Forecast demand correction device and forecast demand correction method

Abstract

By adjusting predicted sales figures to account for complex price elasticity that changes with the product lifecycle and seasons, we can forecast product demand with high accuracy. [Solution] The present invention provides a forecast demand correction device that takes forecast sales volume and sales price as inputs and outputs a forecast sales volume after correction. This device creates multiple learning correction functions by learning them using forecast sales volume, sales price, and actual sales volume data, which are training data for the forecast sales volume after correction, for past generations of products. From the multiple learning correction functions, a selection correction function is selected based on their prediction accuracy. The forecast sales volume and sales price for future generations of the product to be predicted are input to the selected selection correction function, and the forecast sales volume after correction output by the selected selection correction function is obtained.

Description

Technical Field

[0001] The present invention relates to a predicted demand correction device and a predicted demand correction method for predicting the demand of mass-produced products in which product series with similar basic specifications such as size continue due to the generation change between old products and new products which are successor products of the old products.

Background Art

[0002] For manufacturers, it is important to appropriately formulate sales strategies such as product prices and product launch times in order to maximize profits. In order to formulate and evaluate sales strategies, high-precision demand prediction based on the sales strategy is required.

[0003] The demand prediction device of Patent Document 1 predicts the demand for old products (obsolete products) and new products. Patent Document 1 describes that "a performance value of combined-generation demand obtained by combining the performance values of the demand for obsolete products and the demand for new products in time series is generated, and the prediction of the combined-generation demand is performed using the combined-generation demand performance value and generation change parameters including the release interval from when the obsolete product is released until the new product is released and the change period required for the generation change between the obsolete product and the new product, and from the prediction result of the combined-generation demand, it is separated into the predicted demand values for each of the obsolete product and the new product using the generation change parameters."

[0004] The apparatus described in Patent Document 2 corrects the predicted market share of the next model product by the market share improvement rate resulting from a model change from an old model to a new model, and predicts the demand for the next model product. Patent Document 2 states that it "includes means for storing utility values ​​for each level belonging to each attribute that constitutes the product for each consumer, means for calculating and storing the predicted market share of the current model product, means for calculating and storing the predicted market share of the next model product, means for calculating and storing the predicted market share improvement rate due to the model change, and means for calculating and storing the demand for the next model product for each unit period by multiplying the predicted market share improvement rate due to the model change to the next model product, the actual market share stored in the storage means, the adaptation function stored in the storage means, and the seasonal variation adjustment coefficient stored in the storage means." [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2024-58984 [Patent Document 2] Japanese Patent Publication No. 2007-316899 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] The demand forecasting device described in Patent Document 1 forecasts combined generational demand using generational change parameters, and then separates the forecasted demand values ​​for older and newer products. However, while the demand forecasting device in Patent Document 1 uses sales volume-based information, it does not handle value-related information such as sales price. Therefore, a separate measure was needed to consider the difference in value between older and newer products.

[0007] The device described in Patent Document 2 stores utility values ​​for each level belonging to each attribute that constitutes a product for each consumer, and corrects the predicted market share of the next model product based on the utility values. Because the relationship between utility values ​​and sales price is not clear, the device described in Patent Document 2 only performs demand forecasting without considering the effect of price elasticity (correlation between sales volume and sales price) on the time series sales volume from the time series sales price.

[0008] Therefore, the present invention aims to predict product demand with high accuracy by correcting the predicted sales volume by taking into account the complex price elasticity that changes with the product life cycle, seasons, etc. [Means for solving the problem]

[0009] The present invention provides a correction function learning processing unit that creates a plurality of learning correction functions by learning a correction function that takes a predicted sales volume and sales price as input and outputs a corrected sales volume prediction, using the predicted sales volume, sales price, and actual sales volume values ​​which are training data for the corrected sales volume prediction of past generation products in past periods; a correction function selection processing unit that inputs the predicted sales volume and sales price of the target generation product in past periods to each of the plurality of learning correction functions, calculates the prediction accuracy based on the difference between the corrected sales volume prediction output by each of the plurality of learning correction functions and the actual sales volume values ​​of the target generation product in past periods, and selects a selected correction function from the plurality of learning correction functions based on the calculated prediction accuracy; and a correction function utilization processing unit that inputs the predicted sales volume and sales price of the target generation product in future periods to the selected selected correction function and obtains the corrected sales volume prediction output by the selected selected correction function. Other means will be described within the descriptions of embodiments for carrying out the invention. [Effects of the Invention]

[0010] According to the present invention, by correcting the predicted sales volume to account for complex price elasticities that change with the product lifecycle, seasons, etc., it is possible to predict product demand with high accuracy.

Brief Description of the Drawings

[0011] [Figure 1] It is a diagram showing the configuration of the predicted demand correction device. [Figure 2] It is a diagram showing an example of sales-related information before the learning process. [Figure 3] It is a diagram showing an example of sales-related information after the learning process. [Figure 4] It is a diagram showing an example of sales-related information before the selection process. [Figure 5] It is a diagram showing an example of sales-related information after the selection process. [Figure 6] It is a diagram showing an example of sales-related information before the utilization process. [Figure 7] It is a diagram showing an example of sales-related information after the utilization process. [Figure 8] It is a diagram showing an example of correction function information after the learning process. [Figure 9] It is a diagram showing an example of correction function information after the selection process. [Figure 10] It is a diagram showing an example of correction function information after the utilization process. [Figure 11] It is a flowchart of the processing procedure. [Figure 12] It is a diagram showing an example of the display screen.

Embodiments of the Invention

[0012] Hereinafter, embodiments of the present invention (referred to as "the present embodiments") will be described in detail with reference to the drawings. The product in the present embodiments is a mass-produced product in which generation replacements are continuously carried out in the same product grade.

[0013] For example, in the case of a refrigerator, various product grades are set according to the target user group, such as "for a two-person family, the upper door opens to one side, the middle and lower parts are drawer-type, and the capacity is 300 liters". Focusing on individual product grades, model changes such as "First Generation", "Second Generation",... are carried out every few years. The new-generation refrigerator has an appearance that seems the same as that of the old-generation refrigerator at first glance. However, the new-generation refrigerator is designed to reflect the latest technology and is superior to the old-generation refrigerator in terms of, for example, power consumption and quietness.

[0014] As another example, in the case of a central control unit (CPU) of a computer, various product grades are set according to the target user group, such as "for a notebook computer, a high-end model for creators, multi-core". Focusing on individual product grades, model changes such as "First Generation", "Second Generation",... are carried out every few years. The new-generation CPU is designed to reflect the latest technology and is superior to the old-generation CPU in terms of, for example, computing speed and power consumption.

[0015] The products of this embodiment include finished products represented by refrigerators and components represented by CPUs. There are multiple manufacturers participating in the market for such products, and there are competing products. The manufacturer continues to sell the old-generation products (obsolete products) for some time after the start of selling the new-generation products.

[0016] (Configuration of the predicted demand correction device) FIG. 1 is a diagram showing the configuration of the predicted demand correction device 1. The predicted demand correction device 1 is a general computer. The predicted demand correction device 1 includes a central control unit 11, an input device 12 such as a keyboard and a mouse, an output device 13 such as a display and a speaker, a main storage device 14, an auxiliary storage device 15, and a data input / output device 16 that exchanges data with peripheral devices (not shown). These are interconnected by a bus.

[0017] The auxiliary storage device 15 stores sales-related information 31 and correction function information 41. The sales-related information 31 includes sales-related information before learning processing 32a, sales-related information after learning processing 32b, sales-related information before selection processing 33a, sales-related information after selection processing 33b, sales-related information before utilization processing 34a, and sales-related information after utilization processing 34b. The correction function information 41 includes correction function information after learning processing 42, correction function information after selection processing 43, and correction function information after utilization processing 44. The auxiliary storage device 15 also stores multiple learning correction functions 51 and selection correction functions 52. Details of these will be described later.

[0018] In the main memory 14, the correction function learning processing unit 21, the correction function selection processing unit 22, the correction function utilization processing unit 23, the revenue processing unit 24, and the display processing unit 25 are programs. In the following description, when the processing entity is indicated as "the ○○ unit," it means that the central control unit 11 reads each program from the auxiliary storage device 15 into the main memory 14 and executes the information processing (details described later) that is pre-written in each program. The auxiliary storage device 15 may be configured within the housing of the forecast demand correction device 1, as shown in Figure 1, or it may be configured in a separate housing from the forecast demand correction device 1.

[0019] (Sales-related information) Figures 2 to 7 show examples of sales-related information 31. Sales-related information 31 includes, as its "transition forms," ​​the aforementioned pre-learning sales-related information 32a, post-learning sales-related information 32b, pre-selection sales-related information 33a, post-selection sales-related information 33b, pre-utilization sales-related information 34a, and post-utilization sales-related information 34b. Each of these transition forms is a so-called relational database, sharing the same column structure. However, the values ​​in some columns change depending on the processing progression. This processing progression is expressed by the terms "pre-learning," "post-learning," ..., and "post-utilization." The representation format of sales-related information 31 is arbitrary and may include, for example, NoSQL (the same applies to the correction function information 41 described later).

[0020] (Information related to sales before learning processing) Figure 2 shows an example of pre-learning sales-related information 32a. In pre-learning sales-related information 32a, the product generation serial number is stored in the product generation serial number column 102, the product generation type is stored in the product generation type column 103, the period classification is stored in the period classification column 104, the week start date is stored in the week start date column 105, the time elapsed since launch is stored in the time elapsed since launch column 106, the predicted sales volume is stored in the sales volume forecast value column 107, the corrected sales volume forecast value is stored in the sales volume corrected forecast value column 108, the actual sales volume is stored in the actual sales volume column 109, the sales price (current generation) is stored in the sales price (current generation) column 110, the sales price (previous generation) is stored in the sales price (previous generation) column 111, the sales price (competitor product) is stored in the sales price (competitor product) column 112, and the learning correction function is stored in the learning correction function column 113.

[0021] The product model in the product model field 101 is an identifier that uniquely identifies the product model. For example, in "PDR-K1", "PDR" may identify the product grade, and "K1" may identify the generation. The product generation serial number in the product generation serial number column 102 represents the generation of the product, starting with "1" to indicate the "first generation," and continuing with "2, 3, ....". The product generation type in the product generation type column 103 is either "past generation," meaning a generation prior to the latest generation, or "target generation," meaning the latest generation for which the sales volume-adjusted forecast value should be calculated. In Figure 2, the product generation type is "past generation."

[0022] The period classification in the period classification column 104 is either a "past period," meaning that the week start date described later corresponds to the past, or a "future period," meaning that it corresponds to the future. In Figure 2, the period classification is "past period." The week start date in column 105 is the first day of the unit period (one week in the example in Figure 2) in which the quantities and prices described below are stored and calculated. The week start date in Figure 2 is a past date. The "Time elapsed since launch" column 106 indicates the number of weeks in the week that the product's launch date corresponds to.

[0023] The sales volume forecast value in column 107 is the number of units of a product that the manufacturer has artificially predicted in advance, and it is the total value for that unit period. The sales volume forecast value may also be the number of units predicted by another information processing device without considering the sales price. The corrected sales volume forecast value in column 108 is the sales volume of the product predicted by the correction function learning processing unit 21 using the learned correction function 51. In Figure 2, the corrected sales volume forecast value is "-" (no data). This means that the corrected sales volume forecast value has not yet been calculated. The sales figures in column 109 represent the actual sales figures for the product, compiled retrospectively by the manufacturer, and represent the total sales for that period. The manufacturer obtains the sales figures from POS information.

[0024] The "Current Sales Price (Current Generation)" column (110) shows the actual price (unit price, hereafter the same) of the current generation product sold in the market. The longer the time elapsed since launch, the lower the "Current Sales Price (Current Generation)" tends to be. Manufacturers obtain the "Current Sales Price (Current Generation)" from POS information. The "Selling Price (Previous Generation)" column 111 represents the price of the product from the generation immediately preceding the current generation that was actually sold in the market. In Figure 2, the current generation is the "1st Generation." In this case, the selling price (previous generation) is the price of the product that the manufacturer determines to be the immediate preceding generation of the 1st Generation. As the value at the time of release increases, the selling price (previous generation) also tends to decrease. However, at the same point in time, the selling price (previous generation) is rarely higher than the selling price (current generation). The manufacturer obtains the selling price (current generation) from POS information. The "Selling Price (Competitor Products)" column 112 shows the actual market price of competing products of the current generation. Competitor products are similar products sold by other manufacturers targeting the same user base. Manufacturers conduct market research to obtain this information on competing products.

[0025] The learning correction function in the learning correction function column 113 is an identifier that uniquely identifies the learning correction function 51, which was trained using data from release time points 1 to 12. In Figure 2, the learning correction function is "-" (no data). This means that the learning correction function 51 has not yet been trained.

[0026] (Sales-related information after learning process) Figure 3 shows an example of sales-related information 32b after the learning process. Figure 3 differs from Figure 2 in the following respects. • The corrected sales volume forecast column 108 stores a specific value. This is stored by the correction function learning processing unit 21 as the output from the learning correction function 51 (details below). • The learning correction function column 113 stores a specific learning correction function. This means that some kind of learning correction function has been learned.

[0027] The correction function learning processing unit 21 creates n (n>1) learning correction functions 51 after changing the learning conditions (details below), and creates n post-learning sales-related information 32b. When focusing on one post-learning sales-related information 32b, the "Learning Result ○" stored in the learning correction function column 113 of all records (rows) is the same. When focusing on multiple post-learning sales-related information 32b simultaneously, the "Learning Result ○" differs from one another, and the corrected predicted sales volume also differs from one another.

[0028] (Sales-related information before selection processing) Figure 4 shows an example of sales-related information 33a before selection processing. Figure 4 differs from Figure 3 in the following respects. • The product generation serial number field 102 contains the value "2". • The "Predicted Generation" is stored in the Product Generation Type field 103. The week start date in column 105 is approximately one year later than the date in Figure 3. In other words, the seasons or product life cycles for launch points "1", "2", "3", etc. in Figure 4 correspond to the seasons or product life cycles (in this example, the start of the new school year) for launch points "1", "2", "3", etc. in Figure 3. • The corrected sales forecast value column 108 contains a "-". This means that the corrected sales forecast value has not yet been calculated. The values ​​in the "Forecast Sales Volume" column (107), "Actual Sales Volume" column (109), "Current Generation Sales Price" column (110), "Previous Generation Sales Price" column (111), and "Competitor's Product Sales Price" column (112) naturally differ from those in Figure 3, since the product generation serial numbers are different.

[0029] The correction function selection processing unit 22 corresponds to n (n>1) learned correction functions 51 and creates n pre-selection sales-related information 33a.

[0030] (Sales-related information after selection process) Figure 5 shows an example of sales-related information 33b after selection processing. Figure 5 differs from Figure 4 in the following respects. • The corrected sales volume forecast value column 108 stores a specific value. This is stored by the correction function selection processing unit 22 as the output from the learned correction function 51 (learning result 2) (details below).

[0031] The correction function selection processing unit 22 corresponds to n (n>1) learned correction functions 51 and creates n post-selection sales-related information 33b.

[0032] (Sales-related information before processing) Figure 6 shows an example of pre-processing sales-related information 34a. Figure 6 differs from Figure 5 in the following respects. • The period classification column 104 contains the entry for “Future Period”. • The week start date column (105) stores a future date in the Western calendar, based on the current time. • The week start date in column 105 is approximately two months later than the date in Figure 5. The release points "10," "12," "13," etc., are approximately two months after the release points "1," "2," "3," etc., shown in Figure 5. • The corrected sales forecast value column 108 contains a "-". This means that the corrected sales forecast value has not yet been calculated. • The sales volume data for column 109 contains a "-". This means that it is impossible to calculate sales volume data for future periods (no data exists). • The value of 110 in the "Sales Price (Current Generation)" column has decreased compared to Figure 5. This is because the value at the point of release has increased compared to Figure 5. • The value in the "Sales Price (Previous Generation)" column (111) has decreased compared to Figure 5. This is because the value at the point of release has increased compared to Figure 5. • In the learning correction function column 113, "Learning Result 2" is the selection result with the highest prediction accuracy.

[0033] The correction function utilization processing unit 23 corresponds to one selected correction function 52 and creates one pre-utilization sales-related information 34a.

[0034] (Information related to sales after processing) Figure 7 shows an example of post-processing sales-related information 34b. Figure 7 differs from Figure 6 in the following respects. • The corrected sales volume forecast column 108 stores a specific value. This is stored by the correction function utilization processing unit 23 as the output from the selected correction function 52 (learning result 2) (details below).

[0035] The correction function utilization processing unit 23 corresponds to one selected correction function 52 and creates one post-utilization sales-related information 34b.

[0036] (Correction function) The correction function (not shown) in this embodiment is, for example, a four-layer neural network. The correction function has the weights W of each layer as parameters. ij (Weight between the output in row i and the input in column j), and bias bi It has a bias (of the output of the i-th row). The correction function is learned when the weight W ij and bias b i The goal is to optimize the correction function, specifically the weights W as parameters. ij and bias b i The optimized version is the "learning correction function 51," and among the multiple learning correction functions 51, the one specifically selected for its high prediction accuracy in certain cases is the "selection correction function 52" ("learning result 2" in Figures 6 and 7).

[0037] The input data to the correction function (before learning) consists of the sales volume forecast, sales price (current generation), sales price (previous generation), and sales price (competitor product) from the sales-related information 32a before the learning process. The output data from the correction function is the corrected sales volume forecast from the sales-related information 32b after the learning process. A feature of this embodiment is that the input data includes both the number of units and the price. Such a correction function can reflect the price elasticity (correlation between sales volume and sales price) that time-series sales prices have on time-series sales volume.

[0038] Generally, when a function predicts a certain value, the input data (input variables) of that function are known, while the output data (output variables) are often unknown. In that sense, it may seem unfamiliar at first glance that a value containing the word "prediction" is both the input data and the output data of the correction function. However, as mentioned above, in this embodiment, the predicted sales volume is known, while the predicted sales volume after correction is unknown. And the idea that "a computer corrects an artificially predicted value" is reflected in the word "correction."

[0039] Since the output data of the correction function is the corrected sales volume forecast, and learning is performed to improve the prediction accuracy of that corrected sales volume forecast, the correct labels (training data) for the corrected sales volume forecast output by the correction function are required during the learning phase. This training data is the actual sales volume value from the sales-related information 32a before the learning process.

[0040] The input data for the learning correction function 51 (after learning) consists of the sales volume forecast, sales price (current generation), sales volume (previous generation), and sales price (competitor product) from the pre-selection sales-related information 33a. The output data from the learning correction function 51 is the corrected sales volume forecast from the post-selection sales-related information 33b. The actual sales volume from the pre-selection sales-related information 33a is used in the same way as the training data in terms of information processing. However, the purpose is to select a selection correction function 52 with high prediction accuracy from among the multiple learning correction functions 51, depending on a specific case (e.g., season).

[0041] The input data for the selection correction function 52 consists of the sales volume forecast, sales price (current generation), sales volume (previous generation), and sales price (competitor product) from the pre-use sales-related information 34a, while the output data from the selection correction function 52 is the sales volume forecast after correction from the post-use sales-related information 34b.

[0042] (Correction function information) Figures 8 to 10 show examples of correction function information 41. As mentioned above, the correction function information 41 includes correction function information after learning processing 42, correction function information after selection processing 43, and correction function information after utilization processing 44 as its "transition forms". Each of the transition forms is a so-called relational database and shares the same column structure. However, the values ​​of some columns change depending on the progression of processing. This progression of processing is expressed by the terms "after learning processing", "after selection processing", and "after utilization processing".

[0043] (Information on correction function after learning process) Figure 8 shows an example of post-learning correction function information 42. In the post-learning correction function information 42, the product generation serial number is stored in the product generation serial number field 152, the product generation type is stored in the product generation type field 153, the period classification is stored in the period classification field 154, the learning correction function is stored in the learning correction function field 155, the correction function selection flag is stored in the correction function selection field 156, the prediction accuracy is stored in the prediction accuracy field 157, the number of layers is stored in the number of layers field 158, information about layer 1 is stored in the layer 1 field 1591, ..., and information about layer 4 is stored in the layer 4 field 1594.

[0044] The explanation for Product Model Column 151 to Learning Correction Function Column 155 is the same as the explanation for Product Model Column 101 to Period Classification Column 104 and Learning Correction Function Column 113 in Figure 2. In Figure 8, the Product Generation Serial Number is "1", the Product Generation Type is "Past Generation", and the Period Classification is "Past Period". The correction function selection flag in the correction function selection field 156 is either "Selected," meaning the learned correction function was selected as the selected correction function, or "-," meaning it was not selected. In Figure 8, the correction function selection flag is "-." The prediction accuracy in column 157 is "100 - Mean Absolute Percentage Error". Prediction accuracy may simply be the Mean Absolute Percentage Error, or the Root Mean Squared Error may be used instead. The Mean Absolute Percentage Error is the ratio of the average of the absolute values ​​of the difference between the corrected predicted sales value and the actual sales value to the average of the actual sales value. The Root Mean Squared Error is the ratio of the average of the square roots of the squares of the difference between the corrected predicted sales value and the actual sales value to the average of the actual sales value. The number of levels in the "Number of Levels" column, 158, represents the number of levels in the learning correction function 51.

[0045] The Layer 1 column 1591 contains information about the first layer from the input side. The Output Count column 1591a stores the number of outputs, the Input Count column 1591b stores the number of inputs, the Weight W11 column 1591c stores the weights, the Bias b1 column 1591d stores the biases, and the Activation Function column 1591e stores the activation function. The output count in the output count column 1591a represents the number of data points that the first layer outputs to the second layer (the number of neurons in the second layer). The input number in the input number field 1591b is the number of data points accepted by the first layer. Note that the input number corresponds to the number (dimensions) of input data in the correction function. For example, if the input data is 3-dimensional, consisting of "predicted sales volume, sales price (current generation), and sales price (previous generation)", the input number for layer 1 will be "3", as shown in Figure 8. If the input data is 4-dimensional, consisting of "predicted sales volume, sales price (current generation), sales price (previous generation), and sales price (competitor product)", the input number for layer 1 will be "4".

[0046] The weight in column W11, 1591c, is the weight of the first layer. As mentioned above, weight W ij This is defined as "the weight between the output in row i and the input in column j". Therefore, a 3-dimensional weight vector is stored here across 50 columns (represented as "..."). The bias in column b1, 1591d, is the bias of the first layer. As mentioned above, bias b i This is defined as the "bias of the i-th line output". Therefore, a 50-dimensional bias is stored here. The activation function in the activation function column 1591e is the type of activation function provided by the learning correction function 51 as a neural network. The activation function is defined for each neuron in the neural network and expresses the relationship between input and output values. Depending on which range of input values ​​you want to emphasize, there are activation functions such as the ReLU (Rectified Linear Unit) function and the sigmoid function.

[0047] Column 1594 of Layer 4 contains information about the fourth layer (final layer) from the input side, and its explanation is the same as the explanation for Column 1591 of Layer 1. Note that the number of outputs corresponds to the number (dimensions) of output data of the correction function. Since the output data is one-dimensional, "predicted sales volume after correction," the number of outputs for Layer 4 is "1," as shown in Figure 8. Figure 8 shows the post-learning correction function information 42 separated into two rows, upper and lower, which is solely due to space limitations (the same applies to Figures 9 and 10). When the number of hierarchies is m, the post-learning correction function information 42 consists of columns for hierarchies 1 to m, and becomes quite long horizontally.

[0048] (Information on correction function after selection processing) Figure 9 shows an example of the post-selection correction function information 43. Figure 9 differs from Figure 8 in the following respects. • The product generation serial number field 152 contains the value "2". • The "Predicted Generation" is stored in the Product Generation Type field 153. • There is a record (row) in the correction function selection field 156 where "Selected" is stored. The learning correction function 51 corresponding to "Learning Result 2" is selected. This is because the prediction accuracy of Learning Result 2 is higher than the others.

[0049] (Information on correction function after processing) Figure 10 shows an example of post-processing correction function information 44. Figure 10 differs from Figure 9 in the following respects. • The period classification column 154 contains the entry for "Future Period". • The prediction accuracy field 157 contains the value "-". Prediction accuracy evaluates the error (the difference between the predicted value and the actual value) retrospectively from the present time; therefore, there is no prediction accuracy for future periods.

[0050] (Processing procedure) Figure 11 is a flowchart of the processing procedure. As a prerequisite for starting the processing procedure, it is assumed that the pre-learning processing sales-related information 32a, pre-selection processing sales-related information 33a, and pre-utilization processing sales-related information 34a are stored in the auxiliary storage device 15 in the state shown in Figures 2, 4, and 6, respectively.

[0051] In step S201, the correction function learning processing unit 21 of the forecast demand correction device 1 accepts a start command input by the user via the input device 12. The correction function learning processing unit 21 may start automatically in the event of a predetermined event, or it may start automatically at a predetermined date and time (batch processing), without waiting for a start command from the user.

[0052] In step S202, the correction function learning processing unit 21 acquires pre-learning sales-related information 32a (Figure 2). Specifically, the correction function learning processing unit 21 acquires pre-learning sales-related information 32a from the auxiliary storage device 15.

[0053] In step S203, the correction function learning processing unit 21 learns a correction function. Specifically, the correction function learning processing unit 21 learns a correction function using the sales volume forecast value, sales price (current generation), sales price (previous generation), sales price (competitor product), and actual sales volume value of the pre-learning sales-related information 32a as supervised learning data.

[0054] At this time, the correction function learning processing unit 21 creates multiple learning correction functions 51 by repeatedly performing the process while gradually changing the learning conditions. The learning conditions are, for example, as follows: <1> Target and degree of dropout (inactivation of a portion of neurons) <2> Target and extent of mini-batches (updating parameters in groups) <3> Selection of activation function type and target neurons <4> Selection of a range of week start dates corresponding to supervised learning data (e.g., new school term, transfer season, immediately after a heatwave, immediately after a major disaster, immediately after a significant change in sales price (competitor products), immediately after a consumer policy such as a tax cut) <5> Selection of the price range for selling supervised learning data (current generation)

[0055] The correction function learning processing unit 21 creates multiple candidates from each of these (1) to (5) and creates n combinations of these candidates. The correction function learning processing unit 21 creates n learning correction functions 51 through n iterations. In this way, overfitting of the correction functions is avoided and differences in the characteristics of each learning correction function 51 are created. In each iteration, the correction function learning processing unit 21 uses an arbitrary optimization algorithm (Adam, gradient descent, etc.). In other words, the correction function learning processing unit 21 performs p iterations to optimize the parameters in each of the n iterations with different learning conditions. p is the number of times the parameter values ​​are swapped.

[0056] In step S204, the correction function learning processing unit 21 creates post-learning sales-related information 32b (Figure 3). Specifically, the correction function learning processing unit 21 uses the corrected sales volume prediction values ​​output in step S203 to create n pieces of post-learning sales-related information 32b and stores them in the auxiliary storage device 15.

[0057] In step S205, the correction function learning processing unit 21 creates post-learning correction function information 42 (Figure 8). Specifically, the correction function learning processing unit 21 uses the optimized parameter values ​​and calculated prediction accuracy from step S203 to create post-learning correction function information 42 having n rows of records, and stores it in the auxiliary storage device 15.

[0058] This completes the learning phase. During the learning phase, the correction function learning processing unit 21 essentially uses known data (number of units, price) of previous generation products from past periods to create multiple learning correction functions 51 that are suitable for use on a case-by-case basis.

[0059] In step S206, the correction function selection processing unit 22 of the forecast demand correction device 1 acquires pre-selection sales-related information 33a (Figure 4). Specifically, the correction function selection processing unit 22 acquires the pre-selection sales-related information 33a from the auxiliary storage device 15.

[0060] In step S207, the correction function selection processing unit 22 selects a learned correction function with high prediction accuracy. Specifically, firstly, the correction function selection processing unit 22 inputs the sales volume prediction value, sales price (current generation), sales price (previous generation), and sales price (competitor product) from the pre-selection sales-related information 33a to each of the n learned correction functions 51 learned in step S203. Secondly, the correction function selection processing unit 22 obtains the sales volume corrected predicted value from each of the n learned correction functions 51 learned in step S203. Thirdly, the correction function selection processing unit 22 calculates the prediction accuracy based on the difference (error) between the corrected sales volume prediction value obtained in step S207 “second” and the actual sales volume value of the sales-related information 33a before selection processing. Fourth, the correction function selection processing unit 22 selects the one with the highest prediction accuracy from among the n learned correction functions 51 learned in step S203 as the selected correction function 52.

[0061] In step S208, the correction function selection processing unit 22 creates post-selection sales-related information 33b (Figure 5). Specifically, the correction function selection processing unit 22 uses the corrected sales volume forecast values ​​obtained in the "second" step of step S207 to create n post-selection sales-related information 33b and stores them in the auxiliary storage device 15. Alternatively, the correction function selection processing unit 22 may create only one post-selection sales-related information 33b using the corrected sales volume forecast values ​​output by the selection correction function 52 selected in the "fourth" step of step S207 and store it in the auxiliary storage device 15.

[0062] In step S209, the correction function selection processing unit 22 creates post-selection correction function information 43 (Figure 9). Specifically, the correction function selection processing unit 22 creates post-selection correction function information 43 by overwriting the post-learning correction function information 42 with the product generation serial number "2", the product generation type "prediction target generation", and the correction selection result in "step 4" of step S207, and stores it in the auxiliary storage device 15.

[0063] This completes the selection phase. In the selection phase, the correction function selection processing unit 22 essentially uses known data (actual sales figures) of the new generation product from past periods to select a learning correction function 51 that can withstand specific cases. Note that "past period" here refers to the most recent past period from the present time.

[0064] In step S210, the correction function utilization processing unit 23 of the forecast demand correction device 1 acquires pre-processing sales-related information 34a (Figure 6). Specifically, the correction function utilization processing unit 23 acquires the pre-processing sales-related information 34a from the auxiliary storage device 15.

[0065] In step S211, the correction function utilization processing unit 23 calculates the corrected predicted sales volume for the target generation in the future period. Specifically, firstly, the correction function utilization processing unit 23 inputs the predicted sales volume, sales price (current generation), sales price (previous generation), and sales price (competitor product) from the pre-utilization sales-related information 34a to the selected correction function 52 selected in "step 4" of step S207. Secondly, the correction function selection processing unit 22 obtains the sales volume-corrected forecast value from the selected correction function 52 selected in step S207, "step 4".

[0066] In step S212, the correction function utilization processing unit 23 creates post-processing sales-related information 34b (Figure 7). Specifically, the correction function utilization processing unit 23 uses the corrected sales volume forecast value output by the selected correction function 52 selected in "step 4" of step S207 to create post-processing sales-related information 34b and stores it in the auxiliary storage device 15.

[0067] In step S213, the correction function utilization processing unit 23 creates post-processing correction function information 44 (Figure 10). Specifically, the correction function utilization processing unit 23 creates post-processing correction function information 44 by overwriting the period classification "future period" and prediction accuracy "-" with the post-processing correction function information 43, and stores it in the auxiliary storage device 15.

[0068] This concludes the utilization phase. In the utilization phase, the correction function utilization processing unit 23 essentially calculates the corrected sales volume forecast for a specific case by inputting known data of the new generation of products in the future period into the selection correction function 52.

[0069] In step S214, the revenue processing unit 24 of the forecast demand correction device 1 forecasts revenue. Specifically, the revenue processing unit 24 calculates revenue by multiplying the forecast value after correcting the number of units sold, obtained in the "second" step of S211, by the sales price (current generation) of the pre-use sales-related information 34a. Here, "revenue" refers to the sales amount excluding costs. The revenue processing unit 24 may further calculate "net revenue" that takes costs into account by multiplying this revenue by a predetermined rate of return.

[0070] In step S215, the display processing unit 25 of the forecast demand correction device 1 displays a display screen. Specifically, the display processing unit 25 displays the learning and selection status of the correction function and the forecast revenue display screen (abbreviated as "display screen") on the output device 13. After that, the processing procedure ends.

[0071] (display screen) Figure 12 shows an example of the display screen 1001. The display screen 1001 includes display areas 1002, 1003, and 1004. The horizontal axis of display area 1002 represents time (week start date 1051), and the time elapsed since release, such as "Time 1", is also indicated next to the week start date, such as "2023 / 04 / 03", on the horizontal axis. The vertical axis of display area 1002 represents the sales price 1040 (upper row) and the number of units sold 1050 (lower row).

[0072] With the date "2024 / 04 / 01" as the boundary, the left side displays a graph showing the time-series trends of the previous generation product "PRD-K1": sales price (current generation) 1011 (solid line), sales price (previous generation) 1012 (single dotted line), sales price (competitor product) 1013 (double dotted line), actual sales volume 1014 (solid line), projected sales volume 1015 (thin dotted line at the starting point of the arrow), and projected sales volume after adjustment 1016 (thick dashed line at the ending point of the arrow).

[0073] To the right of the boundary, a graph is displayed showing the time-series trends of the following for the target generation product "PRD-N2": sales price (current generation) 1021, sales price (previous generation) 1022, sales price (competitor product) 1023, actual sales volume 1024, projected sales volume 1025, and projected sales volume after adjustment 1026.

[0074] "2023 / 04 / 03" 1031 represents the release date of the previous generation product. "2024 / 04 / 01" 1033 represents the release date of the predicted generation product. "2024 / 05 / 27" 1034 represents the present time.

[0075] Based on data from "2023 / 04 / 03" (1031) to "2023 / 11 / 27" (1032), the correction function was trained. This period is the training period (1036). Based on data from "2024 / 04 / 01" (1033) to "2024 / 05 / 27" (1034), a selective correction function 52 was selected from among several trained correction functions (51). This period is the selection period (1037). Based on data from "2024 / 06 / 03" (1035) to "2024 / 11 / 25" (1039), the selective correction function 52 output the corrected predicted sales volume for that period. This period is the future period (1038). The training period, selection period, and future period are displayed on the same time series axis. Note that the “Past Period” in the Period Classification Column 104 of Sales-Related Information 31 corresponds to the Learning Period 1036 and Selection Period 1037 in Figure 12. The “Future Period” in the Period Classification Column 104 of Sales-Related Information 31 corresponds to the Future Period 1038 in Figure 12.

[0076] Comparing the period from 10:31 on "2023 / 04 / 03" to 10:32 on "2023 / 11 / 27" with the period from 10:33 on "2024 / 04 / 01" to 10:39 on "2024 / 11 / 25", although the product generations are different, the product lifecycles and seasons of both products correspond perfectly.

[0077] First, let's focus on the learning period 1036. The display processing unit 25 displays a graph showing that, for the previous generation product (PRD-K1), "a correction function was learned using the sales price (current generation) 1011, the sales price (previous generation) 1012, the sales price (competitor product) 1013, the predicted sales volume 1015, and the actual sales volume 1014 as training data, and the corrected predicted sales volume 1016 was obtained as the output data of the learned correction function."

[0078] In the example shown in Figure 12, the display processing unit 25 displays only the learning result 2 (1060), which is the selected correction function 52, among the multiple learning correction functions 51, in the display area 1003. The display processing unit 25 may also display learning correction functions other than the selected correction function 52 simultaneously or by switching between them. For example, the display processing unit 25 displays the learning period prediction accuracy (90.5%) 1061 in the display area 1004. The display of the graph and prediction accuracy allows the user to visually and quantitatively grasp the correction status during the learning period 1036.

[0079] Next, we focus on the selection period 1037. The display processing unit 25 displays graphs of the corrected sales volume forecast value 1026 and the actual sales volume value 1024, which are calculated by inputting the following into the sales price (current generation) 1021, sales price (previous generation) 1022, sales price (competitor product) 1023, sales volume forecast value 1025, and selection correction function (learning result 2) 1060 for the target generation product (PRD-N2). The graph of the actual sales volume value 1024 is interrupted at the current point 1034. The display processing unit 25 displays, for example, the selection period forecast accuracy (90.3%) 1062 in the display area 1004. By displaying the graphs and forecast accuracy, the user can visually and quantitatively grasp the correction status in the selection period 1037.

[0080] Furthermore, we focus on the future period 1038. The display processing unit 25 displays a graph of the corrected sales volume forecast value 1026, which is calculated by inputting the following into the sales price (current generation) 1021, sales price (previous generation) 1022, sales price (competitor product) 1023, sales volume forecast value 1025, and selection correction function (learning result 2) 1060 for the target generation product (PRD-N2). For example, the display processing unit 25 displays the forecast revenue "¥1234567890" 1063 in the display area 1004, which is obtained by multiplying the sales price (current generation) 1021 of the target generation product (PRD-N2) 1020 by the corrected sales volume forecast value 1026 of the target generation product (PRD-N2) 1020. The forecast revenue may be displayed for each sales point in time, or it may be displayed for the entire future period. The display of graphs and prediction accuracy allows users to visually and quantitatively understand the correction status in the future period 1038 where the selection correction function is used.

[0081] (Effects of this embodiment) As described above, the forecast demand correction device 1 of this embodiment can perform highly accurate demand forecasting by learning a correction function that takes into account the complex correlation between the number of units sold and the selling price, which changes depending on the product life cycle, season, etc., and by using it to calculate a forecast value after correcting the number of units sold. Furthermore, the revenue processing unit 24 predicts revenue by multiplying the sales volume-adjusted forecast value by the sales price. Users can refer to this result to quantitatively evaluate the sales strategy in terms of revenue. Furthermore, the display processing unit 25 displays the status of learning and selection of the correction function in a graph based on sales-related information 31 and correction function information 41. Users can refer to this graph to visually confirm how the sales figures have been corrected.

[0082] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are included. For example, the embodiments described above are described in detail for the purpose of explaining the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace a part of the configuration of each embodiment with other configurations. In addition, some or all of the above configurations, functions, processing units, processing means, etc., may be realized in hardware, for example, by designing them as integrated circuits. Furthermore, each of the above configurations, functions, etc., may be realized in software by having a processor interpret and execute a program that realizes each function. Information such as programs, tables, files, etc. that realize each function can be stored in memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD. [Explanation of Symbols]

[0083] 1. Forecast Demand Correction Device 11 Central Control Unit 12 Input devices 13 Output device 14 Main memory 15 Auxiliary storage 16. Data Input / Output Device 21 Correction function learning processing unit 22 Correction function selection processing unit 23 Correction function utilization processing unit 24 Revenue Processing Section 25 Display Processing Unit 31 Sales-related information 32a Pre-training sales-related information 32b Sales-related information after learning processing 33a Sales-related information before selection processing 33b Sales-related information after selection processing 34a Pre-sale information 34b Sales-related information after processing 41 Correction Function Information 42 Post-learning correction function information 43 Post-Selection Correction Function Information 44. Correction function information after processing 51 Learning Correction Function 52 Selection Correction Functions 1001 Screen displaying the learning and selection status of the correction function and the predicted earnings.

Claims

1. A correction function learning processing unit creates multiple learning correction functions by learning a correction function that takes predicted sales volume and sales price as input and outputs corrected predicted sales volume, using past sales volume predictions, sales price, and actual sales volume values ​​which are training data for corrected predicted sales volume for past generations of products over past periods. For each of the aforementioned learning correction functions, the predicted sales volume and sales price for the target generation product over the past period are input. The prediction accuracy is calculated based on the difference between the corrected sales volume prediction value output by each of the aforementioned multiple learning correction functions and the actual sales volume value of the target generation product in the past period. A correction function selection processing unit selects a correction function from among the multiple learning correction functions based on the calculated prediction accuracy, For the selected selection correction function, input the projected sales volume and sales price of the target generation product over the future period. A correction function utilization processing unit that obtains the sales volume correction forecast value output by the selected selection correction function, A forecast demand correction device characterized by comprising the following features.

2. The aforementioned selling price is Includes the selling price of the previous generation product. A forecast demand correction device according to claim 1, characterized by the following:

3. The aforementioned selling price is Includes the selling price of competing products. A forecast demand correction device according to claim 1, characterized by the following:

4. The system includes a revenue processing unit that calculates the revenue of the target generation product by multiplying the sales volume-corrected forecast value obtained by the correction function utilization processing unit by the sales price of the target generation product in the future period. A forecast demand correction device according to claim 1, characterized by the following:

5. A graph showing the sales price, projected sales volume, actual sales volume, and adjusted projected sales volume for the aforementioned past generation products over a past period, A graph showing the sales price, projected sales volume, actual sales volume, and corrected projected sales volume for the aforementioned target generation product over a past period, A graph showing the sales price, sales volume forecast, and sales volume adjusted forecast for the aforementioned target generation product over a future period, It includes a display processing unit that displays them in the same time series. A forecast demand correction device according to claim 1, characterized by the following:

6. The correction function learning processing unit of the forecast demand correction device is: Multiple learning correction functions are created by training a correction function, which takes predicted sales volume and sales price as input and outputs a corrected sales volume forecast, using the actual sales volume data, which serves as training data for past sales volume forecasts, sales prices, and corrected sales volume forecasts for past generations of products over past periods. The correction function selection processing unit of the forecast demand correction device is: For each of the aforementioned learning correction functions, the predicted sales volume and sales price for the target generation product over the past period are input. The prediction accuracy is calculated based on the difference between the corrected sales volume prediction value output by each of the aforementioned multiple learning correction functions and the actual sales volume value of the target generation product in the past period. Based on the calculated prediction accuracy, select a correction function from among the multiple learning correction functions. The correction function utilization processing unit of the forecast demand correction device is: For the selected selection correction function, input the projected sales volume and sales price of the target generation product over the future period. Obtain the sales volume correction forecast value output by the selected selection correction function. A forecast demand correction method characterized by the following.

Images