Prediction device and prediction method

The prediction device improves sales volume forecasting by segmenting and combining similar sales data segments, addressing the challenge of insufficient training data and improving prediction accuracy.

WO2026234745A1PCT designated stage Publication Date: 2026-11-12NT T INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2025/016766
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Filing Date
2025-05-07
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing sales forecasting methods struggle to accurately predict sales volume, especially in small and medium-sized stores or newly opened stores, due to insufficient training data and fragmented sales performance data, leading to inaccurate predictions.

Method used

A prediction device and method that acquires sales performance data, generates features, divides the data into short-term segments with similar trends, combines similar segments, and trains a prediction model using the combined data to improve accuracy.

Benefits of technology

The method stabilizes the mean and variance of combined data sets, ensuring sufficient samples and reducing variability, thereby enhancing the accuracy of sales volume predictions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025016766_12112026_PF_FP_ABST
    Figure JP2025016766_12112026_PF_FP_ABST
Patent Text Reader

Abstract

This prediction device includes: an acquisition unit that acquires sales performance data showing a time series of sales performance including features and sales volume related to a product; a generation unit that generates a feature from the sales performance data; a splitting unit that splits the sales performance data into a plurality of pieces of short-term data; a combining unit that combines pieces of short-term data in which the similarity of features extracted from the short-term data is equal to or greater than a predetermined threshold to create combined data; a training unit that trains a prediction model for predicting the sales volume of the product by using the combined data as training data; and a prediction unit that predicts the sales volume of the product in a prediction period by using the prediction model.
Need to check novelty before this filing date? Find Prior Art

Description

Prediction device and prediction method

[0001] The technology disclosed herein relates to a prediction device and a prediction method.

[0002] Techniques for predicting demand, such as the number of products sold, using machine learning models are known. For example, Non-Patent Document 1 describes a sales forecasting technique using macroeconomic indicators. Non-Patent Documents 2 and 4 describe constructing a sales forecasting model using sales figures, dates, day-of-the-week labels, and weather data. Non-Patent Document 3 describes a sales forecasting technique that takes into account holidays and seasonal factors.

[0003] Furthermore, there are known methods to improve prediction accuracy by splitting sales performance data based on time-varying sales trends (demand) before performing machine learning. However, if the sales performance data is split too much, a problem arises: insufficient samples for training data. This problem is more likely to occur in small and medium-sized stores where sales performance data has not been accumulated, as well as in newly opened stores.

[0004] In contrast, for example, Non-Patent Document 5 describes a method of hierarchically clustering and retraining primary prediction residuals, which fluctuate less than sales figures (i.e., are less likely to show different trends from period to period). Patent Document 1 describes obtaining first prediction data from the sum of trend component data and pattern data calculated for each time-series cluster of actual data, and then performing regression learning on the difference between actual data and first prediction data. Patent Document 2 describes adding labels indicating whether or not a product was sold to short-term data obtained by dividing sales data, and generating a classifier that outputs a score indicating the likelihood of a product being purchased based on long-term data that includes all short-term data on the dates when the product was sold, and long-term data that includes all short-term data on the dates when the product was not sold.

[0005] Japanese Patent Publication No. 2023-124238 Japanese Patent Publication No. 2009-251779

[0006] G. Verstraete, E.H. Aghezzaf, B. Desmet (2020), “A leading macroeconomic indicators’ based framework to automatically generate tactical sales forecasts”, Computers & Industrial Engineering, Volume 139, 106169Y. Suimon, H. Tanabe, K. Izumi (2023), “Using weather-based machine learning approach to estimate retail sales and interpret weather factors”, Proceedings of 14th IIAI International Congress on Advanced Applied Informatics, pp. 725-727H. Jiang, J. Ruan, J. Sun (2021), “Application of Machine Learning Model and Hybrid Model in Retail Sales Forecast”, Proceeding of IEEE 6th International Conference on Big Data Analytics (ICBDA), pp. 69-75Nita, S (2015), “Application of big data technology in support of food manufacturers commodity demand forecasting”, NEC Technical Journal, Volume 10(1), pp. 90-93Li, Jiawen, et al. (2024), “A Hierarchical RF-XGBoost Model for Short-Cycle Agricultural Product Sales Forecasting”, Foods 13.18: 2936

[0007] Even with conventional methods that consider sales trends, it was sometimes impossible to predict sales volume with sufficient accuracy. For example, the technology described in Non-Patent Document 5 did not perform segmentation based on sales volume, so its effect on improving the accuracy of sales volume prediction was limited. In the technology described in Patent Document 1, the actual data was sometimes too fragmented by clustering, making it impossible to process subsequent data such as extracting trend component data and pattern data. In the technology described in Patent Document 2, it was useful for predicting whether or not a product could be sold, but it could not predict the number of sales.

[0008] The disclosed technology was developed in view of the above points, and aims to provide a forecasting device and forecasting method that can improve the accuracy of sales volume forecasts.

[0009] A first aspect of this disclosure is a prediction device comprising: an acquisition unit that acquires sales performance data representing a time series of sales performance including product features and sales volume; a generation unit that generates features from the sales performance data; a division unit that divides the sales performance data into a plurality of short-term data; a joining unit that combines short-term data where the similarity of the features extracted from the short-term data is above a predetermined threshold to form joined data; a learning unit that learns a prediction model for predicting the sales volume of the product using the joined data as learning data; and a prediction unit that uses the prediction model to predict the sales volume of the product during a prediction period.

[0010] A second aspect of this disclosure is a prediction method, wherein a computer performs the following processes: acquires sales performance data representing a time series of sales performance including product features and sales volume; generates features from the sales performance data; divides the sales performance data into a plurality of short-term data; combines short-term data where the similarity of the features extracted from the short-term data is above a predetermined threshold to form combined data; trains a prediction model to predict the sales volume of the product using the combined data as training data; and uses the prediction model to predict the sales volume of the product during the prediction period.

[0011] The disclosed technology can improve the accuracy of sales forecasts.

[0012] This is a block diagram showing an example of the hardware configuration of a prediction device. This is a block diagram showing an example of the functional configuration of a prediction device. This is a diagram showing an example of sales performance data. This is a diagram showing an example of features. This is a diagram to explain the processing of the splitting section. This is a diagram to explain the processing of the splitting section. This is a diagram to explain the processing of the joining section. This is a diagram to explain the processing of the learning section. This is a flowchart showing an example of the prediction processing flow. This is a diagram to explain the splitting and joining of data related to the embodiment. This is a diagram to explain the splitting of data related to the comparative example. This is a diagram to explain the effects related to the embodiment and the comparative example.

[0013] An example of an embodiment of the disclosed technology will be described below with reference to the drawings. In each drawing, identical or equivalent components and parts are given the same reference numerals. Furthermore, the dimensional ratios in the drawings are exaggerated for illustrative purposes and may differ from actual ratios.

[0014] This embodiment relates to a forecasting device 10 that predicts the number of products sold during a forecast period using a forecasting model that predicts the number of products sold. The forecasting device 10 aims to accurately predict the number of products sold by also considering sales trends (demand) that may fluctuate over time. "Sales trends" refers to a classification of multiple factors that may affect the number of products sold, such as price, number of customers, number of items on display, and weather, with each combination of values ​​being uniquely represented.

[0015] Figure 1 is a block diagram showing an example of the hardware configuration of the prediction device 10. The prediction device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, storage 14, an input unit 15, a display unit 16, and a communication interface 17. Each component is connected to the others via a bus 19 so as to be able to communicate with each other.

[0016] The CPU 11 is a central processing unit that executes various programs and controls various parts. Specifically, the CPU 11 reads a program from the ROM 12 or storage 14 and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculations according to the program stored in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores a prediction program for executing the prediction process described later. The prediction program may be a single program or a group of programs composed of multiple programs or modules.

[0017] ROM 12 stores various programs and data. RAM 13 temporarily stores programs or data as a working area. Storage 14 consists of storage devices such as HDD (Hard Disk Drive) and SSD (Solid State Drive) and stores various programs and data, including the operating system.

[0018] The input unit 15 includes a pointing device such as a mouse, a keyboard, and buttons, and is used for various types of input. The display unit 16 is, for example, a liquid crystal display and displays various types of information. The display unit 16 may also function as the input unit 15 by employing a touch panel system.

[0019] Communication I / F17 is an interface for communicating with other devices. For this communication, wired communication standards such as Ethernet (registered trademark) or FDDI (Fiber Distributed Data Interface), or wireless communication standards such as 4G, 5G, or Wi-Fi (registered trademark) may be used.

[0020] Next, the functional configuration of the prediction device 10 will be described. Figure 2 is a block diagram illustrating an example of the functional configuration of the prediction device 10. The prediction device 10 includes respective functional units of an acquisition unit 20, a generation unit 22, a division unit 24, a combination unit 26, a learning unit 28, and a prediction unit 30. Each functional unit is implemented by the CPU 11 reading out a prediction program stored in the ROM 12 or the storage 14, expanding it into the RAM 13, and executing the program.

[0021] Furthermore, the prediction device 10 includes respective storage units of a sales performance data storage unit 21, a feature value storage unit 23, a short-term data storage unit 25, a combined data storage unit 27, and a prediction model storage unit 29. Each storage unit is implemented as a non-volatile storage or memory area, or a database logically configuring these, and stores and manages various types of data.

[0022] The acquisition unit 20 acquires sales performance data representing a time series of sales performance including feature values related to products and sales quantities. "Sales performance data" is multi-dimensional data including sales quantities aggregated for each predetermined period (for example, 1 day or 1 hour) for a certain product or product category, and a plurality of elements that may affect the sales quantities. For example, the sales performance data may include elements related to product features such as price, number of display units, inventory quantity, and production information (producer name, production area, raw materials, etc.). Also for example, the sales performance data may include elements related to store environment such as number of store visitors, weather data (temperature, precipitation, humidity, etc.), sales promotion information (presence or absence of coupon issuance, special sales, etc.), and calendar information (day of week, national holiday, event day, etc.).

[0023] Specifically, the acquisition unit 20 receives respective data indicating these elements from an external POS server, inventory management system, weather server, and the like via the communication I / F 17. Further, the acquisition unit 0 arranges each received data into the format of the sales performance data, and stores the formatted data in the sales performance data storage unit 21.

[0024] In the upper part of Figure 3, an example of sales performance data for a certain product A is shown. In this example, the sales performance data includes elements of the number of sales counted for each day from February 1st onwards, price, number of store visitors, number of displays, and weather data. The date labels are set as 2 / 1, 2 / 2, ... according to the dates. The daily sales quantities are denoted as y1, y2, .... The daily prices are denoted as xa1, xa2, .... The daily numbers of store visitors are denoted as xb1, xb2, .... The daily numbers of displays are denoted as xc1, xc2, .... The daily weather conditions are denoted as xd1, xd2, ....

[0025] The prediction device 10 generates a prediction model by using, in addition to the raw data (for example, the price itself) included in the sales performance data, processed data of the raw data as feature quantities. A feature quantity is a quantitative index used as an explanatory variable of the prediction model which is a machine learning model, and is obtained by converting elements that can affect the number of sales that is the output of the prediction model into numerical values or category values.

[0026] The generation unit 22 generates feature quantities from the sales performance data. As shown in the lower part of Figure 3, the generation unit 22 generates at least one feature quantity in the aggregation unit (for example, one day unit) of the sales performance data. In the example of Figure 3, three feature quantities a to c are generated for each day.

[0027] Figure 4 shows an example of feature quantities. This example shows nine examples of feature quantities generated using the sales performance data from February 1st to February 14th and each element of February 15th itself as original data when the prediction period is set to February 15th. Figure 4 lists the original data, data processing methods, examples of feature quantity definitions, and examples of feature quantity values. As the elements of February 15th itself, for example, items that can be grasped in advance such as date label, price, number of displays, inventory quantity, sales promotion information, and production information can be used. The generation unit 22 generates the feature quantities of No. 1 to No. 9 in Figure 4, or any other arbitrary feature quantities as the feature quantities a to c in Figure 3.

[0028] Features No. 1 to No. 3 are based on sales figures y1 to y14. Feature No. 1 uses a lag to directly adopt the sales figures from 7 days prior to the forecast period. For example, the feature corresponding to February 15th is set to the sales figure y8 from February 8th. This is to capture the trend for the same day of the week in the previous week. Feature No. 2 uses a moving average to calculate the average sales figures for the two days 7 to 8 days prior to the forecast period. For example, the feature corresponding to February 15th is set to the average of the sales figures y7 from February 7th and y8 from February 8th. This is to smooth out short-term noise and capture local trends. Feature No. 3 uses a weekday average to calculate the average sales figures for the same day of the week up to 14 days prior to the forecast period. For example, the feature corresponding to February 15th is set to the average of the sales figures y1 from February 1st and y8 from February 8th. This is to capture seasonal trends for each day of the week.

[0029] Feature No. 4 is a feature whose raw data is the prices xa1 to xa15. Feature No. 4 is an indicator of price fluctuation, and it calculates the ratio of the selling price during the forecast period to the maximum price during the observation period, including the day of the forecast period. For example, if the observation period is from February 1st to February 15th, and the maximum price during the observation period is MAX(xa1 to xa15), then feature No. 4 is expressed as xa15 / MAX(xa1 to xa15). This is done to quantify the degree of price decreases and increases and to capture price elasticity.

[0030] Feature No. 5 is a feature whose source data is the number of visitors xb1 to xb14. Feature No. 5 uses a lag to directly adopt the number of visitors from 7 days prior to the prediction period. For example, the feature corresponding to February 15th is set to the number of visitors on February 8th, xb8. This is to capture the trend for the same day of the previous week.

[0031] Features No. 6 to No. 8 are derived from date labels 2 / 1 to 2 / 15. Feature No. 6 converts the day of the week into a 7-bit One-Hot vector and provides categorical information indicating the day of the week as unordered data. For example, if the forecast period is Monday, it is given [1,0,0,0,0,0,0], and if the forecast period is Sunday, it is given [0,0,0,0,0,0,1]. Feature No. 7 is a flag indicating whether or not it is a public holiday. For example, it is set to 1 if the forecast period is a public holiday, and to 0 if it is not. Features No. 6 and No. 7 are intended to capture demand on specific days. Feature No. 8 is assigned by counting the number of days elapsed from the start date of observation of sales performance data to the forecast period. For example, if the first day of sales performance data is February 1st, the counter for February 15th will be 15. This is intended to capture long-term linear trends.

[0032] Feature No. 9 is calculated by multiplying the number of visitors on the previous day by the price on the current day, as an example of a combination of variables that is thought to have a synergistic effect on sales volume. For example, the feature corresponding to February 15th is the product of the number of visitors on the previous day, February 14th (xb14), and the price on the current day, February 15th (xa15). This is done to capture a nonlinear relationship.

[0033] In addition, for example, by calculating features using inventory levels as the source data, it is possible to capture events caused by stockouts. Also, by calculating features using the number of items displayed as the source data, it is possible to capture events related to the number of items displayed (for example, if the number of items displayed decreases, the visibility of the product decreases, and sales decrease). Furthermore, by calculating features using the number of competing products displayed as the source data, it is possible to capture events related to the competition for sales between competing products. And by calculating features using weather as the source data, it is possible to capture events related to temperature (for example, sales of hot drinks increase on cold days).

[0034] In this way, the generation unit 22 generates feature quantities that represent periodicity, seasonality, external factors, and interactions with sales volume by applying various preprocessing steps to the raw data. The generation unit 22 stores each of the generated feature quantities in the feature quantity storage unit 23.

[0035] The dividing unit 24 divides the sales performance data into a plurality of pieces of short-term data. The "short-term data" refers to a set of continuous or discontinuous periods in which sales trends are estimated to be the same or similar. An example of the dividing process by the dividing unit 24 will be described with reference to FIG. 5 and FIG. 6. Here, it is assumed that the sales performance data is tabulated as daily data from February 1 to February 9, and each block with a date represents an observation for one day. The short-term data shall include data for at least one day.

[0036] First, the dividing unit 24 determines the first optimal dividing point in the sales performance data. In FIG. 5, the first dividing point is indicated by a thick solid line. The dividing unit 24 enumerates all candidates for division patterns by sliding the dividing point in units of one day, and setting the parts before and after the dividing point as segment 1 and segment 2, respectively. In this example, since the observation period is 9 days, 8 patterns from division pattern 1 to division pattern 8 are candidates.

[0037] The dividing unit 24 calculates the cost of each candidate. An example of the calculation formula for the cost C in the case of one division (two segments) is shown in Formula (1). The calculation formula of the distance function R(i,j) included in the same formula is shown in Formula (2). Here, the set of daily data belonging to segment 1 is defined as P 1 . The set of daily data belonging to segment 2 is defined as P 2 . Each feature value (see FIG. 4) on date i is defined as f ai , f bi , ... . Each feature value on date j is defined as f aj , f bj , ... . When date i and date j are elements of the set P 1 (segment 1), the median of the distance function R(i,j) is defined as σ 1 . When date i and date j are elements of the set P 2 (segment 2), the median of the distance function R(i,j) is defined as σ 2 . In the following description related to the formulas, to avoid redundant explanations, the same symbols are used for items with the same definition, and partial duplicate explanations of the symbols are omitted.

[0038]

[0039] The distance function R(i,j) is the Euclidean distance in the feature space and represents the difference in sales trends between date i and date j (i.e., any date pair). Cost C is the sum of the sales trend distances for each date pair within the same segment, which are decayed exponentially. A smaller value means that the sales trends within the segment are more uniform (less variability). Therefore, the splitting unit 24 determines that the splitting point (splitting pattern) with the minimum cost C is optimal. In the example shown in Figure 5, we will explain assuming that the splitting point (splitting pattern 6) set between February 6th and February 7th is determined to be optimal.

[0040] Next, the division unit 24 determines the second optimal division point in the sales performance data. In Figure 6, the first optimal division point is shown with a thick solid line, and the second division point is shown with a thick dashed line. The division unit 24 fixes the first optimal division point and slides the second division point in one-day increments, setting segments 1 to 3 to enumerate all candidate division patterns. However, the first and second division points will not overlap. In this example, seven division patterns 6-1 to 6-7 are candidates as derivatives of division pattern 6.

[0041] The division unit 24 calculates the cost of each candidate. An example of the formula for calculating the cost C in the case of two divisions (three segments) is shown in equation (3). The formula for calculating R(i, j) in equation (3) is the same as in equation (2). Here, the set of daily data belonging to segment 3 is P 3 The dates i and j are set P. 3 The median of the distance function R(i,j) when it is an element of (segment 3) is σ 3 That is what they say.

[0042]

[0043] The division unit 24 determines the candidate with the minimum cost C as the second optimal division point. If the cost C at the second optimal division point is lower than the cost C at the first optimal division point, the division unit 24 proceeds to search for the third optimal division point. On the other hand, if the cost C at the second optimal division point is higher than the cost C at the first optimal division point, the division unit 24 terminates the division after only the first division.

[0044] In other words, if the cost C at the m-th optimal division point (where m is an integer greater than or equal to 2) is lower than the cost C at the (m-1)th optimal division point, the division unit 24 attempts to further reduce the cost C by recursively repeating the division. On the other hand, the division unit 24 stops the recursive process when the cost C no longer decreases, or when all segment lengths fall below a predetermined threshold (for example, when all segments contain only one day's worth of data). Note that the cost C from the third division onward can be generalized by increasing the corresponding terms, as can be seen from equations (1) and (3).

[0045] In this embodiment, the final segment determined by the above procedure is defined as "short-term data". The segmentation unit 24 stores information about the short-term data, that is, information indicating how the sales performance data was segmented (for example, a label for each short-term data), in the short-term data storage unit 25.

[0046] Note that the method of division is not limited to the procedure described above (the so-called binary search algorithm). For example, by using a decision tree algorithm to group events such as "sale days" into the same node, discontinuous periods such as February 3rd and February 6th can be assigned to the same segment (short-term data). Thus, short-term data does not necessarily include continuous sales periods. Alternatively, sales performance data could be simply divided into arbitrarily defined time intervals (e.g., every three days, and by day of the week) under the assumption that sales trends are identical or similar.

[0047] The merging unit 26 combines short-term data sets whose feature similarity is above a predetermined threshold to create combined data. Since feature similarity can also be considered similarity of sales trends, combined data can be considered a group of short-term data sets with similar sales trends. When the number of types of features extracted from sales performance data increases, sales trends can also take on various variations, so the division by the splitting unit 24 tends to become finer. As a result, the number of days included in each short-term data set decreases, and it may not be possible to secure the sample size necessary for training the prediction model. Therefore, the merging unit 26 generates combined data with a sufficient sample size by combining similar short-term data sets, and uses this as training data for the prediction model.

[0048] Specifically, the joining unit 26 performs a join determination by comparing the similarity of all pairs of short-term data. For example, as shown in Figure 7, suppose the sales performance data from February 1st to February 12th is divided into four short-term data sets 1 to 4 by the splitting unit 24. In this case, the joining unit 26 will determine whether to join six pairs of short-term data: (1,2), (1,3), (1,4), (2,3), (2,4), and (3,4).

[0049] In Figure 7, an "X" mark is placed on the arrows connecting each pair if the similarity is below the threshold, and a circle mark is placed if the similarity is above the threshold. In this example, if short-term data 1 and 3 are to be joined, and short-term data 3 and 4 are also to be joined, even if short-term data 1 and 4 are not directly to be joined, they are joined in a chain reaction, and a single joined data is generated that combines short-term data 1, 3, and 4. According to this join determination, even short-term data with discontinuous dates can be joined if their sales trends are similar. On the other hand, short-term data that is not similar to any pair may become a single joined data on its own.

[0050] Specifically, the merging unit 26 preferably merges short-term data sets where the similarity of the variances of the features is above a predetermined threshold, and the similarity of the mean of the features is above a predetermined threshold. In other words, the merging unit 26 may perform merge determination using two indicators: the similarity of the variance structure (spread) of the features and the similarity of the central position (mean) of the features. By merging only pairs that satisfy both of these criteria simultaneously, short-term data sets that are similar in both shape (variability) and position (mean) can be comprehensively merged.

[0051] Referring to equations (4) to (8), an example of a method for calculating the similarity of feature variances between short-term data s1 and short-term data s2 will be explained. Equation (4) is the 2x2 covariance matrix corresponding to the short-term data s1. Here, the features f in the short-term data s1 a The variance of Var s1 (f a ) is defined as the feature f in the short-term data s1. b The variance of Var s1 (f b ) is defined as the feature f in the short-term data s1. a and feature f b Covariance of Cov s1 (f a , f b ) is the formula. Equation (5) is the determinant of the covariance matrix and serves as an indicator of the magnitude of the variance structure. Note that while the explanation for equations (4) and (5) pertains to the short-term data s1, the values ​​and equations pertain to the short-term data s2 are the same.

[0052] Equation (6) is the Wilks lambda statistic, which is the value obtained by dividing the smaller of the two determinants by the larger of the two determinants. The closer the lambda statistic is to 1, the more similar the variance structures of the two determinants are, and the closer it is to 0, the greater the difference between the two determinants. Equation (7) is the lambda statistic approximated by the F statistic. Here, the number of dimensions of the feature is n. f (In this example, n f (2) The sample size (number of observation days) in the short-term data s1 is n 1 This is the case. The sample size (number of observation days) within the short-term data s2 is n. 2 That is what they say.

[0053] Equation (8) is the one-sided p-value for the F-statistic F, and represents the significance probability for the null hypothesis that the variance structures of the two short-term data points s1 and s2 are equal. A larger p-value means that there is no difference between the two short-term data points s1 and s2. F cdf This is the cumulative distribution function of the F-distribution.

[0054]

[0055] For example, the joint 26 has a p-value on one side. F If the confidence level (e.g., 5%) is above, the distribution structure between the two short-term data sets is considered to be similar, and it is decided to combine those short-term data sets. That is, the combining unit 26 determines the one-sided p-value p F The similarity of the variances of the features may be quantitatively evaluated using this method.

[0056] Referring to equations (9) to (13), an example of how to calculate the average similarity of features between short-term data s1 and short-term data s2 will be explained. Equation (9) is the Mahalanobis distance between the feature vector at date i and the average feature vector. A larger Mahalanobis distance means that the sales trend for that day is farther from the average in the short-term data. Here, the feature f at date i included in short-term data s1 ai , f bi The array (vector) of ... is Z i This is the case. The feature quantities f for all dates included in short-term data s1 and short-term data s2. a , f b The average array (vector) of ... is Z Ave This is the case. T represents the transpose matrix.

[0057] Equation (10) is an aggregate vector of Mahalanobis distances for each day. Equation (11) is a vector of absolute values ​​of the difference between the Mahalanobis distance and the median for each day. Here, D included in the short-term data s1 i D is the median of (see equation (9)). Mid、s1This is the explanation. Regarding equations (9) to (11), the explanation concerned the short-term data s1, but the same applies to the values ​​and equations related to the short-term data s2.

[0058] Equation (12) is the Welch F-statistic, which is the standardized value of the difference between the mean vectors of each short-term data s1 and s2 by the variance. A smaller W means a smaller mean difference (i.e., the centers are more similar). Here, Y corresponds to the short-term data s1. s1 The average of each element in (see equation (11)) is Y Ave、s1 This is the case. Y corresponds to the short-term data s2. s2 The average of each element is Y Ave、s2 This is the case. Y corresponds to the short-term data s1. s1 , and Y corresponding to the short-term data s2 s2 The overall average of each element is Y Ave That is what they say.

[0059] Equation (13) is the one-sided p-value for the F-statistic W, and represents the significance probability for the null hypothesis that the central positions (means) of the two short-term data points s1 and s2 are equal. A larger p-value means that there is no difference between the two short-term data points s1 and s2.

[0060]

[0061] For example, the joint 26 has a p-value on one side. W If the confidence level (e.g., 5%) is above, the center positions of the two short-term data points are considered to be similar, and it is decided to combine those short-term data points. That is, the combining unit 26 determines the one-sided p-value p W The average similarity of the features may be quantitatively evaluated using this method. The merging unit 26 stores information about the merged data, that is, information indicating how the short-term data were merged (for example, a label for each merged data), in the merged data storage unit 27.

[0062] The learning unit 28 uses the combined data as training data to train a predictive model that predicts the number of product sales. The training data includes, for example, a combination of features and sales figures generated for each aggregation unit of sales performance data (e.g., one day) and representative values ​​of the features within the combined data to which that data belongs (e.g., mean, median, maximum, and minimum). In other words, by training with a two-layer structure of features for each aggregation unit and features for the combined data unit, it is possible to train the predictive model to simultaneously learn a stable sales trend within the combined data while retaining variability at the aggregation unit level. This reduces noise while ensuring effective degrees of freedom, thereby improving prediction accuracy and generalization performance.

[0063] As prediction models, machine learning models such as convolutional neural networks, recurrent neural networks, multilayer perceptrons, and gradient boosting decision trees can be applied as appropriate. The learning unit 28 stores the trained prediction models in the prediction model storage unit 29.

[0064] The prediction unit 30 predicts the number of products sold during the prediction period using a trained prediction model. Specifically, the prediction unit 30 inputs the same features used in training during the training period related to the prediction period into the prediction model, thereby obtaining the number of products sold during the prediction period as output. The features corresponding to the prediction period can be generated using sales performance data up to the day before the prediction period, and elements that can be known in advance from each element on the day of the prediction period, as described above. The "number of products sold" predicted by the prediction device 10 may be the number of products sold or the number of products sold for each category to which each product belongs. The "prediction period" is a period during which the number of product sales is unknown, and this period may be on a daily, hourly, or weekly basis.

[0065] (Modification 1) To further improve prediction accuracy, training may be performed with emphasis on combined data where the prediction period and sales trend are similar. Specifically, the combining unit 26 may associate the prediction period with any of the combined data based on the similarity between the feature corresponding to the prediction period and the feature extracted from the combined data. The similarity between features can be determined using at least one of the similarity of the variances of the features and the similarity of the means of the features, as described above.

[0066] The learning unit 28 may train the prediction model by making the weights of the linked data associated with the prediction period larger than the weights of the other linked data. For example, as shown in Figure 8, suppose February 15th, which is part of the prediction period, is associated with linked data 4, and February 16th, which is also part of the prediction period, is associated with linked data 2. In this case, the learning unit 28 makes the weight coefficients of linked data 2 and 4 larger than the weight coefficients of linked data 1 and 3. For example, the weight coefficients of linked data 1 and 3 may be set to 1, and the weight coefficients of linked data 2 and 4 may be empirically set in the range of 2 to 5, and if overfitting is observed, they may be reduced by cross-validation. Also, if the prediction period changes, the learning unit 28 may reset the weight coefficients and retrain the prediction model.

[0067] (Modification 2) When using weighted learning for each combined data set as in Modification 1, the training data consists only of combinations of features and sales figures generated for each aggregation unit of sales performance data (e.g., one day), and representative values ​​of features within the combined data may be omitted. In other words, the combined data may be used only for calculating the weight coefficients. In this case, while improving prediction accuracy through weighted learning, the prediction model itself operates with simple inputs, thus reducing the amount of computation.

[0068] Next, the operation of the prediction device 10 will be described. Figure 9 is a flowchart showing the flow of the prediction process by the prediction device 10. The CPU 11 reads the prediction program from the ROM 12 or storage 14, loads it into the RAM 13, and executes it, thereby performing the prediction process. The prediction process is an example of the prediction method of this disclosure.

[0069] In step S10, the CPU 11, acting as the acquisition unit 20, acquires sales performance data. In step S12, the CPU 11, acting as the generation unit 22, generates feature quantities from the sales performance data. In step S14, the CPU 11, acting as the division unit 24, divides the sales performance data into multiple short-term data sets.

[0070] In step S16, the CPU 11, acting as a merging unit 26, merges short-term data sets whose feature similarity is above a predetermined threshold to create merged data. In step S18, the CPU 11, acting as a learning unit 28, uses the merged data as training data to train a prediction model that predicts the number of product sales. In step S20, the CPU 11, acting as a prediction unit 30, uses the trained prediction model to predict the number of product sales during the prediction period. Once step S20 is completed, the prediction process ends.

[0071] As described above, the prediction device 10 according to this embodiment includes an acquisition unit 20 that acquires sales performance data representing a time series of sales performance including product features and sales figures, a generation unit 22 that generates features from the sales performance data, a division unit 24 that divides the sales performance data into a plurality of short-term data, a joining unit 26 that combines short-term data where the similarity of the features extracted from the short-term data is above a predetermined threshold to form joined data, a learning unit 28 that learns a prediction model to predict the sales figures of a product using the joined data as learning data, and a prediction unit 30 that uses the prediction model to predict the sales figures of a product during the prediction period.

[0072] The prediction device 10 can combine short-term data sets that are estimated to have similar sales trends, thereby ensuring a sufficient number of samples within the combined data (group). This stabilizes the mean and variance of each combined data set, reducing variability when training the prediction model. Furthermore, it avoids insufficient training data even if the prediction model has many explanatory variables (features). Consequently, the accuracy of sales volume prediction can be improved.

[0073] (Examples) Examples of using the technology of the present disclosure will be described with reference to Figures 10 to 12. Figure 10 shows combined data 1 to 6 generated by the technology of the present disclosure based on sales performance data including the number of units sold and the average price for each day. In the graph of Figure 10, the dotted line shows the trend of the number of units sold each day, and the solid line shows the trend of the average price each day. The combined data No. is shown at the top of the graph. In Figure 10, identical combined data are color-coded so that they have the same background color.

[0074] Figure 11 shows a comparative example in which sales performance data similar to that in Figure 10 is simply divided into short-term data, with dashed lines indicating the boundaries between the short-term data. The method of dividing into short-term data is the same as the division process by the division unit 24. As can be seen from Figures 10 and 11, it can be seen that discontinuous short-term data are combined using the technology of this disclosure.

[0075] Figure 12 shows the difference between predicted sales figures and actual sales figures for both an example and a comparative example using the technology of this disclosure. The example is shown with a solid line as the "proposed method," one of the comparative examples, a method using "machine learning only," is shown with a dashed line, and another comparative example, a method using "machine learning + period splitting," is shown with a dashed-dotted line. The method using "machine learning only" is a method that uses actual sales data directly as training data. The method using "machine learning + period splitting" is a method that uses short-term data split in the same way as the splitting process by the splitting unit 24 as training data. As can be seen particularly in area A of Figure 12, it can be seen that the accuracy of predictions is improved by the technology of this disclosure.

[0076] In addition, the prediction processing that the CPU reads and executes in each of the above embodiments may be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) whose circuit configuration can be changed after manufacturing, such as FPGAs (Field-Programmable Gate Arrays), and dedicated electrical circuits that have a circuit configuration specifically designed to execute a particular process, such as ASICs (Application Specific Integrated Circuits). Furthermore, the learning process and / or analysis process may be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0077] Furthermore, although the above embodiment describes a configuration in which the prediction program is pre-stored (installed) in ROM 12 or storage 14, the invention is not limited to this. The program may be provided in a form stored on a non-transitor storage medium such as CD-ROM (Compact Disk Read Only Memory), DVD-ROM (Digital Versatile Disk Read Only Memory), and USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that is downloaded from an external device via a network.

[0078] The technology of this disclosure can also be appropriately combined with the above-described embodiments and modifications. The descriptions and illustrations shown above are detailed explanations of the parts relating to the technology of this disclosure and are merely examples of the technology of this disclosure. For example, the above descriptions of the configuration, function, operation, and effect are examples of the configuration, function, operation, and effect of the parts relating to the technology of this disclosure. Therefore, it goes without saying that unnecessary parts may be deleted, new elements added, or replaced from the descriptions and illustrations shown above, as long as they do not deviate from the spirit of the technology of this disclosure.

[0079] The following additional information is disclosed regarding the embodiments described above.

[0080] (Note 1) A prediction device comprising: a memory; and at least one processor connected to the memory, wherein the processor acquires sales performance data representing a time series of sales performance including product features and sales volume; generates features from the sales performance data; divides the sales performance data into a plurality of short-term data; combines the short-term data where the similarity of the features extracted from the short-term data is above a predetermined threshold to form combined data; uses the combined data as training data to train a prediction model that predicts the sales volume of the product; and uses the prediction model to predict the sales volume of the product during the prediction period.

[0081] (Appendix 2) A non-temporary storage medium storing a program executable for performing prediction processing by a computer, wherein the prediction processing includes: acquiring sales performance data representing a time series of sales performance including product features and sales volume; generating features from the sales performance data; dividing the sales performance data into a plurality of short-term data; combining the short-term data where the similarity of the features extracted from the short-term data is above a predetermined threshold to form combined data; training a prediction model to predict the sales volume of the product using the combined data as training data; and predicting the sales volume of the product during the prediction period using the prediction model.

[0082] 10 Prediction device 11 CPU 12 ROM 13 RAM 14 Storage 15 Input unit 16 Display unit 17 Communication interface 19 Bus 20 Acquisition unit 21 Sales performance data storage unit 22 Generation unit 23 Feature quantity storage unit 24 Splitting unit 25 Short-term data storage unit 26 Combining unit 27 Combined data storage unit 28 Learning unit 29 Prediction model storage unit 30 Prediction unit

Claims

1. A prediction device comprising: an acquisition unit that acquires sales performance data representing a time series of sales performance including product features and sales figures; a generation unit that generates features from the sales performance data; a division unit that divides the sales performance data into a plurality of short-term data sets; a joining unit that combines short-term data sets whose similarity of features extracted from the short-term data is above a predetermined threshold to form combined data; a learning unit that uses the combined data as training data to train a prediction model that predicts the sales figures of the product; and a prediction unit that uses the prediction model to predict the sales figures of the product during the prediction period.

2. The prediction device according to claim 1, wherein the coupling unit associates the prediction period with one of the combined data based on the similarity between the feature corresponding to the prediction period and the feature extracted from the combined data, and the learning unit learns the prediction model by making the weight of the combined data associated with the prediction period greater than the weight of the other combined data.

3. The prediction device according to claim 1, wherein the coupling unit combines short-term data sets in which the similarity of the variances of the features is equal to or greater than a predetermined threshold, and the similarity of the mean of the features is equal to or greater than a predetermined threshold.

4. A computer-based prediction method comprising: acquiring sales performance data representing a time series of sales performance including product features and sales figures; generating features from the sales performance data; dividing the sales performance data into multiple short-term data sets; combining short-term data sets whose similarity of features extracted from the short-term data sets is above a predetermined threshold to form combined data; training a prediction model to predict the sales figures of the product using the combined data as training data; and using the prediction model to predict the sales figures of the product during the prediction period.