JPH08221384A - Method and device for predicting sales amount - Google Patents

Abstract

PURPOSE: To accurately predict sales amount over the long period of time based on a few data. CONSTITUTION: This device for predicting the sales amount over the period B from this month with a sales result quantity in the past over the period A from a prediction object month as input data is provided with an input means 101 for inputting the sales result quantity, a data storage means 102 for storing the sales result quantity, an arithmetic means 103 for optimizing a coupling coefficient for forming a neural network by the learning of the sales result of a previous year and predicting the sales amount for the period B at the time of prediction based on it, a neuro data table 104 for storing the coupling coefficient optimized by the arithmetic means 103 and an output means 105 for outputting a predicted value by the arithmetic means 103. When the input data to the neural network are normalized by a normalization coefficient calculated from the data of both years, prediction accuracy is improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to prediction of future sales volume using a neural network,
In particular, the present invention relates to a method and a device for accurately predicting a long-term sales forecast based on sales performance data of the previous year.

[0002]

2. Description of the Related Art In recent years, a variety of products have been sold in the manufacturing industry, and new products have come into the market one after another. Therefore, it is becoming difficult for both the producer side and the seller side to predict the sales volume of all products and make a detailed production or sales plan. However, since it is under such circumstances, a detailed sales plan is indispensable. Therefore, conventionally, the sales volume is generally forecasted by each person in charge only for the product of interest, and other forecasts are made. For products, the actual sales figures for the previous fiscal year are directly used as the sales forecast quantities. However, in the forecast by the person in charge, there are considerable individual differences due to the person's sales forecasting ability, information gathering ability, time spent for forecasting, etc. Is unintentionally taken in,
Accurate prediction is impossible.

Therefore, a device having a prediction ability equivalent to or higher than that of a skilled engineer has been devised by performing information processing by using a neural network which imitates human nerve cells. For example, learning the trend of past demand volume using a neural network,
Japanese Patent Laid-Open No. 4-372046 “Demand Demand Forecasting Method and Device” and Japanese Patent Laid-Open No. 5-18995 “Power Demand Forecasting Device” are used to predict future demand based on the actual tendency.

Japanese Unexamined Patent Publication No. 4-372046 “Demand amount forecasting method and apparatus” describes factors affecting the demand amount.
Divide into “continuous influence factors” and “discontinuous influence factors”, and for each “discontinuity influence factor (or combination)”,
Calculate the correlation function between "continuity influencing factors" and past demand,
The demand amount at the time of the prediction target is predicted according to the function, and a multiple regression model or a neural network model is adopted when obtaining the correlation.

[0005] In Japanese Patent Laid-Open No. 5-18995, "Power Demand Forecasting Device", a statistical model representing the relationship between weather data and power demand is created based on past weather data and power demand data. Using this statistical model, the total power demand Q ₁ on the prediction day is predicted. On the other hand, the difference between the total power demand Q ₁ and the past power demand Q ₂ up to the forecast day
(Q ₃ = Q ₁ −Q ₂ ) is calculated, and the relationship between the difference amount Q ₃ and the error between the past weather data up to the forecast day, the power demand, and the calendar day information is learned by a neural network. , Based on the learning data, the total power demand Q _{1 on the} forecast day
Estimate the correct value Q ₁ 'corrected by

Furthermore, not only the past demand amount (sales amount) but also the events that may occur in the future are input to the neural network, the input data are learned, and the sales amount is predicted based on the learning data. For example, in Japanese Unexamined Patent Publication No. 6-12399 “Prediction System”, a scenario is created for each event that can occur after the prediction time point, and a prediction value is calculated by a neural network for each scenario. ing.

[0007]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention
Both 2046 and Japanese Patent Laid-Open No. 5-18995 focus on making short-term forecasts on a daily basis, so they can handle small daily fluctuations, but they are not suitable for long-term forecasts necessary for long-term sales strategies. Not suitable. In the first place, the subject of these publications is that the quantity of water and electric power is fixed to some extent, and the demand patterns are similar in each year, and it is intended to predict it.

On the other hand, the products to be predicted, which the present application mainly intends to deal with, are released as new products one after another within a short period of time, and therefore have a relatively short sales period (life). For such products, because there is no accumulated sales data for several years, the forecasts such as the above publication cannot provide accurate forecasts due to lack of data, and there is a certain amount of forecasted demand for water and electricity. However, since the amount of fluctuation is large (the sales quantity may increase several times in several months), the input value to the neural network exceeds the allowable upper limit input in the prediction method described in the above publication, and the learned neural The network cannot make accurate predictions.

[0009] In the "prediction system" of Japanese Patent Laid-Open No. 6-12399, a scenario must be created for each product whose sales are to be predicted. Therefore, it is impossible to set a scenario because there are too many fluctuation factors and there are too many items.

Therefore, according to the present invention, it is possible to make a prediction based on the sales data of only the previous year, to make a prediction for a long period of time, to make a prediction even when the sales number fluctuates greatly, and to deal with a large number of items
The purpose is to:

[0011]

According to the sales volume forecasting method of the present invention, the past sales forecast volume over a period A before the sales forecast month is used as input data a, and the sales forecast is performed by a neural operation using a neural network. A method of predicting the sales volume over a period B from the target month, wherein the actual sales amount in the period A ′ of the previous year corresponding to the period A is used as the input data a ′, and the period of the previous year corresponding to the period B is set. The sales record quantity in B'is used as teacher data b ',
Learning is performed using a neural network, the coupling coefficient forming the neural network is optimized, and then the neural network to which the optimized coupling coefficient is added is used to calculate the sales amount for the period B for the input data a. It is characterized in that the predicted quantity is obtained by a neuro operation.

In the sales volume forecasting apparatus of the present invention, the past sales record quantity over the period A before the sales forecast month is used as the input data a, and the period B after the sale forecast month is calculated by the neural operation using the neural network. An apparatus for predicting the sales volume over, and an input means (101) for inputting the sales actual quantity before the target month of the sales prediction,
A data storage means (102) for storing the sales record quantity input from the input means (101) and a neuro data table to which the coupling coefficient is added at the time of prediction by optimizing the coupling coefficient forming the neural network at the time of learning. Using period B
Calculating means (103) for predicting the sales amount to
A neuro data table (104) in which the coupling coefficient optimized by (103) is stored, and an output means for displaying or printing out the result predicted by the computing means (103).
(105) and the arithmetic means (103), in detail, during learning: The sales record quantity in the period A ′ of the previous year corresponding to the period A is used as the input data a ′ and corresponds to the period B. The sales actual quantity in the period B ′ of the previous year is read as the teacher data b ′ from the data storage means (102), learning is performed using the neural network, the coupling coefficient is optimized, and at the time of prediction: the data storage means (102) ), A neural network in which a coupling coefficient read out from the neuro data table (104) is applied to the input data a read out from (1) is used to obtain a forecast quantity of sales volume for the period B by a neuro calculation.

[0013]

According to the conventional prediction method using the neural network, as shown in the conventional example, the prediction method is based on the accumulated data over the past several years. According to the prediction device according to claim 4 based on the above, it is possible to make a prediction based on the sales actual quantity of only the previous year,
And it can be predicted over the long term.

As described in claim 2 (and claim 5), the correction means (103a) normalizes the input data a ′ and the teacher data b ′ at the time of learning by the learning normalization coefficient k ₁ , and further, Since the input data a at the time of the prediction is normalized by the prediction normalization coefficient k ₂ , it is possible to deal with a case where the sales number fluctuates greatly and to make an accurate prediction.

When learning the sales record quantity of the previous year, the coupling coefficient is sequentially optimized by using a random number generator, but the connection coefficient is calculated based on the sales record quantity of the previous year. For example, as described in claim 3 (and claim 6), it is possible to optimize in a short time by modifying the coupling coefficient.

[0016]

FIG. 2 is a block diagram showing an embodiment of a prediction device based on the prediction method of the present invention. Reference numeral 1 is an input device for inputting the past daily sales record quantity for each product, and the data input from the input device 1 is registered in the first database 2. Reference numeral 3 denotes a central processing unit, which learns the past sales record quantity of the sales prediction target product by using a neural network, and optimizes the coupling coefficient of each neuron constituting the neural network model.
A second database 4 stores the coupling coefficient optimized by learning. Further, when predicting the sales amount, the central processing unit 3 reads out a necessary coupling coefficient from the second database and performs a neuro calculation using a neural network to which the coupling coefficient is given to predict the sales number in the sales forecast month.

Further, the central processing unit 3 calculates the daily sales record quantity registered in the first database 2 by 1 for each.
For the month, the weekly division is performed every 7th day from the end of the 4th week, the 3rd week, the 2nd week, and the remaining days are included in the 1st week, and the totaling means for collecting the data on a weekly basis will be described later. As described above, the learning normalization coefficient k ₁ and the prediction normalization coefficient k ₂ are obtained, and the correction means normalizes the input data and the teacher data.

Reference numeral 5 is a CRT display device for displaying a graph of the forecast sales volume calculated by the neuro processing by the central processing unit 3, and 6 is a table display of the forecast sales volume,
It is also a thermal transfer printer that prints out the display contents of the CRT display device 5. An alarm device 7 outputs an alarm when the predicted value is abnormal.

The neural network is an artificial simulation that imitates the learning function of brain cells, and its model is shown in FIG. 3 to be referred to later. The model of FIG. 3 is composed of three layers: an input layer into which the sales record quantity is input, one intermediate layer, and an output layer from which a neural operation result by this neural network is output.
Each layer is composed of a large number of neurons (marks). For example, one neuron N in the intermediate layer (the function of which is common to all neurons regardless of layers) is the input of the previous layer as shown in FIG. The output (x ₁ , x ₂ , ...) Of each neuron in the layer is input with its weight (w ₁ , w ₂ , ...). The neuron mutually adds the inputs, and outputs a single response signal in the range of 0 to 1 for the added value by using a sigmoid function having a quasi-linear saturation type response characteristic.
The output is further weighted (w ₁ ′, w ₂ ′, ...) And input to each neuron in the lower output layer. These weights are coupling coefficients that indicate the strength of coupling between layers, and the value of this coupling coefficient is less than or equal to the specified value when the difference between the neural network output value from the neural network and the teacher signal is the input value. Is optimized so that This function is the learning, and in particular, the learning performed by using the teacher signal in the hierarchical neural network as shown in FIG. 3 is called pack propagation.

FIG. 5 shows the prediction in this device having the above-mentioned configuration.
It will be described according to the flowchart of. In the first data table 2, the sales forecast target product (in this embodiment, our laundry detergent "A large box"), from the sales forecast target month (first month for long-term forecast, that is, the current month), The daily sales record quantity for at least the past 15 months is stored. First,
In step S1, the sales forecast target product and the current month as the forecast target month are input from the input device 1. However, from this month 6
Even if you forecast the sales volume up to several months ahead, the forecast result cannot be verified, so here, as shown in the graph of FIG.
For example, April 1994 is selected as the target month for prediction using the past sales actual quantity (marked with ●) in 1993 and 1994, and it is decided to predict the number of sales from April to December 9 months ahead. . Therefore, in this case, the first database 2 stores the past sales quantity for 15 months from January 1993 to March 1994 as past reference data.

When the sales forecast target month is input, the daily sales actual quantity from January 1, 1993 to March 31, 1994 is read from the first database 2 in step S2, and the central sales amount is read. The totalizing means of the arithmetic processing unit 3 divides one month into four weeks (weekly division), and totals the sales actual quantity on a weekly basis. It should be noted that the amount of data that will be enormous when trying to present the daily sales record quantity is shown in the table below.

[0022]

[Table 1]

In step S3, the correction means normalizes the input data (past sales record quantity) and the teacher signal, which are used in the following learning. That is, in FIG. 6, it is not possible to adopt the same weight because there is a time lag between the sales actual quantity of 1993 and the actual sales quantity of 1994. Therefore, in this embodiment, the data is normalized by using a normalization coefficient.
Specifically, the past 3 before the sales forecast target month of April 1994
Monthly (a total of the weekly sales actual quantity in the period A from January to March is a, and a total of the weekly sales actual quantity in the period A ′ in the 93 fiscal year corresponding to the period A is a ′. , The weekly sales actual quantity of the period B ′ in 1993 corresponding to the sales forecast period B for 9 months from April 1994 is “teacher signal”, where k ₁ = a ′ / ( a
+ A ') is used as a normalization coefficient (referred to as a learning normalization coefficient in order to distinguish it from a normalization coefficient used in a neuro calculation described later), and a period B input to the neural network
Normalization was performed by dividing the sales record quantity and the teacher signal by the learning normalization coefficient k ₁ .

In step S4, data for earlier (that is, 1992) data is learned for the sales forecast month, and the resultant coupling coefficient is the second database 4
If it is the first learning, the coupling coefficient is randomly generated in step S6, and the normalized input value is calculated in step S8.
Learning is repeated until the difference between the value input to the input layer of the neural network model shown in (1) and output from the output layer at that time on a weekly basis and the teacher signal generated by the random number is equal to or less than a predetermined value. On the other hand, when the learning is performed in the past and the coupling coefficient is accumulated, the coupling coefficient is read in step S5, the coupling coefficient is corrected in step S7, and the learning is similarly performed in step S8. Be seen. In this case, the time until the coupling coefficient is optimized (converged) is significantly shortened compared with the time when the learning is the first time.

In the coupling coefficient optimized by such learning, information indicating the sales tendency in the period B'for the period A'in 1993 is held. This coupling coefficient is registered (updated if registered) in the second database 4 in step S9. As described above, the “learning” for the prediction is completed, and the “prediction” of the sales forecast target product is performed based on the learning result.

In step S10, first, the coupling coefficient of the sales forecast target product is read from the second database 4.
Then, in step S11, as the input data used at the time of forecasting, the sales actual quantity in week units of the past three months (period A from January to March 1994) before the sales forecasting month (April 1994) is read out. Then, the value is normalized. Normalization coefficient here (predicted normalization coefficient k ₂ )
Is given by k ₂ = a / (a + a ′).

FIG. 7 shows a neural network model in which a coupling coefficient obtained by learning is added to the neural network model of FIG. When the sales record quantity of 12 weeks in the period A normalized by the predictive normalization coefficient k ₂ is input to the input layer of the neural network model, the neural network model produces
Nieuro calculation is performed, and a value in the range of 0 to 1, which is the response value of the sigmoid function described above, is output from the output layer every predicted week, that is, 36 values are output. By multiplying each of these output values by the prediction normalization coefficient k ₂ , the sales forecast quantity in 36 weeks in the forecast period B from the sales target month to 9 months ahead can be obtained. These sales forecast quantities are in step S
It is stored in the first database 2 at 13. These predicted values are shown in Table 2.

[0028]

[Table 2]

These predicted sales quantities can be displayed as a graph on the CRT display device 5 as shown in FIG. 6 or can be printed out by the printer 6. When the sales forecast quantity obtained in this way deviates greatly from the numerical value expected in advance, the corresponding display location is displayed in a different color or blinks on the CRT display device 5. Since an alarm is output from the alarm device 7, it will not be overlooked. In FIG. 6, the sales forecast quantity for each week predicted for the forecast period B is also shown by a mark, and it was possible to almost match the actual sales actual quantity (mark ●).

In this embodiment, the unit of input data is 1
It is set as a week, and one month is divided into four weeks. The reason for this is described below. If the unit of input data is one day, the sales quantity will inevitably be predicted in units of one day. However, as the number of input data increases, the number of coupling coefficients will become huge, making learning impossible. Become. Further, in FIG. 6, the sales quantity in the period B ′ in 1993 is used as a teacher signal, and the sales quantity in the period B in 1994 corresponding to this period B ′ is predicted. The days of 1993 and 1994 do not correspond because the days of the week are different, while the daily sales volume changes greatly depending on the day of the week. Therefore, the daily forecast includes a change in the sales amount depending on the day of the week as an error, and an accurate forecast cannot be performed. Therefore, it seems appropriate that the unit of input data is one week or ten days.

Therefore, if the unit of input data is set to 10 days and it is attempted to predict the sales quantity every 10 days, it is convenient because one month can be equally divided into three, but it is unsuitable for the following reasons. Reason 1: Depending on the day of the week, there are days that can be sold (for example, Sunday) and days when it cannot be sold, and if there is a 10-day discount, then every 10 days will include one day that can be sold (for example, Sunday, the same can be said for unsold days). Sometimes included twice, so each 10
The sales volume varies from day to day. Reason 2: For example, during 1st to 10th August 1993,
Even though only one Sunday is included, two Sundays may be included between the 1st and 10th days of August 1994, which corresponds to this period. The included days of the week do not correspond. Therefore, it is inappropriate to set the unit of input data to 10 days.

On the other hand, if the unit of input data is one week, each week includes evenly each day from Sunday to Saturday, so that there is no variation in the input data, and the reason 1 is solved. In addition, the day of the week in 1993 differs from the corresponding week in 1994, but each day is 1
Since every week includes every day from Sunday to Saturday, there are no variations in each week in both years, and because the two weeks will be fully addressed, Reason 2 will be resolved.

However, if you divide every 7 days, at the end of the month,
One week may extend to the next month, and in this case, therefore, the sales forecast month will be predicted, ..., 20th week, 21st week, and so on. However,
After all, since it is more convenient to make predictions on a weekly basis with a month as a delimiter, one month is divided into four weeks in this embodiment. In that case, if weekly allocation is performed every 1st to 7th day of the beginning of the month, the fourth week becomes 10th day in the month having 31st day.
In general, the sales quantity tends to increase at the end of the month, so that the sales quantity may be about double that of other weeks during the 10th day of the fourth week. In this way, since the sales actual quantity greatly varies in each week and smooth learning cannot be performed, in the present embodiment, every 7 days from the end of the month, the fourth week, the third week, and the second week are divided into weeks. Therefore, the number of remaining days was included in the first week, and the actual sales volume for each week was made uniform.

Further, in this embodiment, the input data is "three months" before the sales forecast target month, but this will also be examined. As described above, it is best to input the input data on a weekly basis. These three months are closely related to weekly input.
If it is longer than (12 pieces of input data), the input data increases and learning for optimizing the coupling coefficient becomes difficult.
It becomes difficult to capture the recent sales trends. On the other hand, if the input data period is shorter than this, learning will be easier, but
Due to lack of data, accurate prediction cannot be performed for a long time. As a result of repeating the prediction by adjusting the number of input data, the applicant was able to predict relatively well when the period of the input data was 3 months. However, the number of input data is not limited to 12 (weeks), and may be around 12, but on a monthly basis (the number of input data is 8, 12,
Since it is more convenient to divide it by 16 pieces, ... On the other hand, it is not necessary to limit the period for which the input data for these three months can be predicted to 9 months, and it was possible to accurately predict the period from a short period of a few months to a maximum of 12 months.
For example, when predicting 12 months (48 weeks), it is necessary to set the number of teacher data to 48 and optimize the coupling coefficient again.

The upper graph of FIG. 8 shows the sales forecast quantity (marked in FIG. 6) in the forecast period B of FIG. 6 (the item to be forecast is the laundry detergent “A large box”, and the period A of the input data was 3 months). ) Is summarized and re-posted by month. On the other hand, the lower graph of FIG. 8 shows the sales forecast quantity when the period of input data is 6 months for the same forecast target product. The prediction accuracy in this case did not differ much between the two graphs.

FIG. 9 shows a case in which the item to be predicted is the dish detergent "M refill", and the upper graph is the period A of the input data.
Is 3 months, and the graph below shows the input data period is 6 months. In this case, the input data period is 3 months.
Higher prediction accuracy was obtained in months.

[0037]

As described above, according to the present invention, when forecasting the sales volume over several months from the sales forecast target month, the sales actual quantity before and after the sales forecast target month in the previous year is input data, As output data, learning is performed using a neural network, and based on the learning results, the sales volume for several months after the sales forecast target month in this year is predicted, so it is possible to predict relatively long term with a small amount of data. It can be performed. At that time, if the data used in learning and the data used in prediction are normalized by using a normalization coefficient, accurate prediction can be performed even when the number of sales varies greatly. In addition, since a small amount of data can be used for the prediction, that is, a large amount of labor is not required for the prediction, it is possible to support the prediction of many items.

[Brief description of drawings]

FIG. 1 is a diagram illustrating a complaint correspondence diagram of a sales volume prediction device according to the present invention.

FIG. 2 is a control block diagram showing an embodiment of a sales volume prediction device of the present invention.

FIG. 3 is a diagram showing a model of a neural network in which a coupling coefficient is optimized by learning.

FIG. 4 is a diagram showing the functions of neurons forming a neural network.

FIG. 5 is a flowchart showing a prediction operation performed by this device.

FIG. 6 is a diagram showing input data and teacher data adopted during learning, as well as input data used during prediction and a prediction result in a table format.

FIG. 7 is a diagram showing a model of a neural network that predicts sales volume by using optimized coupling coefficient.

FIG. 8 is a graph showing sales forecast results when the input data is 3 months and 6 months.

FIG. 9 is a graph showing sales forecast results when input data is set to 3 months and 6 months for another product.

[Explanation of symbols]

 1 Input Device 2 First Database 3 Central Processing Unit 4 Second Database 5 CRT Display Device 6 Printer 7 Alarm Device

Claims (15)

[Claims] 1. A method of predicting a sales volume for a period B after the sales forecast target month by a neuro calculation using a neural network, using past sales actual quantity over a period A before the sales forecast target month as input data a. In addition, the sales record quantity in the period A ′ of the previous year corresponding to the period A is input data a ′, and the sales record quantity in the period B ′ of the previous year corresponding to the period B is teacher data b ′. Learning is performed using a neural network, the coupling coefficient forming the neural network is optimized, and thereafter, the neural network to which the optimized coupling coefficient is given is used to calculate the sales amount for the period B for the input data a. A sales volume forecasting method, characterized in that the forecast volume is obtained by a neuro operation. 2. The learning normalization coefficient k ₁ = a ′ / (a + a ′) is used in advance for the input data a ′ in the period A ′ and the teacher data b ′ in the period B ′, which are adopted in the above learning. The input data a of the period A, which is used for the above prediction, is normalized with a prediction normalization coefficient of k ₂ = a / (a + a ′), and the neural network using the neural network is used. _{3. An} accurate sales forecast quantity is obtained by multiplying the forecast sales quantity obtained by calculation by the forecast normalization coefficient k 2.
Sales volume forecast method described in. 3. When the coupling coefficient when the sales quantity of the period B'is predicted by the periods A'and B'of the previous year and the periods A "and B" of the years before the two corresponding to these periods is calculated. The sales quantity forecasting method according to claim 1 or 2, wherein in the learning, optimization is performed by modifying the coupling coefficient. 4. A device for predicting a sales volume for a period B after the sales forecast target month by a neuro calculation using a neural network, using past sales actual quantity over a period A before the sales forecast target month as input data a. And, the input means (101) for inputting the sales actual quantity before the target month of the sales forecast, the data storage means (102) for storing the sales actual quantity input by the input means (101), At the time of learning, the calculation means (103) for optimizing the coupling coefficient forming the neural network, and the prediction means for predicting the sales volume for the period B using the neuro data table to which the coupling coefficient is given, and the calculation means (103). A neuro data table (104) in which the optimized coupling coefficient is stored, and an output means (105) for displaying or printing out the result predicted by the calculating means (103). In detail, the calculation means (103) is as follows: During learning: The actual sales quantity in the period A ′ of the previous year corresponding to the period A is used as the input data a ′, and the period of the previous year corresponding to the period B is set. The sales actual quantity in B ′ is read as the teacher data b ′ from the data storage means (102), learning is performed using a neural network, the coupling coefficient is optimized, and at the time of prediction: read out from the data storage means (102). A sales volume forecasting device, characterized in that a forecast quantity of sales volume for a period B is obtained by a neuro calculation by using a neural network to which a coupling coefficient read from the neuro data table (104) is applied to the input data a. 5. The learning normalization coefficient k ₁ = a ′ / (a + a ′) is used in advance for the input data a ′ of the period A ′ and the teacher data b ′ of the period B ′, which are adopted in the above learning. Period A, which has been normalized and is used for the above forecast
Input data a of the above is normalized with a predictive normalization coefficient k ₂ = a / (a + a ′), and the predictive normalization coefficient k ₂ is added to the sales forecast quantity obtained by the neuro operation using the neural network. The sales volume forecasting apparatus according to claim 4, further comprising a correction means (103a) for multiplying the value to obtain an accurate sales forecast volume. 6. The combination coefficient when the sales quantity of the period B'is predicted by the periods A'and B'of the previous year and the periods A "and B" of the years before the two corresponding to these periods. In the learning, the sales quantity prediction device according to claim 4 or 5, wherein optimization is performed by modifying the coupling coefficient. 7. An aggregate means for using data aggregated on a weekly basis as the input data and the teacher data.
7. The sales quantity prediction device according to claim 4, further comprising (103b). 8. The aggregating means (103b), when obtaining the data aggregated on a weekly basis, reads each month from the end of the fourth month.
7. The sales quantity forecasting apparatus according to claim 4, wherein the sales quantity forecasting apparatus is divided into weekly, third weekly, second weekly and every seven days, and the remaining days are included in the first week. 9. The sales quantity forecasting device according to claim 4, wherein the period A is 3 months. 10. The sales quantity forecasting apparatus according to claim 9, wherein the sales forecasting period B is between 1 and 12 months. 11. The sales quantity forecasting apparatus according to claim 10, wherein the period B for forecasting sales is set to 9 months. 12. The sales quantity predicting device according to claim 4, wherein the output means (105) displays the predicted sales quantity as a trend in a graph in comparison with the sales quantity of the previous year. 13. The sales quantity predicting device according to claim 12, wherein when the predicted sales quantity is far from the expected value, the display on the output means (105) is displayed in a different color or blinks. 14. When the predicted sales quantity is far from the expected numerical value, the display on the output means (105) is displayed in a different color or blinks and the output means (105) is displayed.
13. The sales quantity prediction device according to claim 12, which outputs an alarm from the device. 15. A device for predicting a sales volume for a period B after the sales forecast target month by a neuro operation using a neural network, using past sales actual quantity over a period A before the sales forecast target month as input data a. And, the input means (101) for inputting the sales actual quantity before the target month of the sales forecast, the data storage means (102) for storing the sales actual quantity input by the input means (101), In order to use the input sales volume as input data and teacher data that are aggregated on a weekly basis, each month is divided into 4th week, 3rd week, 2nd week, and 7th week from the end of the month, and the remaining The aggregation means (103b) for including the number of days in the first week, and the input data a ′ of the period A ′ and the teacher data b ′ of the period B ′, which are adopted at the time of learning, are previously set as k ₁ = a ′ / With a learning normalization coefficient of (a + a ') -Normalized; then, then subjected to during the prediction, the input data a time period A, k ₂ = a /
(a + a ') is normalized with a prediction normalization coefficient, and the sales prediction quantity obtained by the neuro operation using the neural network is multiplied by the prediction normalization coefficient k ₂ to obtain an accurate sales prediction quantity. Correction means (103a)
An arithmetic means (103) for optimizing the coupling coefficient forming the neural network at the time of learning and for forecasting the sales amount for the period B using the neuro data table provided with the coupling coefficient at the time of prediction; ), A neuro data table (104) storing the coupling coefficient optimized, and an output means (105) for displaying or printing out the result predicted by the arithmetic means (103), the arithmetic means (103) For details, during learning: Input the sales data quantity in the period A ′ of the previous year corresponding to the period A as the input data a ′, and the sales data volume in the period B ′ of the previous year corresponding to the period B in the teacher data. As b ′, each is read from the data storage means (102), learning is performed using a neural network, the coupling coefficient is optimized, and at the time of prediction: read from the data storage means (102) For the input data a, the neural network to which the coupling coefficient read from the neuro data table (104) is added is used to obtain the forecasted sales volume of the sold products for the period B by the neuro calculation, and the forecasted sales volume is characterized. apparatus.