JP2016091294A - Calculation server and method for cash balance prediction - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for predicting a cash balance in a terminal, in a case where data of a plurality of terminals is used, with consideration for a factor common to the terminals and a factor depending on the respective terminals.SOLUTION: A calculation server comprises: a data collection section for collecting transaction information from the terminals at a certain base; a learning section for calculating a prediction model including a first component common to the terminals and a second component depending on the respective terminals on the basis of the transaction information; and a prediction section for using the prediction model to predict a cash variation in the respective terminals.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a calculation server and method for predicting a cash balance in a terminal.

  Automatic cash transaction terminals such as ATMs and CDs operated at financial institutions or the like, or counter terminals for bank window tellers need to be filled with cash in advance for transactions. In these terminals, the amount of cash in the terminal always fluctuates due to the transaction. The cash in the automatic cash transaction terminal can be regarded as being procured in some way from the financial market, and the procurement cost (interest cost) is incurred. Therefore, it is desirable not to pack cash in the terminal as much as possible.

  On the other hand, when there is little cash packed in the terminal, the time when cash runs out will be advanced, and the number of times of cash transportation will increase. Thereby, the cost (cash transport cost) related to cash transport increases. Therefore, it is desired to suppress the procurement cost and the cash transportation cost by accurately predicting how much cash is required at each terminal and filling the terminal with necessary and sufficient cash according to the prediction result. There is Patent Literature 1 as a conventional method for predicting the balance of cash.

JP 2013-167936 A

  The conventional cash demand prediction method has taken the structure which estimates a cash balance from the history of the change of the past balance of a terminal independently for every terminal. From the viewpoint of cash balance prediction, the more historical data of the past cash balance, the higher the prediction accuracy. In many cases, a branch of a financial institution has a plurality of terminals, and it is desirable to predict a cash balance using data of a plurality of terminals at the same base.

  A plurality of terminals installed in a branch (base) of a financial institution or a certain region may show a similar change tendency of the cash balance. This is because the balance in the terminal may fluctuate based on a common factor among a plurality of terminals. On the other hand, even in the same base or the same region, the location where the terminal is installed is different, so it is assumed that the usage tendency will differ (for example, a terminal near the entrance of a branch office or a terminal installed near a station) Etc. tend to be used). This is because the balance in the terminal may fluctuate based on factors dependent on each terminal (for example, the position of the terminal). Conventionally, factors common to terminals and factors dependent on each terminal are not considered separately, and the accuracy of cash balance prediction has not been sufficient.

  Therefore, the present invention provides a technique for predicting a cash balance in a terminal in consideration of factors common to terminals and factors dependent on each terminal when data of a plurality of terminals is used.

  For example, in order to solve the above-mentioned problem, the configuration described in the claims is adopted. The present application includes a plurality of means for solving the above-described problem. To give an example, a data collection unit that collects transaction information from a plurality of terminals at a certain base, and the plurality of the plurality of means based on the transaction information. A learning unit that calculates a prediction model including a first component common to the terminals and a second component that depends on each terminal; and a prediction unit that predicts a cash change amount of each terminal using the prediction model. A calculation server is provided.

  According to another example, a method for predicting a cash change amount of each terminal at a certain base using a calculation server, wherein the calculation server collects transaction information from a plurality of terminals at the base. And a learning step in which the calculation server calculates a prediction model including a first component common to the plurality of terminals and a second component depending on each terminal based on the transaction information; And a prediction step of predicting a cash change amount of each terminal using the prediction model.

  According to another example, a data collection unit that collects transaction information from a plurality of terminals at each site in a certain area, and a first common to the plurality of terminals at each site based on the transaction information A calculation server comprising: a learning unit that calculates a prediction model including a component and a second component that depends on each base in the region; and a prediction unit that predicts the amount of cash change at each base using the prediction model. Provided.

ADVANTAGE OF THE INVENTION According to this invention, when using the data of a some terminal, it becomes possible to estimate a cash balance with high precision in consideration of the factor common between terminals and the factor depending on each terminal.
Further features related to the present invention will become apparent from the description of the present specification and the accompanying drawings. Further, problems, configurations and effects other than those described above will be clarified by the description of the following examples.

It is a block diagram of the cash balance prediction system in 1st Example. It is a block diagram of the calculation server in 1st Example. It is a terminal layout example of a terminal installation base. It is a functional block diagram of the learning part of the calculation server in 1st Example. It is a flowchart which shows the learning flow in 1st Example. It is a functional block diagram of the prediction part and operation plan calculation part of the calculation server in 1st Example. It is a flowchart which shows the prediction flow and operation plan calculation flow in 1st Example. It is the example which showed the transaction frequency for every terminal in time series. It is an example of the format of a prediction result. It is a flowchart which shows the learning flow in 2nd Example.

  Embodiments of the present invention will be described below with reference to the accompanying drawings. The accompanying drawings show specific embodiments in accordance with the principle of the present invention, but these are for the understanding of the present invention, and are never used to interpret the present invention in a limited manner. is not.

[First embodiment]
The following embodiment provides a function of predicting the amount of cash change of a terminal such as ATM / CD of a financial institution such as a bank. The following embodiments also provide a function for optimizing the operation plan for the amount of cash filling the terminal and the delivery timing.

  A branch (base) of a certain financial institution has a plurality of terminals. In this case, the prediction model of the cash balance of each terminal is calculated using the first component common to the plurality of terminals and the second component depending on each terminal. Said 1st component contains the element regarding the total visitor number of a certain base, and the element regarding the predicted value of the cash variation | change_quantity for every transaction. Said 2nd component contains the element regarding the ratio (transaction frequency) that a certain visitor uses a corresponding terminal. The cash balance of each future terminal is predicted using the prediction model.

  In the following examples, the product of three components of the total number of visitors at a certain base, the rate at which a certain customer uses the corresponding terminal (transaction frequency), and the predicted value of the amount of cash change for each transaction, Calculate the predicted value of the terminal cash balance. The total number of visitors is defined as the sum of transactions of all terminals existing in the branch. The amount of change in cash for each transaction can be considered as a factor determined without depending on the terminal. Therefore, by using the past histories of a plurality of terminals, prediction can be performed with higher accuracy than in the conventional configuration (a technique for predicting independently for each terminal). On the other hand, the transaction frequency can be considered as a factor determined depending on each terminal. Therefore, it is desired to estimate for each terminal.

FIG. 1 is a configuration diagram of a cash balance prediction system in the first embodiment. In this example, a plurality of terminal installation bases 1 to N are provided. Each terminal installation base includes a plurality of terminals. For example, the terminal installation base 1 includes a plurality of terminals 101-1-1 to M ₁ . Here, it is assumed that each terminal is a terminal to which cash such as ATM or CD is deposited or withdrawn. The cash balance prediction system includes a calculation server 102. The calculation server 102 collects transaction information from each terminal at each base and predicts a cash balance. Here, the transaction information includes the number of transactions for each terminal time zone, balance information for each terminal time zone, and the like. The unit of the time zone may be, for example, 1 hour or 10 minutes. Also, the time zone may be divided in units of morning / afternoon. The balance information may be a change amount of the balance based on the previous time zone.

  FIG. 2 is a configuration diagram of the calculation server 102 in the first embodiment. The calculation server 102 includes an input unit 201, an output unit 202, a CPU 203, a memory 204, and a storage medium 205. The input unit 201 is a human interface such as a mouse and a keyboard, and receives an input to the calculation server 102. The output unit 202 is a display or printer that outputs a calculation result by the calculation server 102.

  The storage medium 205 is a storage device that stores programs for realizing various processes by the calculation server 102, execution results of the various processes, and the like, and is, for example, a non-volatile storage medium (magnetic disk drive, non-volatile memory, or the like). .

  A program stored in the storage medium 205 is expanded in the memory 204. The CPU 203 is an arithmetic device that executes a program loaded in the memory 204. The CPU 203 executes the processes and operations described below.

  The calculation server 102 of the present embodiment may be configured by a single computer or a plurality of computers configured logically or physically. Various processes of the calculation server 102 may operate on separate threads on the same computer, or may operate on virtual computers constructed on a plurality of physical computer resources.

  The program executed by the CPU 203 may be provided to the calculation server 102 via a removable medium (CD-ROM, flash memory, etc.) or a network. Therefore, the calculation server 102 may include an interface for reading a removable medium.

  The calculation server 102 includes a database 206. The database 206 includes a transaction information storage unit 221, a learning parameter storage unit 222, and an operation constraint information storage unit 223.

  The storage medium 205 includes a data collection unit 211, a learning unit 212, a prediction unit 213, and an operation plan calculation unit 214 as various programs. The data collection unit 211 collects transaction information for each time zone from each terminal at each terminal installation base, and stores the collected transaction information in the transaction information storage unit 221. The data collection unit 211 collects transaction information together with identification information (terminal ID) for identifying a terminal. Further, the data collection unit 211 may collect a transaction type (payment / withdrawal) and a transaction amount as transaction information. Further, the data collection unit 211 may collect terminal location information together with the terminal ID. Information that associates each terminal ID with location information may be stored in advance on the calculation server 102 side. In this case, the data collection unit 211 may not collect location information.

  The learning unit 212 performs learning using past transaction information (transaction information up to the present time) stored in the transaction information storage unit 221. Here, “learning” means calculating a prediction model to be described later. The learning unit 212 stores the calculated parameters of the prediction model in the learning parameter storage unit 222.

  The prediction unit 213 calculates the future balance amount of each terminal using the prediction model calculated by the learning unit 212. The predicted balance amount may be related to an arbitrary time, and may be, for example, the cash balance amount on the next day of each terminal or the cash balance amount after one week of each terminal.

  The operation plan calculation unit 214 calculates an operation plan from the future balance amount of each terminal calculated by the prediction unit 213. Here, “operation plan” means a cash filling amount and a delivery timing for each terminal. The amount of cash filling is, for example, the number of cash to be packed for each cassette in the terminal and the type of cash. The delivery timing means the delivery timing to the terminal at each base. For example, the delivery timing is information indicating whether to deliver the next day or when to go within one week.

  The operation plan calculation unit 214 calculates at least one of a cash transport cost and a procurement cost. Furthermore, the operation plan calculation unit 214 corrects the operation plan in consideration of the constraint information stored in advance in the operation constraint information storage unit 223 and the above-described cost, and outputs the corrected operation plan. Here, the “constraint information” is information that becomes a constraint when determining an operation plan, and is, for example, constraint information regarding a cash filling amount and a delivery timing for each terminal. For example, the constraint information is information on a minimum amount of cash filling that should be maintained at each terminal, information such as delivery at least once a month.

  FIG. 3 is a terminal layout example of a certain terminal installation base. The plurality of terminals 301-1 to 301-6 are installed at different distances from the entrance / exit, respectively. Therefore, in a quiet time zone, it is assumed that a terminal (for example, terminal 301-5) at a short distance from the entrance is frequently used. On the other hand, during the busy season, it is assumed that all terminals are used at the same frequency. Thus, the ease of use of the terminal varies depending on the installation position of the terminal, and the usage tendency of each terminal also varies depending on the congestion situation. In this embodiment, paying attention to this property, the ease of use of the terminal is estimated as a variable that varies depending on the degree of congestion of the entire terminal installation base.

  FIG. 4 is a functional block diagram of the learning unit 212 of the calculation server 102. The learning unit 212 includes a learning block 401. The learning block 401 receives past transaction information stored in the transaction information storage unit 221 and the latest transaction information collected by the data collection unit 211, and learns various parameters of the prediction model using these transaction information. To do. The learning block 401 stores the parameters of the prediction model in the learning parameter storage unit 222.

  FIG. 5 is a flowchart showing a learning flow in the learning block 401. In the following description, the functional block of FIG. 4 will be described as the subject, but the program is executed by the CPU 203 and the processing determined by the CPU 203 is performed using the memory and the communication port (communication control device). The explanation may be as follows.

  The learning block 401 starts processing (501). Next, the learning block 401 calculates the total number of visitors for each site based on the input transaction information for each terminal (502). Here, the “total number of visitors” is the total number of transactions for each time zone at each base. The learning block 401 calculates the total number of transactions at each base for each time zone. Therefore, the calculated total number of visitors has one value for each base and every time.

  Next, the learning block 401 estimates a mapping function between the number of transactions for each terminal and the total number of visitors (503). First, the learning block 401 calculates the number of transactions for each terminal for each time zone. In addition, a mapping function with the total number of visitors for each time zone and each base already calculated is defined as follows, for example.

(Equation 1)
Number of transactions in a certain time zone per terminal = Total number of visitors in a certain time zone per base × G

  Here, G may be a ratio of the number of transactions per terminal to the total number of visitors per base, or may be defined as some function that does not depend on time. The learning block 401 may determine G from the number of transactions for each terminal and the total number of visitors for each base, for example, by the least square method.

  FIG. 8 is an example showing the number of transactions for each terminal in time series. As shown in FIG. 8, when the total number of visitors at a certain base is small, it is easy to use the terminal near the entrance of the base, whereas almost all terminals are used during the busy season. . That is, in a busy period, G is assumed to approach a value that does not depend on each terminal. Therefore, a configuration may be adopted in which G is estimated as a function G (total number of visitors) that depends on the total number of visitors. In other words, G may be defined as a function that depends on time. In this example, G uses the total number of visitors that changes over time as a variable.

  In this case, the learning block 401 may set G using a linear model, a neural network model, a kernel function, or the like. In the case of a linear model, G may be estimated using the least square method. In the case of a sigmoid function model, G may be estimated using a neural network model. In the case of a kernel function, G may be estimated using support vector regression or the like.

  Next, the learning block 401 calculates a parameter of the prediction formula for the total number of visitors (504). For example, the prediction formula is defined as follows.

(Equation 2)
Total number of visitors (t + D, A base) = f (total number of visitors (t, A base), total number of visitors (t-1, A base), ...)

  According to the above formula, it is possible to predict the total number of visitors D units time ahead of the time zone t at the A site by inputting the total number of visitors at the A site up to the t time zone in the function f. it can. f is a prediction function for calculating a predicted value from an explanatory variable. The prediction function f is defined using, for example, a linear model, a sigmoid function model, a kernel function, or the like.

  Equation 2 is an autoregressive model of the total number of visitors. For example, the total number of visitors depends on the weather ahead of the D unit time, the day of the week before the D unit time, or whether it is early, mid or late. The number is expected to change. In consideration of this, a configuration may be adopted in which explanatory variables corresponding to the above items are increased in the prediction function f. Hereinafter, factors such as weather and day of the week are referred to as “external factor information”.

  The learning of the prediction function f may be performed using the least square method if it is a linear model. In the case of a sigmoid function model, learning of the prediction function f may be performed using a neural network model. In the case of a kernel function, learning of the prediction function f may be performed using support vector regression or the like.

Next, the learning block 401 estimates the amount of cash change for each transaction (505). For each time period, the sum (A ₁ ) of the amount of cash change for every transaction of all terminals in the A base is taken. In addition, the sum (B ₁ ) of the number of transactions of all terminals in the base for each time zone is taken. As an example, the learning block 401 calculates a value obtained by dividing the A ₁ by B ₁ as cash change amount for each transaction.

  Next, the learning block 401 calculates the parameter of the prediction formula of the cash change amount for each transaction (506). For example, the prediction formula is defined as follows.

(Equation 3)
Cash change amount for each transaction (t + D, A base) = h (day of t + D, weather at t + D, whether t + D is a bonus payment date, etc.)

  According to the above formula, the amount of cash change for each transaction in the D unit time ahead of the time zone t at the A base can be predicted by inputting various information at the time point t + D to the function h. h is a prediction function for calculating a predicted value from an explanatory variable. The prediction function h is defined using, for example, a linear model, a sigmoid function model, a kernel function, or the like. Since the amount of cash change for each transaction is determined by how much cash the customer needs, it is estimated by external factor information that affects the cash demand such as day of the week, weather, and bonus payment date. Note that explanatory variables other than those described above may be added with respect to the prediction function h.

  The learning of the prediction function h may be performed using the least square method if it is a linear model. In the case of a sigmoid function model, learning of the prediction function h may be performed using a neural network model. In the case of a kernel function, learning of the prediction function h may be performed using support vector regression or the like.

  After estimating the prediction functions G, f, and h, the learning block 401 stores the parameters for determining the shapes of the prediction functions G, f, and h as parameters of the prediction model in the learning parameter storage unit 222, and ends the processing ( 507).

  FIG. 6 is a functional block diagram of the prediction unit 213 and the operation plan calculation unit 214 of the calculation server 102. FIG. 7 is a flowchart showing a prediction flow and an operation plan calculation flow in the configuration of FIG.

  The prediction block 601 of the prediction unit 213 starts processing (701). The calculation of the predicted amount of the cash balance may be started at an arbitrary timing. For example, the processing may be started via an input from the input unit 201. The calculation of the predicted amount of the cash balance may be performed for each day of the week, or may be performed for each time zone of each day of the week.

  The prediction block 601 receives transaction information from the data collection unit 211 and the transaction information storage unit 221 as input. For example, the prediction block 601 calculates the predicted value of the cash balance demand of each terminal at the A base (702). The predicted value here is, for example, the cash balance of each terminal on the next day, the cash balance of each day until one week later, or the cash balance for each time zone of each day of the week. Suppose.

  The prediction block 601 uses the parameters of the prediction functions G, f, and h stored in the learning parameter storage unit 222 to calculate the amount of change in cash at the D unit time ahead of each terminal as follows.

(Equation 4)
Cash change amount (t + D, terminal with A base) = G x h x f

  Here, various explanatory variables are required to calculate the values of the prediction functions G, h, and f. For example, if the explanatory variable is the weather at time t + D, the prediction block 601 may acquire the weather forecast at time t + D from an external resource (not shown) via the network. The prediction block 601 acquires the value of the explanatory variable using an external resource or the like. Note that when the prediction function G is modeled so as to depend on the total number of visitors, the parameter estimated by the prediction function f may be used as the parameter of the total number of visitors given to the prediction function G.

  FIG. 9 is an example in which the cash change amount of each terminal is calculated for each unit time. Here, the unit time zone D is one day. In the format of FIG. 9, for each terminal, the balance change amount up to one week later is output as a difference value from the current cash balance of each terminal. The numerical value in FIG. 9 may be, for example, the number of a certain banknote or may be arbitrarily set such as an actual amount. The prediction block 601 outputs the format of FIG. 9 to the operation plan initialization block 602 of the operation plan calculation unit 214 as output data.

  Next, the operation plan initialization block 602 calculates an initial plan of the cash operation plan (703). The operation plan initialization block 602 calculates the initial plan of the cash operation plan on the basis of, for example, minimizing the number of times of cash transport within a range that does not cause a cash out using the format of FIG. This cash management plan includes information on when and how much cash is to be packed and collected in which terminal. The operation plan initialization block 602 outputs information on the initial plan of the cash operation plan to the cost simulator 603.

  Next, the cost simulator 603 calculates the cost using the initial plan information of the cash management plan (704). For example, the cost simulator 603 calculates the cash transportation cost using information on the initial plan of the cash operation plan. The cost simulator 603 may calculate the procurement cost.

  Next, in order to minimize the calculated cost, the operation plan correction block 604 corrects the cash operation plan (705). Specifically, the operation plan correction block 604 uses the quasi-Newton method to determine which terminal when the cost value calculated by the cost simulator 603 (at least one of the cash transportation cost and the procurement cost). Change how much cash you pack and collect. In the quasi-Newton method, parameters can be optimized as long as there is a mapping function between parameters and cost values. Here, the parameter is information on which terminal, when and how much cash is to be packed and collected. In this configuration, the mapping function is the cost simulator 603. The operation plan optimization is not limited to the quasi-Newton method, and other known optimization methods such as mixed integer programming may be used.

  At this time, the operation plan correction block 604 corrects the cash operation plan using the information in the operation constraint information storage unit 223. The operation constraint information storage unit 223 stores in advance information that becomes a constraint when determining an operation plan. For example, if there is constraint information such as the need to deliver at least once a month for each terminal, optimize the cash management plan by the constrained quasi-Newton method with the constraint of delivering once a month Is possible. In the case of using mixed integer programming, the operation plan correction block 604 may correct the cash operation plan by the constrained mixed integer programming. The operation plan correction block 604 outputs the corrected operation plan to the correction end determination block 605.

  Next, the correction end determination block 605 determines whether to end the correction of the operation plan (706). When the modification of the operation plan is to be ended, the modification end determination block 605 outputs the corrected operation plan to the output unit 202 (707) and ends the processing (708). On the other hand, when the correction of the operation plan is continued, the process returns to step 705.

  The determination in the correction end determination block 605 may be performed based on, for example, whether the number of corrections in the operation plan correction block 604 has been made a predetermined number or more. Further, the determination may be performed based on whether the parameter correction amount in the quasi-Newton method is less than a predetermined amount (that is, the parameter correction amount has converged).

  According to the above embodiment, the transaction frequency is estimated for each terminal as a factor depending on the terminal, and the predicted value of the fluctuation amount of the cash balance for each transaction and the total number of visitors are estimated as a factor independent of the terminal. By taking the configuration, it becomes possible to estimate the cash balance by efficiently using data of a plurality of terminals.

  In cash balance prediction, using information on the same base terminal that shows the same trend of balance balance will improve the prediction accuracy, but even if it is the same base, the location where the terminal is installed is different, so for example near the entrance Terminals are likely to be used differently, such as being easy to use. However, conventionally, even when data of a plurality of terminals are used, the cash balance is only predicted independently for each terminal. On the other hand, in the present embodiment, it is possible to estimate the cash balance in consideration of the difference in usage tendency of each of the plurality of terminals. In the present embodiment, by utilizing data of a plurality of terminals at the same base, it becomes possible to predict a cash balance with higher accuracy than a configuration in which a cash balance is predicted independently for each terminal. Therefore, procurement cost and cash transportation cost could not be suppressed well.

  In the above example, the cash change amount is predicted for each terminal using the terminal ID, but the terminal position information may be used. In this case, the prediction unit 213 can calculate how much cash the terminal at which position in the terminal installation base has.

  In the above-described example, the prediction is made as the amount of change in cash handled as a combination of deposits and withdrawals. However, deposits and withdrawals may be predicted separately. In this case, the data collection unit 211 collects a transaction type (payment / withdrawal) and a transaction amount as transaction information. Using these pieces of information, the learning unit 212 may create a prediction model for each type of transaction. Thereby, the prediction unit 213 can individually predict the deposit for each terminal in a certain time zone at a certain base and the withdrawal for each terminal in a certain time zone at a certain base.

[Second Embodiment]
In the following embodiments, an example will be described in which cash balance information is predicted using information for the entire region. In this example, the prediction model further includes a third component common to a plurality of bases in a certain region and a fourth component that depends on each base in the region. FIG. 10 is a flowchart showing a learning flow of the learning unit 212 in the second embodiment.

  The learning block 401 starts processing (1001). Next, the learning block 401 calculates the total number of visitors in the entire area based on the inputted transaction information for each terminal (1002). Here, the “total number of visitors” is the total number of transactions for each time zone of the entire region (hereinafter, described as region A). The learning block 401 calculates the total number of transactions at each site for each time period, and calculates the total number of transactions at all sites belonging to the A area. Thereby, the transaction total number of A area for every time slot | zone can be obtained.

  Next, the learning block 401 estimates a mapping function between the number of transactions for each base and the total number of visitors in the entire area (A area) (1003). A mapping function with the total number of visitors for each time zone and each region that has already been calculated is defined as follows, for example.

(Equation 5)
Total number of visitors in a certain time zone at a certain base = Total number of visitors in a certain time zone of the entire region (A region) x J

  Here, J may be a ratio of the total number of visitors at each base to the total number of visitors in the entire region, or may be defined as some function that does not depend on time. The learning block 401 may never use J from, for example, the least square method based on the total number of visitors for each base and the total number of visitors in the entire region.

  In addition, the frequency of use is assumed to change from base to base depending on whether it is busy or quiet. J in the off season depends on the location (for example, it is assumed that the number of visitors fluctuates due to convenience, such as close to the station in the off season), while J in the busy season does not depend on the location. It is assumed that Therefore, a configuration may be adopted in which J is estimated as a function J (total number of visitors in the region) that depends on the total number of visitors in the region. In this case, the learning block 401 may set J using a linear model, a neural network model, a kernel function, or the like. In the case of a linear model, J may be estimated using the least square method. In the case of a sigmoid function model, J may be estimated using a neural network model. In the case of a kernel function, J may be estimated using support vector regression or the like.

  Next, the learning block 401 calculates a parameter of a prediction formula for the total number of visitors in the entire region (1004). For example, the prediction formula is defined as follows.

(Equation 6)
Total number of visitors in the whole area (t + D, A area) = f (total number of visitors in the whole area (t, A area), total number of visitors in the whole area (t-1, A area), ...)

  According to the above formula, it is possible to predict the total number of visitors D units time ahead of the time zone t in the A region by inputting the total number of visitors in the A region up to the time t in the function f. it can. f is a prediction function for calculating a predicted value from an explanatory variable. The prediction function f is defined using, for example, a linear model, a sigmoid function model, a kernel function, or the like.

  Equation 6 is an autoregressive model of the total number of visitors. For example, the total number of visitors depends on the weather ahead of D unit time, the day of the week ahead of D unit time, or whether it is early, mid or late. The number of visitors is expected to change. Considering such external factor information, a configuration may be adopted in which explanatory variables corresponding to the above items are increased in the prediction function f.

  The learning of the prediction function f may be performed using the least square method if it is a linear model. In the case of a sigmoid function model, learning of the prediction function f may be performed using a neural network model. In the case of a kernel function, learning of the prediction function f may be performed using support vector regression or the like.

  Next, the learning block 401 estimates the mapping function G between the number of transactions for each terminal and the total number of visitors at the base by the same method as the method shown in the first embodiment (1005). Similarly, the learning block 401 calculates the cash change amount for each transaction by the same method as that shown in the first embodiment (1006). Further, the learning block 401 performs parameter estimation of the cash change amount prediction function h for each transaction by the same method as shown in the first embodiment (1007). The learning block 401 stores the parameters for determining the shapes of the prediction functions J, f, G, and h as parameters of the prediction model in the learning parameter storage unit 222, and ends the processing (1008).

  Next, processing in the prediction block 601 of the present embodiment will be described. The prediction block 601 uses the parameters of the prediction functions J, f, G, and h stored in the learning parameter storage unit 222 to calculate the amount of change in cash D units ahead as follows.

(Equation 7)
Cash change (t + D, A area) = J x G x h x f

  About the process which creates an operation plan based on the amount of cash changes of this D unit time ahead, it performs with the method same as the method shown in 1st Example.

  In addition, it is assumed that the ATM cash balance in each base shows the same behavior, and the value of G may be set to an equal value for all ATMs in the base.

  Moreover, although the present Example demonstrated the example which estimates the cash variation | change_quantity of each terminal of each base from the total visitor number of the whole area, it is limited to the combination of 1st Example and 2nd Example in this way. Not. For example, the cash change amount for each base in the A area may be calculated without considering the cash change amount for each terminal. In this case, without using G, for example, the cash change amount of each base in the A area may be defined by the following expression.

(Equation 8)
Cash change amount (t + D) = J x h x f

  According to the above embodiment, the amount of cash change for each terminal at each base can be estimated from the total number of visitors in the entire region. In addition, the amount of cash change at each base can be estimated from the total number of visitors in the entire region.

  The present invention is not limited to the above-described embodiments, and includes various modifications. The above embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Also, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment. Moreover, the structure of another Example can also be added to the structure of a certain Example. Further, with respect to a part of the configuration of each embodiment, another configuration can be added, deleted, or replaced.

  The processing described above can also be realized by a program code of software that realizes these functions. In this case, a storage medium in which the program code is recorded is provided to the system or apparatus, and the computer (or CPU or MPU) of the system or apparatus reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiments, and the program code itself and the storage medium storing it constitute the present invention. As a storage medium for supplying such program code, for example, a flexible disk, CD-ROM, DVD-ROM, hard disk, optical disk, magneto-optical disk, CD-R, magnetic tape, nonvolatile memory card, ROM Etc. are used.

  The processes and techniques described herein are not inherently related to any particular device, and can be implemented by any suitable combination of components. In addition, various types of devices for general purpose can be used. It may be beneficial to build a dedicated device to perform the method steps described herein. That is, part or all of the above-described processing may be realized by hardware using electronic components such as an integrated circuit.

  Furthermore, in the above-described embodiments, control lines and information lines are those that are considered necessary for explanation, and not all control lines and information lines on the product are necessarily shown. All the components may be connected to each other.

101-1-1, ···, 101-N- M N: terminal 102: computation server 201: Input section 202: Output section 203: CPU
204: Memory 205: Storage medium 206: Database 211: Data collection unit 212: Learning unit 213: Prediction unit 214: Operation plan calculation unit 221: Transaction information storage unit 222: Learning parameter storage unit 223: Operation constraint information storage unit 401: Learning block 601: Prediction block 602: Operation plan initialization block 603: Cost simulator 604: Operation plan correction block 605: Correction end determination block

Claims (15)

A data collection unit that collects transaction information from multiple devices at a location;
Based on the transaction information, a learning unit that calculates a prediction model including a first component common to the plurality of terminals and a second component depending on each terminal;
A prediction unit that predicts the amount of cash change of each terminal using the prediction model;
A calculation server comprising: In the calculation server according to claim 1,
The first component includes a first element relating to the total number of visitors at the base and a second element relating to prediction of a cash change amount for each transaction, and the second component is a certain visitor. A calculation server comprising a third element relating to a ratio of using a terminal. In the calculation server according to claim 2,
The learning unit
As the first element, a first prediction function for predicting the total number of visitors in the time zone where the base is located is calculated.
As the second element, calculate a second prediction function for predicting the amount of cash change for each transaction in the time zone where the base is located,
The calculation server characterized in that, as the third element, a mapping function is calculated between the total number of visitors in the time zone where the base is located and the number of transactions for each terminal. In the calculation server according to claim 3,
The calculation server, wherein the first prediction function and the second prediction function include explanatory variables of external factor information. In the calculation server according to claim 1,
An operation plan calculation unit that calculates an operation plan based on the predicted cash change amount of each terminal,
The said operation plan calculation part corrects the said operation plan based on at least 1 cost of a cash transport cost and a procurement cost, The calculation server characterized by the above-mentioned. In the calculation server according to claim 5,
The said operation plan calculation part corrects the said operation plan using the constraint information registered previously, The calculation server characterized by the above-mentioned. In the calculation server according to claim 1,
The calculation server, wherein the prediction model further includes a third component common to a plurality of bases in a certain region and a fourth component depending on each base in the region. In the calculation server according to claim 1,
The prediction unit predicts deposit and withdrawal individually as the cash change amount. A method for predicting the amount of cash change of each terminal at a certain base using a calculation server,
A data collection step in which the calculation server collects transaction information from a plurality of terminals at the base;
A learning step in which the calculation server calculates a prediction model including a first component common to the plurality of terminals and a second component depending on each terminal based on the transaction information;
A prediction step in which the calculation server predicts a cash change amount of each terminal using the prediction model;
A method comprising the steps of: The method of claim 9, wherein
The first component includes a first element relating to the total number of visitors at the base and a second element relating to prediction of a cash change amount for each transaction, and the second component is a certain visitor. A method comprising a third element relating to the rate of using a terminal. The method of claim 10, wherein
The learning step includes
Calculating a first prediction function for predicting the total number of visitors in the time zone where the base is located as the first element;
Calculating a second prediction function for predicting the amount of cash change for each transaction in the time zone in which the base is located as the second element;
Calculating a mapping function between the total number of visitors in the time zone in which the base is located and the number of transactions for each terminal as the third element;
A method comprising the steps of: The method of claim 9, wherein
An operation plan calculation step of calculating an operation plan based on the predicted amount of cash change of each terminal;
The operation plan calculation step includes modifying the operation plan based on at least one of a cash transportation cost and a procurement cost. The method of claim 12, wherein
The operation plan calculation step includes modifying the operation plan using constraint information registered in advance. A data collector that collects transaction information from multiple devices at each location in a region,
Based on the transaction information, a learning unit that calculates a prediction model including a first component common to the plurality of terminals at each site and a second component that depends on each site in the region;
A prediction unit that predicts the amount of cash change at each site using the prediction model;
A calculation server comprising: In the calculation server according to claim 14,
The first component includes a first factor relating to the total number of visitors in the region and a second factor relating to a prediction of a cash change amount for each transaction, and the second component is a total amount of the entire region. A calculation server comprising a third element relating to the ratio of the total number of visitors at each site to the number of visitors.

Images