JP2021103444A - Demand forecasting system - Google Patents

Abstract

To obtain high prediction accuracy in EC demand forecast by applying machine learning incorporating EC access information.SOLUTION: A demand forecasting system is provided with means for obtaining explanatory variables including information based on the sales performance of a product, and means for forecasting demand of the product in electronic commerce by inputting the acquired explanatory variables into a model generated by machine learning.SELECTED DRAWING: Figure 1

Description

The present invention relates to a demand forecasting system.

A demand forecasting system adapted to actual stores (see, for example, Patent Documents 1 and 2) was born around 2000, and is being used by many companies in the current retail industry to be effective. Recently, with the expansion of the electronic commerce (hereinafter referred to as EC) market, the EC rate by retailers who operate actual stores is on the rise.

JP-A-2017-010436 JP-A-2017-1230888

However, even if the demand forecast system for an actual store is directly applied to the demand forecast in EC, the forecast accuracy as in the actual store cannot be obtained. The reason for this is that since the actual store has a limited trade area, factors such as price, sales promotion, store location conditions, and weather should be taken into consideration, but for EC, the trade area is nationwide and media information (SNS, TV, etc.) Since it is highly linked with the news app), it may not be possible to obtain high accuracy with the factors and methods used in the actual demand forecast for stores.

The present invention has been made in view of these problems, and an object of the present invention is to provide a technique capable of obtaining high prediction accuracy in EC demand forecast by applying AI (machine learning) incorporating EC access information. It is in.

One aspect of the present invention relates to a demand forecasting system. This demand forecasting system predicts the demand for goods in electronic commerce by means of acquiring explanatory variables containing information based on the sales performance of the goods and by inputting the obtained explanatory variables into a model generated by machine learning. Means and means to do so.

It should be noted that any combination of the above components and those in which the components and expressions of the present invention are mutually replaced between devices, methods, computer programs, recording media storing computer programs, etc. are also aspects of the present invention. It is effective as.

According to the present invention, high prediction accuracy can be obtained in EC demand forecast by applying machine learning incorporating EC access information.

It is a schematic diagram for demonstrating the demand forecasting service provided by the demand forecasting system which concerns on embodiment. It is a hardware block diagram of the demand forecasting system of FIG. It is a block diagram which shows the function and structure of the demand forecasting system of FIG. It is explanatory drawing of the explanatory variable used in the demand forecasting system of FIG. It is a schematic diagram for demonstrating the derivation of the forecast sales number using the forecast model of FSXGBT. It is a figure explaining the contribution to the forecast of the demand by the gradient of the number of PV. It is a figure explaining the contribution to the forecast of the demand by the average about PV. It is a flowchart which shows the flow of a series of processing in the demand forecasting system of FIG. It is a figure explaining the application of the FS model to EC. It is a figure which shows the image of a long tail.

Hereinafter, the same or equivalent components, members, and processes shown in the drawings shall be designated by the same reference numerals, and redundant description will be omitted as appropriate. In addition, some of the members that are not important for explanation are omitted in each drawing.

FIG. 1 is a schematic diagram for explaining a demand forecasting service provided by the demand forecasting system 10 according to the embodiment. The demand forecast system 10 collects various data such as data on product sales performance and data on access to product information from service user companies 20 and 3rd party 30 via a network (not shown) such as the Internet. To do. The demand forecasting system 10 creates a forecasting model by machine learning the collected data. The demand forecasting system 10 predicts the number of products sold by inputting the latest sales performance of the product and the number of accesses to the product information as explanatory variables into the created forecast model. The demand forecast system 10 distributes forecast data such as the forecast sales number obtained as a result of the forecast to the service user company 20. Alternatively, the service user company 20 may acquire similar forecast data by using the forecast data API provided by the demand forecast system 10.

The data provided by the service user company 20 to the demand forecasting system 10 is, for example, sales data including the number of products sold and the unit price, data related to product inventory, product ordering data, and product data (product master data). ), Data on access to product information in EC, and campaign data. The service user company 20 can predict the sales of the product from the forecast data provided by the demand forecast system 10, and can optimize the order quantity of the product.

The data related to the access to the product information in the EC may be the frequency information indicating the frequency of the access to the product information via the network, and more specifically, the PV (Page View) of the product introduction page in the EC. ) It may be a number. The PV number is an aspect of attention level information indicating the degree of attention to a product in EC. The number of PVs and the degree of attention have a relationship that the larger the number of PVs, the higher the degree of attention to the product. In another aspect, the attention level information may be, for example, the number and frequency of appearances of the product name in a message service such as SNS or tweet, or may be a conversion rate.

The data provided by the 3rd party 30 to the demand forecasting system 10 includes, for example, event information, weather information, Web behavior data, and search keywords. The 3rd party 30 may include, for example, a commercial database or a database provided by a public institution.

In this embodiment, in particular, the demand forecasting system 10 builds a forecasting model specialized for EC and forecasts the demand for goods in EC. That is, the demand forecasting system 10 acquires the number of PVs associated with the product from the service user company 20 together with the data showing the sales performance of the product, and uses those data as training data to predict the demand in EC. To build. The forecast model constructed in this way outputs the forecast sales number of products in EC by using the number of PVs as an explanatory variable in addition to the sales performance of the latest product. In addition, the explanatory variables of such a prediction model also include variables obtained by processing the sales performance of the product and the number of PVs. According to such a demand forecasting system 10, the accuracy of demand forecasting of goods in EC can be improved.

FIG. 2 is a hardware configuration diagram of the demand forecasting system 10 of FIG. The demand forecast system 10 includes a memory 110, a processor 112, a communication interface 114, a display 108, and an input interface 118. Each of these elements is connected to bus 120 and communicates with each other via bus 120. The demand forecast system 10 may be realized by one server, or may include a plurality of servers having the configuration shown in FIG.

The memory 110 is a storage area for storing data and programs. The data or program may be permanently stored in the memory 110 or may be temporarily stored. The processor 112 realizes various functions of the demand forecasting system 10 by executing a program stored in the memory 110. The communication interface 114 is an interface for transmitting and receiving data to and from the outside of the demand forecasting system 10. The communication interface 114 is connected to the network and exchanges data with the server of the service user company 20 and the server of the 3rd party 30 via the network. The display 108 is a device for displaying various information. The input interface 118 is a device for receiving input from the user.

FIG. 3 is a block diagram showing the functions and configurations of the demand forecasting system 10 of FIG. Each block shown here can be realized by elements such as the CPU of a computer or a mechanical device in terms of hardware, and can be realized by a computer program or the like in terms of software, but here, it is realized by cooperation between them. It depicts a functional block to be done. Therefore, it will be understood by those skilled in the art who have referred to this specification that these functional blocks can be realized in various forms by combining hardware and software.

The demand forecast system 10 includes a data processing unit 130, an explanatory variable acquisition unit 132, a demand forecast unit 134, a forecast result providing unit 136, a model creation / update unit 138, and a forecast model holding unit 140.

The data processing unit 130 acquires information based on the sales performance of the product and information on the degree of attention of the product, and processes the acquired information. Information based on the sales performance of the product includes, for example, the unit price of the product per week over the past four weeks and the number of products sold per week. The attention level information includes, for example, the number of PVs of the product introduction page for each week over the past four weeks. The data processing unit 130 acquires and processes the product master data.

The explanatory variable acquisition unit 132 acquires an explanatory variable to be input to the prediction model. The explanatory variables acquired include the unit price of the product, the number of products sold, the product master data, and the number of PVs. The details of the explanatory variables will be described later.

The demand forecasting unit 134 predicts the demand for goods in EC by inputting the explanatory variables acquired by the explanatory variable acquisition unit 132 into the prediction model generated by machine learning. The demand forecasting unit 134 uses a forecasting model held by the forecasting model holding unit 140.

The machine learning in the present embodiment is machine learning with teacher data, for example, a tree (decision tree (CART), regression tree, random forest, gradient boosted tree (Gradient Boosted Trees)), neural network (perceptor, convolution). Neural network (CNN), recurrence type neural network (RNN), residual network (ResNet)), bayes (naive bayes), clustering (k-nearest neighbor method (KNN)), ensemble learning (boostering, bagging) Etc. may be used. In this embodiment, decision tree-based machine learning, particularly Frequency Boosted Trees (referred to as FSXGBT), which is a kind of gradient boosting tree, is used.

The FSXGBT is a Frequency-Severity model (FS model, known as a two-part model).
https://www.researchgate.net/profile/Xiaoping_Feng/publication/282598735_Dependent_frequency-severity_modeling_of_insurance_claims/links/5b98b9c592851c4ba81220ea/Dependent-frequency-severity-modeling-of-insurance-claims.pdf
XGBT (eXtreme Grandient Boosted Trees)
https://arxiv.org/pdf/1603.02754.pdf
This is the model applied to). The FS model is a method for obtaining "a continuous quantity at the time of event occurrence on the premise of a rare event". For example, when a rare event occurs, the quantity is obtained. It is applied when you want to.

FIG. 9 is a diagram illustrating application of the FS model to EC. The FS model consists of two parts, the first part 902 assessing whether an event occurs. The next part, Part 904, evaluates the continuous quantity when an event occurs. In FSXGBT, the next part 904 is XGBT.

It is known that retailers in EC often handle a large number of products with zero sales. This is known as the "long tail". FIG. 10 is a diagram showing an image of a long tail. In EC, if you drill down to a single item on a daily basis, there is a large amount of data with zero sales. For example, in EC, about 80% of all products handled may have about 0 sales per day. In the EC demand forecast, sales do not occur much, but when sales occur, it is necessary to forecast the number of sales. However, when learning data containing a large amount of data with zero sales by a machine learning algorithm, a phenomenon called "zero boost problem" in which learning cannot be performed well may occur.

Therefore, in the present embodiment, learning / prediction is not performed directly from the data including the long tail, but learning / prediction is performed in two steps by applying the FS model.
(1) First, whether or not sales will occur (estimate the frequency of sales, 906 in FIG. 9)
(2) Next, when sales occur, the number of sales is estimated (regression of the number of sales, 908 in FIG. 9).

In this way, focusing on the similarity between the structure of the FS model and the structure of sales at EC retail against the background of long tails (no basic sales / number of sales at the time of sales), EC demand forecast By applying the FS model to, the accuracy of the EC demand forecasting model can be improved.

In addition, the Tweedie distribution is used as an algorithm optimization index for the purpose of algorithm optimization of the FS model applied to EC demand forecasting. In order to model "presence or absence of sales" and "number of sales at the time of sales" at the same time, the probability of occurrence of a discrete event "presence or absence of sales" (generally according to the Poisson distribution) and "at the time of sales occurrence" It is necessary to handle "the number of sales", which is a continuous probability that can express how many sales events occur within a certain period (generally according to the gamma distribution), at the same time. The Tweedie distribution is a probability distribution that combines the Poisson distribution and the gamma distribution. By using the Tweedie distribution as the population distribution and optimizing the model, it is possible to grasp the sales structure in EC retail against the background of the long tail, such as selling several pieces at the time of sales, although sales do not basically occur.

The forecast result providing unit 136 provides the service user company 20 with the result of the forecast obtained by the demand forecasting unit 134.

The model creation / update unit 138 generates a prediction model of FSXGBT and registers it in the prediction model holding unit 140. Further, the model creation / update unit 138 holds the predicted model holding unit 140 by re-machine learning the data accumulated up to that point as teacher data on a regular basis or when a predetermined amount of new data is accumulated. Update the forecast model to be done.

FIG. 4 is an explanatory diagram of explanatory variables used in the demand forecasting system 10. In FIG. 4, the nth week is the week to be predicted, and the sales performance data and the number of PVs (transaction data) for the past 4 weeks (n-1 week to n-4 weeks) are provided by the service user company 20. Suppose you are. In addition to the sales performance data, the service user company 20 also provides product master data.

The sales performance data includes the unit price of each product one week ago, two weeks ago, three weeks ago, and four weeks ago. The sales performance data includes the number of sales of each product one week ago, two weeks ago, three weeks ago, and four weeks ago. The number of PVs on each product introduction page 1 week ago, 2 weeks ago, 3 weeks ago, and 4 weeks ago is also provided. The unit price, the number of sales, and the number of PVs are quantitative variables that take continuous values.

The product master data includes the JAN code of the product, the store code of the store that sold the product, the category code of the product, and the attribute code of the product. Various codes included in the product master data are qualitative variables that take discrete values such as classifications and flags. The JAN code is a universal product identification number that indicates "which company and which product". The store code is an identifier that identifies the store that sold the product. The category code is an identifier indicating the genre of the product (for example, television, audio, air conditioner, etc.). The attribute code is an identifier indicating the attribute of the product. For example, in the case of USB, the attribute code indicates whether it is type A or type C, and in the case of a battery, it indicates whether it is AA, AA, or other. The attribute code may also indicate the color, shape, weight, etc. of the product.

In the demand forecasting system 10, the following 15 explanatory variables are used for forecasting the demand for a certain product in EC.
(1) Estimated number of sales in the next week corresponding to price elasticity (2) Number of PVs one week ago (3) Number of weeks elapsed from the time of release (4) Estimated number of sales in the next week corresponding to PV gradient ( 5) Price decline rate from the time of release (6) JAN code (7) Average number of PVs over the past 3 weeks (8) Number of PVs 2 weeks ago (9) Store code (10) Number of sales 4 weeks ago (11) ) Estimated number of sales for the next week corresponding to the average performance (12) Number of sales one week ago (13) Category code (14) Number of sales two weeks ago (15) Attribute code

(1), (3), (4), (5), (7), and (11) are explanatory variables obtained by processing by the data processing unit 130. Each processing method will be described below.

(1): Estimated number of sales in the next week corresponding to price elasticity (unit price one week ago, number of sales one week ago), (unit price two weeks ago, number of sales two weeks ago) Linear regression is performed on the model, and the regression equation is derived. By substituting the planned unit price for the next week into the derived regression equation, the estimated number of sales for the next week is calculated. The estimated number of sales calculated in this way is the number of sales at the future unit price calculated by a regression equation based on the relationship between the past unit price and the number of sales, and reflects the price elasticity.

(3) Number of weeks elapsed from the time of release The product release date and time is acquired from the product master data, and the difference (weekly unit) between the acquired release date and time and the current date and time is defined as the number of weeks elapsed from the time of release.

(4) Estimated number of sales in the next week corresponding to the PV gradient First, the average number of PVs in the last 3 weeks is calculated as 3wPV (this is the average number of PVs in the past 3 weeks in (7)). Assuming that the number of PVs one week ago is 1 wPV, the estimated number of sales is calculated by the following formula.
(1wPV / 3wPV) x (Number of sales 1 week ago)
Here, (1 wPV / 3 wPV) corresponds to the slope of the number of PVs. Therefore, the estimated number of sales calculated in this way corresponds to the gradient of PV.

(5) Price decline rate from the time of release The price decline rate is calculated by the following formula.
(Unit price at the time of release-1 week ago unit price) / (Number of weeks elapsed from the time of release)

(11) Estimated number of sales in the next week corresponding to the average performance Estimated sales by calculating the number of sales per PV from the data of the past one month (4 weeks) and multiplying it by the number of PVs one week ago (1wPV) Calculate the number. Specifically, first, the total number of sales in the past month is calculated as 1 m, and the total number of PVs in the past month is calculated as 1 mPV. Next, the estimated number of sales is calculated by the following formula.
(1m sales / 1mPV) x 1wPV
The estimated number of sales calculated in this way corresponds to the average sales performance for the past month.

The explanatory variable acquisition unit 132 of the demand forecasting system 10 acquires the explanatory variables of (1) to (15). The demand forecast unit 134 calculates the forecast sales number for the nth week by inputting the acquired explanatory variables (1) to (15) into the forecast model of FSXGBT.

FIG. 5 is a schematic diagram for explaining the derivation of the forecast sales number using the forecast model of FSXGBT. In the derivation of the predicted sales number using the FSXGBT prediction model, a large number of decision trees are generated, and the results of each decision tree are aggregated and predicted. The decision tree is gradually increased, and the label of the case where the generated decision tree is mistaken is updated to generate a new decision tree. In the example of FIG. 5, M decision trees of 50_1, 50_2, ..., 50_M are generated. The forecast model outputs the total number of forecast sales calculated by each decision tree.

FIG. 6 is a diagram illustrating the contribution of the gradient of the number of PVs to the forecast of demand. In the example of FIG. 6, first, the average number of PVs over the past three weeks is calculated as (10,000 + 20,000 + 30,000) / 3 = 20,000 PV. Next, the PV gradient from the previous week to this week is calculated to be 30,000 / 20,000 = 1.5. The calculated "1.5" indicates that the latest PV has increased "1.5" times with respect to the average. Assume that the PV gradient from this week to next week is the same "1.5", and that PV growth correlates with sales growth. Then, by multiplying the number of sales this week by 1.5, which is the growth rate of PV, the estimated number of sales next week corresponding to the gradient of PV can be calculated. The demand forecasting unit 134 includes the estimated number of sales in the explanatory variables to realize the demand forecasting in consideration of the PV gradient.

FIG. 7 is a diagram illustrating the contribution of the average of PV to the forecast of demand. In the example of FIG. 7, from the total number of PVs for the past 4 weeks (90,000 PV) and the total number of sales for the past 4 weeks (900 pieces), the number of sales per PV is 900 (pieces) / 90,000. It is calculated as (PV) = 0.01 (pieces / PV). Assume that the number of PVs this week and the number of PVs next week are the same. Then, by multiplying the calculated coefficient 0.01 (pieces / PV) by 30,000 PV, which is the planned number of PVs for next week (= the number of PVs for this week), the estimated number of sales for next week corresponding to the average performance can be calculated. .. By including this estimated number of sales in the explanatory variable, the demand forecasting unit 134 realizes the demand forecasting in consideration of the average of PV.

The operation of the demand forecasting system 10 based on the above configuration will be described.
FIG. 8 is a flowchart showing a series of processing flows in the demand forecasting system 10 of FIG. The demand forecast system 10 acquires the sales results and the number of PVs for the past four weeks, and the product master data (S802). The demand forecast system 10 calculates processing variables such as the forecast sales number by processing the data acquired in step S802 (S804). The demand forecasting system 10 inputs an explanatory variable selected from the data acquired in step S802 and the processing variables obtained in step S804 into the forecasting model (S806). The demand forecasting system 10 provides the forecasting result of the number of sales output by the forecasting model to the user or the outside (S808).

In the above-described embodiment, examples of the holding unit are a hard disk and a semiconductor memory. Further, based on the description of this specification, each part is provided by a CPU (not shown), an installed application program module, a system program module, a semiconductor memory for temporarily storing the contents of data read from the hard disk, or the like. What can be achieved is understood by those skilled in the art who have referred to this specification.

According to the demand forecasting system 10 according to the present embodiment, the accuracy of forecasting the demand for goods in EC can be improved by using the forecasting model generated by machine learning.

In the demand forecasting system 10 according to the present embodiment, a decision tree-based forecasting model is used. Therefore, both quantitative and qualitative variables can be used as explanatory variables. For example, if the quantitative variable is the number of PVs, whether the number of PVs is greater than or equal to the threshold value or not may be used as the branching condition of the decision tree, and if the qualitative variable is the genre of the product, the branching of the decision tree may be used. The genre may be used as a condition. As a result, more kinds of variables can be used for the prediction, and the accuracy of the prediction is improved. In particular, the demand forecast of the present embodiment is advantageous in that the qualitative variables can be effectively used as compared with the regression equation-based forecast in which the qualitative variables are difficult to handle.

In the demand forecasting system 10 according to the present embodiment, the number of PVs, which is a parameter peculiar to EC, is used as an explanatory variable. Therefore, it is possible to realize a demand forecast more specialized for EC.

In the demand forecast system 10 according to the present embodiment, in addition to the past sales number and PV number, the forecast sales number corresponding to the PV gradient obtained by processing them and the forecast sales number corresponding to the average actual result are also explanatory variables. It is said. As explained with reference to FIGS. 6 and 7, the contribution of these forecast sales numbers to the demand forecast is high. By processing the variables into such variables with a high degree of contribution and then using them as explanatory variables, the accuracy of prediction can be further improved.

Comparing the conventional statistical model with the AI utilization model according to the present embodiment, the AI utilization model has the flexibility to respond to sudden business changes. That is, the sales tendency in retail such as EC changes depending on factors such as individual item level and product attributes (for example, category). On the other hand, the "effect" and "effectiveness" of these effects differ depending on the product category. Traditional statistical models are strong against regular waveforms created by mathematical algorithms, but weak against irregular waveforms. On the other hand, the decision tree-based prediction model adopted in this embodiment makes it possible to apply the "type of influence" and "how the effect is produced" to "for each product and its classification". ..

Conventional statistical models include multivariate analysis systems such as multiple regression analysis and time series analysis systems such as ARIMA. The former is the idea of explaining the result from the cause, and the latter is the idea of following from the result on the premise that the cause is unknown. In both cases, it is difficult to capture trends for products that are moving rapidly. In both cases, when there is a change in business, it takes a relatively long time to review the model because the calculation method is examined after accumulating and analyzing the results.

On the other hand, the AI utilization model according to the present embodiment is a decision tree-based prediction model such as FSXGBT, and is based on the idea of analyzing and explaining a lot of data with correct answers (teacher data). Since this prediction model has a decision tree structure, it is possible to easily capture sudden movements that are not an extension of the time series trend. In addition, when there is a change in business, the model can be reviewed speedily by updating the model to add features.

For example, if you upload a video complimenting a product that has a YouTuber with more than 1 million subscribers, the product may sell explosively the next day or later. It is not possible to capture such sudden fluctuations in demand with methods that use conventional statistical models. On the other hand, the AI utilization model according to the present embodiment is a decision tree-based model in which the number of PVs is included in the explanatory variables. Therefore, it is possible to predict an explosive increase in demand from the surge in the number of PVs that occurs immediately after uploading a video.

The configuration and operation of the demand forecast system 10 according to the embodiment have been described above. This embodiment is an example, and it will be understood by those skilled in the art that various modifications are possible for each component and each combination of processes, and that such modifications are also within the scope of the present invention.

10 Demand forecasting system, 20 Service users, 30 3rd party.

Claims (8)

A means of obtaining explanatory variables that include information based on product sales performance,
A demand forecasting system including means for predicting the demand for the goods in electronic commerce by inputting the acquired explanatory variables into a model generated by machine learning. The demand forecasting system according to claim 1, wherein the model is a decision tree-based model. The demand forecasting system according to claim 2, wherein the explanatory variable further includes a qualitative variable related to the product. The demand forecasting system according to any one of claims 1 to 3, wherein the explanatory variable further includes attention level information indicating the degree of attention to the product in electronic commerce. The demand forecasting system according to claim 4, wherein the explanatory variables further include a variable based on the average of the attention level information and a variable based on the gradient of the attention level information. The demand forecasting system according to any one of claims 1 to 5, which uses a Frequency-Severity model as machine learning. The Frequency-Severity model has a first part for evaluating whether or not a sale occurs and a second part for evaluating the number of sales when a sale occurs.
The demand forecasting system according to claim 6, wherein the eXtreme Grandient Boosted Trees is applied to the second part. The demand forecasting system according to claim 6 or 7, wherein the Tweedie distribution is used as an index for algorithm optimization of the Frequency-Severity model.

Images