Patent Application
20180260744
The present invention relates to an online forecasting system and an online forecasting method.
An online forecasting system in which a series of flows that new data is input every day, a value of a certain index is forecasted based on the new data and past result data previously stored, a certain optimization or decision making is performed based on the forecasted value, and while processing or observation is actually performed to obtain a result value of the index thereafter, and the obtained result value is stored for future forecasting are repeated is common. A system in which matters that new data is received at each point in time and certain processing is performed for data at that point in time are repeated is usually called an online system, compared to a system in which all data are given at one time and processing is performed at one time.
For example, in a retail company having developed chain stores, commodities purchased in bulk from a maker are generally stored in a physical distribution warehouse, a list of slips in which commodities to be delivered to each store and the number of commodities are written is received every morning, and commodities are delivered to each store as needed. Also, the quantity of commodities delivered in one day is recorded as the result data and the quantity of commodities to be delivered on the day is forecasted based on past result data.
As the background art of the technical field of the present invention, there is PTL 1. In PTL 1, a forecasting model of the index value in which the newest processing result at that point in time is reflected is prepared each time when a new input is obtained. With this, PTL 1 has a configuration in which forecasting following change is performed even in a case where a relationship between the input and the index changes with time.
In the online forecasting system, a tendency of a processing target changes slowly and also an accidentally caused great change in the tendency is not uncommon.
For example, in the physical distribution warehouse, a tendency of received slips at a special day such as in an opening period of a large-scaled event or during long holidays is largely different from a normal day. As such, in a case where there is a special day which is composed of peculiar data and has a relationship between an input x and y which is largely different from other days among the past N days, generally, it is considered that forecasting with high accuracy is possible by preparing a forecasting model using data of remaining days except for the special day. In the example of the physical distribution warehouse, it is considered that forecasting accuracy in forecasting a picking time of a normal day is improved by excluding data of such a special day.
As another example, the relationship between the input x and y may be largely different from a previous one bordering a certain specific day in the latter half of the past N days. In this case, it is considered that forecasting with high accuracy is possible by using only newest data except for data before the specific day, in preparing for a forecasting model of today. In the example of the physical distribution warehouse, matters that a temperature is abruptly raised after the rainy season passes and Christmas business began are considered factors.
In the online forecasting system described in PTL 1, all pieces of result data for the past N days are handled equally to construct the forecasting model of the index and forecasting of the index value of today from input data of today is performed based on the constructed forecasting model. Accordingly, there is a problem to be solved that it is difficult to follow abrupt change in the tendency.
An object of the present invention is to provide an online forecasting system capable of accurately performing forecasting of an index value even in a case where there is abrupt change in a tendency among past result data.
In order to achieve the object described above, the present invention adopts a configuration in which input data of a predetermined day is received, result data of an index value of a day antedating the predetermined day is stored, similarity between the input data of the predetermined day and each of the stored result data is calculated, a weight for each of the result data based on the calculated similarity is calculated, and the index value of the predetermined day is forecasted based on the calculated weight, the result data, and the input data of the predetermined day.
According to the present invention, influence by data of a special day or influence before the specific day is made small to thereby make it possible to improve forecasting accuracy of the index value.
In the following, embodiments of the present invention will be described in detail using the drawings. In respective drawings, common configurations are assigned the same reference numerals as a matter of principle and repetitive explanations will be omitted. In the following Examples, an example of a physical distribution warehouse of a retail company having developed chain stores is used.
In the retail company having developed chain stores, commodities purchased in bulk from a maker are generally stored in a physical distribution warehouse and commodities are delivered to each store as needed. In the physical distribution warehouse, a list of slips in which commodities to be delivered to each store and the number of commodities are written is received every morning. When the list of slips is received, a physical distribution warehouse worker picks up the commodities according to commodity names and the number of shipments described in the slips while going around commodity storage shelves within the physical distribution warehouse, packs all commodities described in a single slip together, and ships the commodities to each store. A delivery deadline time is attached to each slip and shipment processing needs to be completed until that time. In the physical distribution warehouse, matters that man hours required for picking work for the slips of that day, that is, so to speak, total man hours are estimated and the number of the physical distribution warehouse workers needed to end the picking work within a required time is secured, or order of work is changed, based on the past experience, that is, recorded result data, at the point in time of receiving the list of slips to be subjected to shipment processing on that day in the morning are performed. Each time when actual picking work is ended, actual time required for picking processing is recorded and is used for future forecasting.
The past index value's result database PASTDB stores inputs by the past N days, that is, a list of slips x(t−N), x(t−N+1), . . . , x(t−1) and the index values by the past N days, that is, result values of a picking work time y(t−N), y(t−N+1), . . . , y(t−N), when N is a positive integer, as a set.
In the online forecasting system of Example 1, when the index value forecasting model preparation unit FORECAST prepares a forecasting model of the index value, the forecasting model of the index value is prepared by applying weights w(t−N, t), w(t−N+1, t), . . . , w(t−N, t) on data of respective days based on the result data x(t−N), x(t−N+1), . . . , x(t−1) for the past N days and y(t−N), y(t−N+1), . . . , y(t−N) stored in the past index value's result data storing unit PASTDB.
Here, the weight w(t1,t2) is a numerical value indicating how much a weight is applied to be taken into account for result data of the index value of Day(t−1) when the index value forecasting model of Day(t−2) is prepared. With this, when the forecasting model of the index value of today (Day(t)) is prepared, in a case where only Day(t−k) is composed of peculiar data and has the relationship between the input x(t−k) and the y(t−k) which is largely different from other days, for example, for a certain integer k of 1≤k≤N, among result data of the index values of the past N days, matters that the weight w(t−k, t) of past data of Day(t−k) is made smaller to prepare a forecasting model in which influence by peculiar data of Day(t−k) is made small become possible.
In a case where only data of Day(t−k) is different from other days, matters that the weight w(t−k, t) is made small is not necessarily optimum. For example, the input x(t) of today (Day(t)) is similar to only the input x(t−k) of the special Day(t−k) and is not similar to inputs of other days, inversely, the weight w(t−k, t) to be applied to past data of Day(t−k) is made large so as to make it possible to improve forecasting accuracy of the index value. Accordingly, it is considered that the weights w(t−N, t), w(t−N+1, t), . . . , w(t−N, t) for the result data of the past N days used when the forecasting model of today is prepared are preferably applied by being depended on how much the input x(t) of today is similar to the input x of each of the past days.
In the following, operations of the online forecasting system of Example 1 will be described in order. The online forecasting system receives an input of data x(t) to be processed today (Day(t)) every morning. In the example of the physical distribution warehouse, the list of slips (list of slips to be processed today) to be received every morning corresponds to data to be received. Next, the input similarity calculation unit SCALC compares the input x(t) of today with each of inputs x(t−N), x(t−N+1), . . . , x(t−1) for the past N days stored in the past index value's result data storing unit PASTDB to calculate similarity and outputs weight vector w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) composed of N numerical values.
As a calculation of similarity, for example, a method of calculating feature quantity vectors from respective inputs x and regarding the reciprocal of a distance between the feature quantity vectors as similarity may be considered. A feature quantity used for this case may be set in advance, or can be obtained by a configuration for automatically extracting from examples of multiple inputs x by a known “unsupervised machine learning method” such as a self-organization mapping or a deep neural network.
The weight multiplication unit WMUL multiplies the result value y(t−j) of the index of Day(t−j) stored in the past index value's result data storing unit PASTDB by the weight w(t−j, t) calculated by the input similarity calculation unit SCALC and outputs the result to the index value forecasting model preparation unit FORECAST.
Next, the index value forecasting model preparation unit FORECAST prepares a forecasting model used for forecasting the index value of today, that is, an index value forecasting model from the result data of the past index value stored in the past index value's result data storing unit PASTDB. In this case, the result data of Day(t−j) is prepared by applying the weight w(t−j t) thereto. For this, a method of obtaining a regression model for, for example, the forecasting model used for forecasting the index value of today, by a weighted least square method by using the input x(t−j), the result value of the index y(t−j), and the weight w(t−j, t) may be used. Specifically, a regression coefficient vector β which minimizes a weighted least square error S expressed in the following Equation 1 may be obtained.
In Expression 1, the center point indicates an inner product between the vectors. The input x and the index value y may not be a single numerical value but a vector composed of a plurality of numerical values. In Expression 1, a case where the input x(t−j) is a vector is described. In this case, a dimension of the regression coefficient vector β is the same as that of the input x. Otherwise, as another method of preparing the forecasting model, a configuration in which learning is performed by applying a weight to supervisor data in a known “supervised machine learning method” called a neural network and Boltzman machine may be available.
As described above, a configuration in which the index value forecasting model preparation unit FORECAST operates as a learning device having an internal state and learns result data of a newly added index value online every day may be considered, instead that a forecasting model of the index of that day is newly prepared using the result data of the past index every day. As the online learning device, a known “supervised online machine learning method” called an online neural network or online Boltzman machine can be used.
Next, the index value forecasting model preparation unit FORECAST performs forecasting of the index value y(t) of today from the input x(t) of today based on the prepared forecasting model and calculates a forecast value y*(t). In the example of the physical distribution warehouse, this correspond to matters that total working hours of the picking work required for the processing from the list of slips to be processed today are estimated.
Next, the measure decision unit DECISION performs some kind of measure decision based on the calculated forecast value y*(t) of the index. In the example of the physical distribution warehouse, this correspond to matters that the number of workers needed to end work within the request time is calculated. The number of needed workers is calculated based on, for example, a standard work time per a single worker.
Thereafter, the processing execution/observation unit PROC actually executes processing for a target system or observes the target system to acquire a result value y(t) of the index of today. In the example of the physical distribution warehouse, this corresponds to matters that the total picking work time is measured in a case where the worker actually performs the picking work for each slip. When the result value y(t) of the index of today is acquired, the input x(t) of today and the result value y(t) of the index are stored in the past index value's result data storing unit PASTDB as a set. With this, the set of the input x(t) of today and the result value y(t) of the index is taken into account when the index value forecasting model at or after tomorrow (Day(t+1)) is prepared.
With this, the newest result value of the index is reflected in preparation for the index value forecasting model at all times and it becomes possible to timely follow change in the tendency of the result value of the index. In the present Example, only the result data of the past N days is taken into account in preparing for the index value forecasting model, and thus, data of a day before the N day may be appropriately deleted.
In the above-description, a configuration in which the result data of the past N days (N is an integer determined in advance) is taken into account and data before the N day is not taken into account in preparing for the index value forecasting model is adopted, a configuration in which the result data before the N day is taken into account, that is, all pieces of past data stored in the past index value's result data storing unit PASTDB are used in preparation for the index value forecasting model may be considered. In this case, a configuration in which the weight w(t−j, t) for past data which is older is made small may be adopted and, for example, a weight w is made smaller at a fixed rate as data becomes older. This is based on an idea that generally, data of a day closer to today tends to closer to data of today (being similar to data of today).
Otherwise, matters that control called “the older the data, the smaller the weight w” is not performed may be considered. This is because, typically, as the input x(t−j) in the past becomes older, a difference between the input x(t−j) and the input x(t) of today becomes large (similarity is lowered) and thus, the weight w(t−j, t) for past data which is older tends to become naturally smaller value even without controlling the weight w.
As described above, according to the online forecasting system in Example 1, a weight is applied to past result data by similarity between the input x(t) of today and the past input to thereby make it possible to prepare the forecasting model of the index value and improve forecasting accuracy of the forecast value y*(t) of the index value.
In Example 1, in the online forecasting system, the input x(t) of today is compared with respective input x(t−N), x(t−N+1), x(t−1) for the past N days stored in the past index value's result data storing unit PASTDB to calculate similarities and the calculated similarities are set as weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t). This is based on an idea that a large weight is applied to result data of the index in an input similar to the input x(t) of today to prepare a forecasting model of the index value so as to make it possible to improve forecasting accuracy of the forecast value y*(t) of the index of today.
However, similarity between the input x(t) of today and an input x(t−j) of a past certain day Day(t−j) and a weight w(t−j, t) optimum for the Day(t−j), that is, the weight w(t−j, t) by which the forecast value y*(t) of the index of today forecasted by the prepared forecasting model becomes closest to a true value y(t) are not in a simple proportional relationship. Accordingly, like Example 1, when the forecasting model is prepared by using similarity, as it is, as a weight, there is a possibility that forecasting accuracy of the forecast value y*(t) of the index becomes low.
Here, Example 2 is characterized in that the weight estimation model preparation unit WESTIMATE prepares a weight estimation model using the result data of the past optimum weight stored in the past optimum weight's result data storing unit WDB and the weight estimation model preparation unit WESTIMATE estimates an estimation value w*(t−j, t) of the optimum weight w(t−j, t) from the input x(t) of today and the input x(t−j) of the past day Day(t−j).
The result value (true value) of the optimum weight w(t−j, t) can be obtained by calculating the weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) so that an error of the forecast value y*(t) of the index with respect to the result value y(t) is minimized, after the result value y(t) of the index of today is obtained by the processing execution/observation unit PROC.
In the following, the operations of the online forecasting system of Example 2 will be described in order using
In a case where data as illustrated in
As described above, a configuration in which the weight estimation model preparation unit WESTIMATE operates as a learning device having an internal state and learns result data of a newly added optimum weight online every day may be considered, instead that the weight estimation model is newly prepared using past optimum weight result data in a certain point in time every day. As the online learning device, a known “supervised online machine learning method” called an online neural network or an online Boltzman machine can be used.
Next, the weight estimation model preparation unit WESTIMATE calculates the estimation values w*(t−N, t) w*(t−N+1, t), . . . , w*(t−1, t) of the weight vectors from the input x(t) of today and the inputs x(t−N), x(t−N+1), . . . , x(t−1) for the past N days based on the prepared weight estimation model. The estimated weights w*(t−N, t), w*(t−N+1, t), . . . , w*(t−1, t) are displayed in a matrix pattern so as to make it possible for a user to visually grasp change in a tendency of result data of the index value. Details of the way of displaying will be described later in Example 3.
The weight multiplication unit WMUL multiplies the result value y(t−j) of the index of Day(t−j) stored in the past index value's result data storing unit PASTDB by the weight w*(t−j, t) calculated by the weight estimation model preparation unit WESTIMATE and outputs the multiplication result to the weight index value forecasting model preparation unit FORECAST.
Next, the index value forecasting model preparation unit FORECAST prepares the forecasting model, that is, the index value forecasting model used in forecasting the index value of today based on result data in the past stored in the past index value's result data storing unit PASTDB and the estimation value of the weight vector prepared by the weight estimation model preparation unit WESTIMATE. In this case, the result data of Day(t−j) is prepared by applying the weight w*(t−j, t) thereto. A specific method of preparing the index value forecasting model is the same as Example 1.
Next, the index value forecasting model preparation unit FORECAST performs forecasting of the index value y(t) of today from the input x(t) of today using the prepared index value forecasting model to calculate the forecast value y*(t). This is the same as processing of Example 1.
Next, the measure decision unit DECISION performs some kind of measure decision based on the forecast value y*(t). This is the same as processing of Example 1.
Thereafter, the processing execution/observation unit PROC actually executes processing for a target system or observes the target system to acquire the result value y(t) of the index of today and records the input x(t) of today and the result value y(t) of the index of today, as a set, in the past index value's result database PASTDB. This is the same as processing of Example 1.
Next, the weight optimization controller WOPT calculates the optimum values of the weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) from the result value y(t) of the index of today. In the weight estimation model preparation unit WESTIMATE, the result value y(t) of the index of today is known after processing is performed by the processing execution/observation unit PROC and the result value y(t) of the index of today is obtained while the result value y(t) of the index of today is unknown at the point in time when the weight vector is prepared and thus, it is possible to calculate the optimum weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) causing the error of the estimation value y*(t) of the index to the result value y(t) to be minimized. This is executed in such a way that the weight optimization controller WOPT repeatedly activates the past index value's result database PASTDB, the weight multiplication unit and the index value forecasting model preparation unit FORECAST according to an optimization method.
Next, the forecasting model of the index value is prepared by using the current weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) and the forecast value y*(t) of the index is calculated by the forecasting model of the index value prepared.
Next, a difference r=y*(t)−y(t) between the forecast value y*(t) and the result value y*(t) obtained in the processing execution/observation unit PROC described above is calculated. When the difference r converges, the current weight vectors w(t−N, w(t−N+1, . . . , w(t−1, t) are set as the optimum weight vectors to be output and processing is ended. In a case where the difference r does not converge, according to the optimization method which is being used, the current weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) are updated so that the difference r is minimized, and preparing for the forecasting model of the index value and calculating of a forecast value of the index are repeated again.
The optimum weights w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) determined here are displayed in a matrix pattern so as to make it possible for a user to visually grasp change in a tendency of the result data of the index value. Details of the way of displaying will be described later in Example 3.
When the optimum weight vectors w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) are obtained, for j which is an integer of 1 or more to N or less, the input x(t−j) of Day(tj), the input x(t) of today, and a weight w(t−j, t) of Day(t−j) to today are stored in the past optimum weight's result data storing unit WEB, as a set. With this, a set of the x(t−j), the x(t), and the w(t−j, t) is taken into account at the time when the weight estimation model at or after tomorrow (Day(t+1)) is prepared.
By doing as described above, the result value having the newest optimum weight is reflected in the weight estimation model at all times and it becomes possible to timely follow change in the tendency of the result value of the index. Here, only the result data of the optimum weight of the past M days is taken into account in preparing for the weight estimation model and thus, data of days before the past M days may be appropriately deleted.
In processing described above, a configuration in which only the result data having the optimum weight of the past M days, for M which is an integer determined in advance, is taken into account and data of days before the past M days is not taken into account in preparing for the weight estimation model is adopted, a configuration in which the result data of days before the past M days is taken into account, that is, all pieces of past data stored in the past optimum weight's result data storing unit WDB are used in preparing for the weight estimation model may be considered.
The weight estimation model allows to catch a tendency which does not change very large of the processing execution/observation unit PROC compared to matters that the index value forecasting model allows to catch a tendency which changes in a relatively short period of time of the processing execution/observation unit PROC and thus, it is considered that an excellent result is frequently obtained by using all pieces of usable result data including older data in preparing for the weight estimation model.
As described above, according to the online forecasting system in Example 2, the weight between the input x(t) of today and the past input is estimated based on the result data having the optimum weight calculated from the result value of the index so as to make it possible to improve forecasting accuracy of the forecast value y*(t) of the index.
Example 3 is an example in which the weight w(t1,t2) applied to result data of the past index value is displayed in a matrix pattern so as to make it possible for a user to visually grasp change in a tendency of the weight (or change in a tendency of result data of the index value) in the online forecasting system of Example 1 and Example 2. In Example 1 and Example 2, the weight w(t−j, t) indicates how much a weight is applied to be taken into account for data of the result value of the index before j days, that is, for Day(t−j) when the index value forecasting model is prepared in Day(t).
The weight w(t−j, t) indicates what extent the input x(t−j) of Day(t−j) and the result data y(t−j) of the index value are to be taken into account in forecasting the index value y(t) of Day(t) from the input x(t) of Day(t). That is, it is possible to regard the weight w(t−j, t) as indicating what extent the relationship between the input x and the index value y shows similar tendency between Day(t−j) and Day(t).
Like
In
In a case where only the tendency of result data of Day(t0) is different from that of other days, all pieces of past result data, that is, all pieces of result data before Day(t0) are equally useless in forecasting at Day(t0) and thus, a relative difference is hardly generated between weights of the vertical column at t=t0 and weights of cells before t=t0. Accordingly, especially, when the weight is set in Example 2, as illustrated in
In the example of a screen described here in which the weights are displayed in a matrix pattern, j is set to a range from 1 or more to N or less in compliance with matters that N weights w(t−N, t), w(t−N+1, t), . . . , w(t−1, t) to be respectively applied to past data for the past N days every day are set. However, as described in Example 1 and Example 2, a method in which the number of days of past data to be taken into account in preparing for the forecasting model of the index value is not determined in advance is also available. In this case, the number of vertical rows may be different for each column in displaying the matrix pattern of the weights of
Until now, although 1 is set as an integer of 1 or more in the weight w(t−j, t), matters that a negative integer can be set as j in the matrix pattern display of the weights can be considered. In a case where j is the negative integer, when it is set that j′=−j, w(t−j, t)=w(t+j′, t) represents a weight in a case where the result data of the index of Day(t+j′), which is after Day(t), can be used in preparing for the forecasting model of the index of Day(t). In the online forecasting system, although it is impossible to use result data of a future index in forecasting the index of today and the w(t+j′, t) is rarely considered, as explained in Example 3, it is useful to take into account the w(t−j′, t) for the purpose of matters that the user visually grasps chancre in a tendency of a consideration target system. Actually, in a case where the weight w(t−j, t) relating to the negative integer j is displayed, a portion where j is positive may be displayed in a lower side of the axis t and a portion where j is negative may be displayed in an upper side of the axis t in
As described above, according to the online forecasting system in Example 3, the weight w(t−j, t) used in preparing for the forecasting model of the index is presented to the user in a matrix shape so that the user is able to visually grasp change in the tendency of the relationship between the input x and the index y.
The embodiments described above are just one instance for realizing the present invention and a technical scope of the present invention is not limited thereto.
Some or all of respective configurations, functions, processing units, and processing means described above may be realized by hardware by designing those in, for example, an integrated circuit. The respective configurations, functions, and the like may be realized by software in such a way that a processor interprets a program realizing the respective functions and executes the program. Information such as the program realizing the respective functions, a table, and a file can be placed in a recording device such as a memory, a hard disk, or a solid state drive (SSD) or a recording medium such as an IC card, an SD card, or a DVD.
The present embodiment is not limited to an example of the physical distribution warehouse and can be generally applied to the online processing system.
As another example, a retail store is considered. In the retail store, the number of store visitors or sales of that day is forecasted from day of the week, weather, information about events held in the vicinity, or the like based on past result data and the number of purchases, a job shift of a clerk, or the like is determined according to the forecasted number of store visitors or sales of that day. At the closing time, the result value of the number of actual store visitors or the actual sales of that day is recorded as result data and is used at or after the present time. In the example of the retail store, it may be considered that day of the week, weather, information about events held in the vicinity, or the like corresponds to input data in the online processing system described above, the number of store visitors or sales corresponds to the index in the online processing system described above, and determining of the number of purchases, a job shift of a clerk, or the like from the forecasted value of the number of store visitors or sales corresponds to measure decision in the online processing system described above.
As still another example, financial investment is considered. In the financial investment, profits to be obtained in the future through investment are forecasted from various financial index values or information related to an investment target in each point in time based on past result data and it is determined whether investments should be actually performed or not. Thereafter, a result value of the actually obtained profits is recorded as result data and is used for forecasting at or after the present time. In the example of financial investment, various financial index values or information related to an investment target in each point in time correspond to input data in the online processing system described above, the profits to be obtained in the future through investment correspond to the index in the online processing system described above, and the determination as to whether investments should be actually performed or not from a forecasted value of future profits corresponds to measure decision in the online processing system described above.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2015/074899 | 9/2/2015 | WO | 00 |
