of the number of test kit orders represented by the performance data. As a result, when a date is input, the base model is trained to output the sum of the value of the trend term for that date, the value of the periodic term for that date, and the value of the event term for that date as the number of test kit orders for that date. The hyperparameters of the base model may be optimized using an algorithm such as Bayesian optimization. In addition, the base model according to other embodiments may be other models having trend terms and periodic terms, such as the Silverkite model, the SARIMA (Seasonal Autoregressive Integrated Moving Average) model, the Holt-Winters model, the TBATS model, or the MSTL model.

[0026] When learning the base model, the base learning unit 105 may optimize the hyperparameters by performing cross-validation using a part of the first period as a learning dataset and the rest as a validation dataset.

[0027] The base estimation unit 106 inputs a date into the trained base model to obtain a base estimate of the number of test kits in demand for that date. The base estimate is an estimate that does not include noise components. The base estimation unit 106 inputs a date of a first period into the trained base model to obtain a base estimate, which is an estimate of the number of test kits in demand for the first period. The base estimate includes an error (noise component) relative to the actual measured value recorded in the database 101. The base estimation unit 106 inputs a date of a second period that is later than the first period into the trained base model to obtain a base estimate of the number of test kits in demand for the second period.

[0028] The residual calculation unit 107 calculates the residual between the base estimate of the demand quantity of test kits for the first period obtained by the base estimation unit 106 and the number of applications for test kits (actual measured value of the demand quantity) recorded in the database 101. In this way, the residual calculation unit 107 obtains a time series of the residual between the estimate and actual measured value of the demand quantity for the first period.

[0029] The residual learning unit 108 performs a learning process for a residual model, which is a machine learning model for estimating the residual value for the base estimate, which is the objective variable, from explanatory variables including date information and external variable information. The residual model may be, for example, a LightGBM model. LightGBM is a gradient boosting framework. That is, the residual learning unit 108 creates a learning dataset using date information and external variable information from the performance data acquired by the performance acquisition unit 104 as input samples and the residual calculated by the residual calculation unit 107 as output samples, and trains the residual model using the learning dataset. The residual learning unit 108 may replace the number of recent test kit orders, which is part of the external variable information, with the residual between the actual measured number of test kits most recently in demand and the base estimate. The residual model may estimate the residual value without using part of the date information and external variable information acquired by the performance acquisition unit 104. Note that the base model according to other embodiments may be other regression models such as a neural network model or a random forest model.

[0030] When learning the residual model, the residual learning unit 108 may optimize the hyperparameters by performing cross-validation using a part of the first period as a learning dataset and the rest as a validation dataset.

[0031] The residual estimation unit 109 inputs date information and external variable information relating to the second period into the trained residual model, thereby obtaining the residual of the demand quantity of test kits relating to the second period.

[0032] The prediction unit 110 calculates the predicted value of the demand quantity for the second period as the sum of the base estimate of the demand quantity of test kits for the second period estimated by the base estimation unit 106 and the residual of the demand quantity of test kits for the second period estimated by the residual estimation unit 109. The predicted value of the demand quantity for the second period calculated by the prediction unit 110 is presented to the testing provider terminal T3 by the service providing unit 103.

[0033] FIG. 2 is a flowchart showing a method for predicting demand for test kits by the recruitment system 1 according to the first embodiment. First, the performance acquisition unit 104 calculates the values ​​of variables for each date in the first period from the information stored in the database 101, and acquires performance data, which is time-series data for the first period regarding demand for test kits (step S1). For example, the number of recent test kit applications can be obtained by counting the number of tuples recorded in the test request table 1013 whose request date falls within a specified period immediately before the target date. The number of job openings can be obtained by counting the number of tuples recorded in the job opening management table 1015 whose work date or creation date matches the target date. The number of stores can be obtained by multiplying the number of unique store IDs in the group of tuples recorded in the job opening management table 1015 whose work date or creation date matches the target date.

[0034] Next, the base learning unit 105 performs a learning process for the base model using the time series of the dates of the performance data for the first period and the number of test kit applications acquired in step S1 as a learning dataset (step S2). Next, the base estimation unit 106 inputs the dates of the first period and the second period into the base model learned in step S2 to obtain base estimates for the first period and the second period (step S3).

[0035] The residual calculation unit 107 calculates a time series of the residual of the demand quantity for the first period based on the base estimate of the demand quantity of test kits for the first period obtained in step S3 and the actual measured number of orders for test kits for the first period obtained in step S1 (step S4).

[0036] The residual learning unit 108 uses the date information and external variable information of the performance data for the first period acquired in step S1 as input samples and the residual for the first period calculated in step S4 as output samples to train a residual model using a learning dataset (step S5).

[0037] The residual estimation unit 109 generates date information and external variable information for each day of the second period (step S6). The external variable information for the second period may use the same value, assuming no change since the last day of the first period, or may use a separately estimated value. The residual estimation unit 109 calculates an estimated value of the residual for the second period by inputting the generated date information and external variable information for the second period into the residual model trained in step S5 (step S7). The prediction unit 110 calculates the sum of the base estimate of the demand quantity for the second period calculated in step S3 and the estimate of the residual for the second period calculated in step S7 as a predicted value of the demand quantity for the target day (step S8). The prediction unit 110 records the predicted value of the demand quantity for the second period in the database 101.

[0038] Thereafter, when the server 10 is accessed by the testing company terminal T3, the service providing unit 103 displays the predicted value of the demand quantity for the second period on the testing company terminal T3. The predicted value of the demand quantity for the second period may be represented, for example, by a graph with the date on the horizontal axis and the predicted value of the demand quantity on the vertical axis.

[0039] As described above, the recruitment system 1 according to the first embodiment performs the following processing. The performance acquisition unit 104 acquires performance data, which is time-series data relating to a first period, including dates and values ​​of a dependent variable on those dates. In the first embodiment, the dependent variable is the number of test kits in demand. The base estimation unit 106 uses a base model, which is a trained machine learning model, to calculate base estimation data, which is time-series data of the dependent variable relating to a second period following the first period, from the performance data. The base model includes a trend term representing fluctuations in the dependent variable during the period relating to the time-series data and a periodic term representing periodic fluctuations in the dependent variable in the time-series data. In the first embodiment, the base model is a Prophet model. The residual estimation unit 109 uses the residual model to calculate residual estimation data, which is an estimate of the residual relating to the second period. The residual model is a machine learning model trained to output a predicted value of the residual using residual data between the dependent variable value of the performance data and the base predicted value. In the first embodiment, the residual model is a LightGBM model. The prediction unit 110 calculates the sum of the base estimate and the residual estimate for the second period as the predicted value of the dependent variable for the second period.

[0040] Because the base model includes a trend term and a periodic term, it excels at capturing the characteristics of medium- to long-term changes in the dependent variable, i.e., the baseline. On the other hand, the base model has difficulty capturing nonlinear fluctuations caused by events or external factors. The residual model is trained to find the residual between the actual measured value of the dependent variable and the estimated baseline. This residual reflects nonlinear fluctuations caused by events or external factors. Therefore, according to the recruitment system 1 of the first embodiment, by summing the output of the base model and the output of the residual model, highly accurate time series prediction is possible by combining the advantages of each model.

[0041] In particular, the residual model according to the first embodiment includes the values ​​of explanatory variables related to the values ​​of the dependent variable in the input, and therefore can appropriately estimate fluctuations caused by external factors.

[0042] The recruitment system 1 according to the first embodiment predicts the future demand for test kits, allowing testing companies to appropriately manage the number of test kits in stock and quickly provide test kits to workers.

[0043] Other Embodiments Although one embodiment has been described in detail above with reference to the drawings, the specific configuration is not limited to the above, and various design modifications are possible. That is, in other embodiments, the order of the above-described processes may be changed as appropriate. Furthermore, some processes may be executed in parallel.

[0044] The server 10 according to the above-described embodiment may be configured by a single computer, or the configuration of the server 10 may be divided into multiple computers that cooperate with each other to function as the server 10. For example, the server 10 may be realized by cloud computing. Furthermore, the above-described functions may be realized by, for example, a combination of the server 10 and the inspection company terminal T3, or may be realized by an application program installed on the inspection company terminal T3.

[0045] The residual model according to the above-described embodiment is trained to input date information and external variable information and output residuals, but is not limited to this. For example, a residual model according to another embodiment may be trained to input a time series of residuals relating to a predetermined period in the past and output an estimated value of the residual following the time series of residuals. In this case, the residual model may be a time series prediction model (deep time series model) such as an RNN, LSTM, Transformer, DeepAR, or GRU.

[0046] <Computer Configuration> FIG. 3 is a schematic block diagram illustrating the configuration of a computer according to at least one embodiment. The computer 90 includes a processor 91 , a main memory 92 , a storage 93 , and an interface 94 . The above-described server 10 is implemented in a computer 90. The operations of the above-described processing units are stored in the form of a program as a prediction program in a storage 93. A processor 91 reads the program from the storage 93, loads it into a main memory 92, and executes the above-described processing in accordance with the program. The processor 91 also allocates storage areas in the main memory 92 corresponding to the above-described storage units in accordance with the program. Examples of the processor 91 include a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), and a microprocessor.

[0047] The program may be for realizing some of the functions to be performed by the computer 90. For example, the program may be combined with other programs already stored in storage or implemented in other devices to perform the functions. In another embodiment, the computer 90 may include a custom LSI (Large Scale Integrated Circuit) such as a PLD (Programmable Logic Device) in addition to or instead of the above configuration. Examples of PLDs include PAL (Programmable Array Logic), GAL (Generic Array Logic), CPLD (Complex Programmable Logic Device), and FPGA (Field Programmable Gate Array). In this case, some or all of the functions realized by the processor 91 may be realized by the integrated circuit. Such an integrated circuit is also an example of a processor. In another embodiment, the computer 90 may be virtualized on one or more computers.

[0048] Examples of storage 93 include a magnetic disk, a magneto-optical disk, an optical disk, and a semiconductor memory. Storage 93 may be an internal medium directly connected to the bus of computer 90, or an external medium connected to computer 90 via interface 94 or a communication line. Furthermore, when this program is distributed to computer 90 via a communication line, computer 90 that receives the program may load the program into main memory 92 and execute the above-described processing. In at least one embodiment, storage 93 is a non-transitory tangible storage medium.

[0049] The program may also be a program for realizing part of the above-described functions. Furthermore, the program may be a so-called differential file (differential program) that realizes the above-described functions in combination with another program already stored in storage 93. [Explanation of symbols]

[0050] 1...Recruitment system 10...Server 90...Computer 91...Processor 92...Main memory 93...Storage 94...Interface 101...Database 102...Management unit 103...Service provision unit 104...Performance acquisition unit 105...Base learning unit 106...Base estimation unit 107...Residual calculation unit 108...Residual learning unit 109...Residual estimation unit 110...Prediction unit 1011...Store table 1012...Worker table 1013...Test request table 1014...Test kit management table 1015...Recruitment management table 1016...Application management table T1...Business operator terminal T2...Store terminal T3...Testing company terminal T4...Worker terminal

Claims

1. a performance acquisition unit that acquires performance data for a specific target, the performance data being time-series data relating to a first period and including a date and a value of a response variable on the date; a base estimation unit that calculates, from the performance data, a base estimate that is an estimate of the dependent variable for a second period following the first period, using a base model that is a trained machine learning model including a trend term that represents a fluctuation of the dependent variable during a period related to time-series data and a periodic term that represents a periodic fluctuation of the dependent variable in the time-series data; a residual estimation unit that calculates an estimate of a residual for the second period using a residual model that is a machine learning model trained to output an estimate of a residual using a residual between the value of the objective variable of the performance data and the base estimate; a prediction unit that calculates the sum of the base estimate and the residual estimate for the second time period as a predicted value of the dependent variable for the second time period; A prediction device comprising:

2. a base learning unit that learns the base model using the performance data; a residual learning unit that learns the residual model using a residual between the value of the objective variable of the performance data and the base estimated value; The prediction device according to claim 1 , comprising:

3. the performance data further includes values ​​of explanatory variables related to the values ​​of the objective variables, The residual model is a machine learning model trained to use values ​​of explanatory variables of the performance data and the residuals to accept inputs of the explanatory variables and output estimated values ​​of the residuals, the residual estimation unit calculates an estimated value of the residual by inputting values ​​of explanatory variables related to the second period into the residual model. The prediction device according to claim 1 .

4. The objective variable is the demand number of test kits for health-related tests, The explanatory variables include the number of jobs or users that require testing using the test kit. The prediction device according to claim 3 .

5. The explanatory variables include the number of users who have experience with testing using the test kit. The prediction device according to claim 4 .

6. A method of calculating a time series data for a specific target, the time series data including a date and a value of a target variable on a first period; a step in which the computer calculates, from the performance data, a base estimate that is an estimate of the dependent variable for a second period following the first period, using a base model that is a trained machine learning model including a trend term that represents a fluctuation of the dependent variable over a period related to time-series data and a periodic term that represents a periodic fluctuation of the dependent variable in the time-series data; a step in which the computer calculates an estimated value of a residual for the second period using a residual model, which is a machine learning model trained to output an estimated value of a residual using the residual between the value of the dependent variable of the performance data and the base estimated value; a step of calculating, by the computer, a sum of the base estimate and the residual estimate for the second time period as a predicted value of the dependent variable for the second time period; A prediction method comprising:

7. Computer, a performance acquisition unit that acquires performance data for a specific target, the performance data being time-series data relating to a first period and including a date and a value of a response variable on the date; a base estimation unit that calculates, from the performance data, a base estimate that is an estimate of the dependent variable for a second period following the first period, using a base model that is a trained machine learning model including a trend term that represents a fluctuation of the dependent variable during a period related to time-series data and a periodic term that represents a periodic fluctuation of the dependent variable in the time-series data; a residual estimation unit that calculates an estimate of a residual for the second period using a residual model that is a machine learning model trained to output an estimate of a residual using a residual between the value of the dependent variable of the performance data and the base estimate; a prediction unit that calculates the sum of the base estimate and the residual estimate for the second time period as a predicted value of the dependent variable for the second time period; A prediction program to function as a

Images