JP2000250888A - Model selection type demand predicting system by predictive purposes - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy in prediction by analyzing the demand characteristics of an article and making a process for extracting a demand model in response to the demand characteristics into a routine. SOLUTION: 'Trend/pattern presence/absence judgement' is processing for judging the presence/absence of trend and pattern by performing trend analysis and pattern analysis on the basis of demand result information. Relating to 'predictive model selection', the candidates of a demand predictive model are selected from the judged result of trend/pattern presence/absence. In order to evaluate the candidate models, predictive values (predictive values for evaluation) in the case of applying respective candidate models in the past are calculated and an evaluation value is calculated in comparison with the real demand result. This is processing for selecting the optimum predictive model on the basis of the result. 'Further predictive value calculation' is processing for calculating a predictive value in future while using the selected predictive model.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for selecting a demand forecast model and calculating a demand forecast value by utilizing actual demand information.

[0002]

2. Description of the Related Art As one of the important themes of unification of production, sales and distribution, improvement of demand forecasting accuracy is often taken up.
The construction of an optimal demand forecasting model is a prerequisite for increasing the accuracy of demand forecasting values. Until now, this forecasting model was created in a form that relied on human know-how. As a well-known example of a prediction model created based on human know-how, there is a “demand amount prediction method” (Japanese Patent Laid-Open No.
-311851).

In this known example, a prediction value of a demand amount composition ratio by day of the month including a prediction target date is calculated by calculating an average value of a demand amount composition ratio by day of the latest same day for each day.
The forecast target date is calculated by taking the weekly demand volume composition ratio of the same month as the month including the forecast target date as the reference weekly demand volume composition ratio and taking into account the latest amount of change in monthly fluctuation characteristics to the reference weekly demand volume composition ratio. Calculate the forecast value of the weekly demand component ratio of the month including the month, calculate the forecast value of the demand component ratio by day and the forecast value of the weekly demand component ratio, and calculate the forecast value of the daily demand component ratio. And a predicted value of the demand amount on the prediction target day is calculated from the predicted value of the monthly demand amount. "

[0004]

Relying on human know-how to create a prediction model means that the know-how is not formulated, and as a result, it has been difficult to maintain prediction accuracy. In addition, under the current situation where the fluctuation cycle of the demand characteristic is short and the variety of product items is increasing, it is difficult in terms of man-hours / time to create a prediction model for all items in a timely manner by human know-how. There was a problem.

[0005]

SUMMARY OF THE INVENTION The present invention is a method and apparatus for analyzing a demand characteristic of a product and formulating a process of extracting a demand model according to the demand characteristic.

[0006]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a system configuration diagram showing a computer, a storage device, and the like for performing the present invention.
10 to 10 show flowcharts for implementing the present invention. FIGS. 11 to 26 show table diagrams of files stored in the respective storage devices constituting FIG.

In FIG. 1, reference numerals 1001 to 1011 denote programs for realizing the method of the present invention.
A control program 1001 and a work area 1003 perform basic processing for executing the program.
Reference numeral 1005 denotes a prediction model module.
Various patterns of prediction programs for calculating predicted values are shown. Reference numeral 1007 denotes a trend / pattern property determination module, 1009 denotes a prediction model selection module, and 1011 denotes a future predicted value calculation module, which is a processing program for implementing the present invention.

Reference numerals 1013 to 1037 denote files for implementing the present invention, 1013 denotes a predicted product master, 1015 denotes a trend / pattern analysis master,
Reference numeral 17 denotes an annual trend corresponding demand actual file, 1019 denotes a monthly trend corresponding demand actual file, 1021 denotes a yearly monthly patternable demand actual result file, 1023 denotes a monthly weekly pattern corresponding demand actual file, and 1025 denotes a correlation coefficient. Test table master, 1027 is F distribution table master, 1029
Is a master corresponding to the trend / pattern characteristic determination result / applied model, 1031 is a master corresponding to the prediction model, 1033 is a theoretical value file for evaluation, 1035 is an evaluation value file, 103
Reference numeral 7 denotes a future predicted value file.

Reference numeral 1039 denotes a display, 1041 denotes an input device, and 1043 denotes a CPU.

FIGS. 11 to 28 show table diagrams of the files shown in FIG. FIG. 11 shows a table of the predicted product master (1013), in which a unit of fineness (prediction unit) and a period for which prediction is required (prediction period) are described for each product.

FIG. 12 shows a table of the trend / pattern / exponential smoothing analysis master (1015). The number of data stored in the performance file and one cycle as shown in FIGS. The number (for example, in the case of monthly data, one rotation is performed from January to December, so 12) is shown. (The number of data / analysis cycle is used when extracting values from the correlation coefficient test table / F distribution table.) FIG. 13 is a table diagram of the annual trend corresponding demand actual file (1017) for the past 10 years. Actual demand values are accumulated on a yearly basis.

FIG. 14 shows a table of a monthly trend corresponding demand actual result file (1019), in which actual demand values for the past December are accumulated on a monthly basis.

FIG. 15 is a table diagram of a yearly monthly patternability correspondence result file (1021), in which demand actual values for the past three years are accumulated on a monthly basis.

FIG. 16 shows a table of an annual index smoothing correspondence performance file (1022), in which actual demand values for the past 20 months are accumulated on a yearly basis.

FIG. 17 shows a table of a monthly index smoothing correspondence performance file (1023), in which demand actual values for the past 46 months are accumulated on a monthly basis.

FIG. 18 shows a table of a trend prediction model parameter transaction file (1024), in which parameters calculated by regression analysis when performing trend analysis are stored.

FIG. 19 is a table diagram of a transaction file (1025) for the judgment result of the trend / pattern property. The calculation result of the correlation coefficient or F value, the test table or the F distribution table is extracted for each analysis section. Values and judgment results are stored.

FIG. 20 shows a correlation coefficient test table master (102
FIG. 7B shows a table diagram in which the values of the correlation coefficient test table are accumulated.

FIG. 21 is a table diagram of the pattern index calculation transaction file (1029), in which the values and the row totals of the result files as shown in FIGS.

FIG. 22 shows a table of the pattern index calculation transaction file (1031), in which pattern ratios are accumulated.

FIG. 23 shows a table of the F distribution table master (1033), in which the values of the F distribution table are accumulated.

FIG. 24 shows a table of the trend / pattern property determination result / applied model correspondence master (by year) (1035). The corresponding master. Note that “*” in the application condition indicates a wild card, and both “○” and “×” apply. Also, the numbers (such as) in the processing method indicate the order of processing.
(For example, in the case of the model 2-3, the processing of “annual forecast value inclination monthly division” is performed after the “annual trend upper annual forecast value calculation” processing.) FIG. Result / Applied model correspondence master (monthly) (103
FIG. 6B is a table diagram showing a table corresponding to the application condition of the trend / pattern property and the processing function of the monthly prediction model.

FIG. 26 shows a theoretical value file for evaluation (103
7 shows a table diagram of 7), in which theoretical values for evaluation are accumulated. Here, the theoretical value for evaluation is a value predicted by the prediction model for the past (for which there is actual data) in order to evaluate the reliability of the prediction model.

FIG. 27 shows a table of the evaluation value file (1039), in which evaluation values are accumulated for each model.

FIG. 28 shows a future predicted value file (104).
1 shows a table diagram of 1), in which future predicted values are accumulated.

FIG. 29 is a screen for selecting an optimal prediction model from the candidate models by the end user.

FIG. 30 is a screen for outputting the prediction value calculated in the prediction model.

FIGS. 2 to 10 show flowcharts for carrying out the present invention. Based on this, the flow of the processing of the present invention will be described.

FIG. 2 is a main flowchart.

"Trend / pattern property determination" (step 2101) is a process of performing trend analysis and pattern property analysis based on demand actual information to determine whether there is a trend or pattern property. (See FIG. 3 for details). The “selection of prediction model” (step 2103) selects a demand prediction model candidate from the result of the trend / pattern property determination. Then, in order to evaluate the candidate model, a prediction value (theoretical value for evaluation) when each candidate model is applied to the past is calculated, and the evaluation value is calculated by comparing with the actual demand actual. This is a process for selecting an optimal prediction model from the result. (See FIG. 9 for details).

"Calculation of future predicted value" (step 2105)
Is a process of calculating a predicted value for the future using the selected prediction model. (See FIG. 10 for details).

FIG. 3 "Trend / Pattern presence / absence judgment"
Is a process of performing trend analysis and pattern analysis based on demand actual information, and determining whether there is a trend or pattern.

The "repeat of trend presence / absence analysis" (step 2201) is a process of repeating all kinds of trend presence / absence analysis. The types of trend presence / absence analysis include types of functions (linear, quadratic curves, etc.) and prediction units (annual, monthly, weekly, daily, etc.) when using regression analysis.

In the detailed description of the invention, only the prediction unit for the year and month is described, but the same processing is performed for the week and day if necessary.

"Demand demand file" (step 220)
3) FIGS. 13 and 14 are for performing trend analysis.
This is a process of reading a corresponding demand result file from the demand result file as described above.

"Trend Presence Analysis" (Step 220)
5) is a process for determining the presence or absence of a trend. (See FIG. 4 for details).

"Repeat of pattern property presence / absence analysis" (step 2207) is a process of repeating various pattern property presence / absence analyzes. The types of pattern presence / absence analysis include monthly in the year, weekly in the month, and the like depending on the prediction unit / cycle.

In the detailed description of the invention, only the prediction unit for the year and month is described, but the same processing is performed for the week and day if necessary.

"Demand actual file" (Step 220)
9) is a process of reading a corresponding demand record file from a demand record file as shown in FIG. 15 in order to perform pattern property analysis.

"Analysis for Presence or Absence of Pattern" (Step 221)
1) is a process for determining the presence or absence of pattern characteristics. (See FIG. 6 for details).

FIG. 4 "Trend presence / absence analysis" is a process for determining the presence / absence of a trend.

The "regression analysis" (step 2301) is a process of performing a regression analysis based on the actual demand file as shown in FIGS. 13 and 14 and registering the calculated parameters in the trend prediction model parameter transaction file (FIG. 18). It is. The regression analysis used here is a general method of statistics, and is described in “Introduction to Prediction Method: Doyukan” or the like.

"Calculation of correlation coefficient r" (step 2303)
Calculates the correlation coefficient based on the actual demand file as shown in FIGS.
This is the process of registering the value. In addition, used here,
The calculation of the correlation coefficient r is a general technique of statistics, and is described in “Statistics for Beginners: Kyoritsu Shuppan Co., Ltd.” or the like.

"Registration of trend presence / absence determination" (step 2305) compares the value of "correlation coefficient" of the trend / pattern presence / absence determination result file (FIG. 19) with the value of the correlation coefficient test table. This is a process of registering the information to be performed in the “determination result” of the trend / pattern property determination result file (FIG. 19). (See FIG. 5 for details).

FIG. 5 “Registration of presence / absence determination of trend” compares the “correlation coefficient” value in the trend / pattern presence / absence determination result file (FIG. 19) with the value of the correlation coefficient test table. Is registered in the “determination result” of the trend / pattern property determination result file (FIG. 19).

"Extraction of Values of Correlation Coefficient Test Table" (Step 2)
401) extracts the “data number” corresponding to the currently performed analysis category from the trend / pattern / exponential smoothing analysis master (FIG. 12), and obtains a correlation coefficient test table master (FIG. 2).
From 0), extract the “test table value” corresponding to the number of data,
This is a process of registering in the “test table value or F distribution table value” of the trend / pattern property determination result file.

"Comparison of values of correlation coefficient r and correlation coefficient test table"
(Step 2403) compares “correlation coefficient r” and “test table value” of the result of the trend / pattern property determination (FIG. 18), and if correlation coefficient ≧ test table value, step 240
5. If the correlation coefficient <test table value, the processing of step 2407 is performed.

"Register ○" (step 2405) is a process for registering に in "determination result" of the trend / pattern property determination result file (FIG. 18).

"Register x" (step 2407) is a process for registering x in "determination result" of the trend / pattern property determination result file (FIG. 18).

FIG. 6 "pattern property presence / absence analysis" is a process for determining the presence or absence of pattern property.

"Pattern index calculation" (step 250)
1) is a process of calculating a pattern index based on a demand record file as shown in FIG. 15 and registering it in a pattern index calculation transaction file (FIG. 22). (Detail is,
(See FIG. 7).

The "F value calculation" (step 2503) is a pattern index calculation transaction file (FIG. 22).
The F-value is calculated based on the “pattern ratio” of the above, and the transaction file (FIG. 1)
This is the process of registering the “correlation coefficient or F value” of 9). The calculation of the F value used here is a general method of statistics, and is described in “Statistics for Beginners: Kyoritsu Shuppan Co., Ltd.” or the like.

"Registration of trend presence / absence determination" (step 2305) compares the value of "F value" in the trend / pattern property determination result file (FIG. 19) with the value of the F distribution table, and finds the corresponding one. / This is a process of registering in the “determination result” of the pattern existence determination result file (FIG. 19). (See FIG. 8 for details).

FIG. 7 “Calculation of pattern index”, a pattern index is calculated based on the actual demand file as shown in FIG. 15, and a pattern index calculation transaction file (FIG. 22).
This is the process of registering in.

"Calculation of total value of each row" (step 2601)
Is a process of calculating the total value (row total value) for one cycle based on the demand actual file as shown in FIG. 15 and registering it in the pattern index calculation transaction file (FIG. 21).

"Each line pattern ratio calculation" (step 26)
03) is a process of calculating the actual demand value (ij) ÷ the total value of one cycle (j) based on the pattern ratio calculation transaction file (FIG. 21) and registering it in the pattern index calculation transaction file (FIG. 22). .

"Calculation of average pattern ratio" (step 26)
05) is a process of calculating the average value of each line pattern ratio based on the pattern index calculation transaction file (FIG. 22) and registering the average value in the pattern index calculation transaction file (FIG. 22).

FIG. 8 “Registration of trend presence / absence judgment result” compares the value of “F value” of the trend / pattern presence / absence determination result file (FIG. 19) with the value of the F distribution table, and stores the corresponding result in the trend / pattern. Sex determination result file (Fig. 19)
This is a process of registering in the “judgment result”.

"Extraction of F distribution table value" (step 270)
1) Extracts “the number of data” and “analysis cycle” corresponding to the currently performed analysis category from the trend / pattern / exponential smoothing analysis master (FIG. 12), and extracts “molecule degree of freedom α =
The analysis period -1 "" the denominator degree of freedom β = the number of data -1-the numerator degree of freedom α "is calculated. From the F distribution table master (Fig. 23)
Extract “F distribution table value” corresponding to “numerator degree of freedom α” and “denominator degree of freedom β” and register it in “Test table value or F distribution table value” in the trend / pattern property presence / absence judgment result file (FIG. 19). I do.

"Comparison between F value and F distribution table value" (step 2)
703) is the trend / pattern property determination result (FIG. 1).
9) The “F value” and the “F distribution table value” are compared, and if F value ≧ F distribution table value, the process of step 2705 is performed; if F value <F distribution table value, the process of step 2707 is performed.

“Register ○” (step 2705) is a process for registering に in the “determination result” of the trend / pattern property determination result file (FIG. 19).

"Register x" (step 2707) is a process for registering x in "determination result" of the trend / pattern property presence / absence determination result file (FIG. 19).

FIG. 9 “Selection of prediction model” selects a demand prediction model candidate from the result of the trend / pattern property determination. Then, in order to evaluate the candidate model, a prediction value when each candidate model is applied to the past is calculated,
The evaluation value is calculated in comparison with the actual demand results. This is a process for selecting an optimal prediction model from the result.

"Predictive specification analysis unit" (step 28)
01) is a process for branching for each target prediction model because the target prediction model differs depending on the unit of analysis to be predicted. Referring to the “predicted unit” in the predicted product master (FIG. 11), the process of step 2803 is performed in the case of the year unit, and the process of step 2811 is performed in the case of the month unit.

In the detailed description of the invention, only the year / month unit is described. However, the same processing as in steps 2803 to 2809 is performed for the week / month unit if necessary.

The "repeated yearly prediction model" (step 2803) is a process of determining a conformity condition for all "predicted models" of the trend / pattern property determination result / applied model correspondence master (yearly) (FIG. 24). It is.

The “comparison of the trend / pattern property judgment result with the matching condition” (step 2805) is a transaction file of the trend / pattern property judgment result transaction file (FIG. 19).
"Judgment result" and "Conformance condition" of trend / pattern property judgment result / applicable model correspondence master (by year) (Fig. 24)
If they match, the processing of steps 2807 and 2809 is performed.

Step 2805 will be described as a specific example in the case where the judgment result of the annual trend is “○”. "Trend / Pattern Judgment Result / Applicable Model Support Master"
From FIG. 24, the “application conditions” are “model 1: ○” and “model 2: * (* is a wild word and both XX can be used)”. 2 "Both are extracted as condition matching.

"Calculation of theoretical value for evaluation" (step 280)
7) is the “processing method” of the trend / pattern property determination result / applied model correspondence master (yearly) (FIG. 24), and “the number of data” of the trend / pattern property analysis master (FIG. 12).
This is the process of calculating the theoretical value in the past for the part (the part with the actual value) and storing it in the theoretical value file for evaluation (FIG. 26).

Step 2807 will be described in a concrete example. The method of calculating the theoretical value of “model 2” is based on the trend prediction model parameter transaction file (FIG. 18).
The theoretical value is calculated with reference to the parameters.

The theoretical value of the “model 1” is calculated by the exponential smoothing method based on the “annual index smoothing demand actual result file” (FIG. 16).

“Calculation of logical value” (step 2809)
Based on the theoretical value file for evaluation (FIG. 26) and the actual demand file as shown in FIG. 13, the evaluation value = {Σ | logic value for evaluation (i) | ÷ | logic value for evaluation (i) −actual demand value ( i) |} ÷ This is a process of calculating the predicted data number t and storing it in the evaluation value file (FIG. 27).

Here, if it is not desired that the prediction deviates greatly at a certain point in time, [{| logical value for evaluation (i) | ² {predicted value (i) −actual demand value (i)} ² ]} The predicted data number t is set.

The "repeated monthly prediction model" (step 2811) is a process of determining a matching condition for all "predicted models" of the trend / pattern property determination result / applied model correspondence master (monthly) (FIG. 25). It is.

The “comparison of the trend / pattern property determination result with the matching condition” (step 2813) is a transaction file of the trend / pattern property determination result file (FIG. 19).
"Judgment result" and trend / pattern judgment result / applicable model correspondence master (monthly basis) (Fig. 25)
If they match, the processing of steps 2815 and 2817 is performed.

Step 2813 will be described as a specific example in the case where the judgment result of the annual trend and the monthly pattern pattern within the year is "O" and the monthly trend is "X". Since the monthly trend is “x”, it is understood that “model 0-2” is not applied. In addition, "Model 2-3" and "Model 0-2"
For “Model 0-1”, since the conditions apply,
“Model 2-3”, “Model 1-3”, and “Model 0-1” are extracted as condition conformance.

"Calculation of theoretical value for evaluation" (step 281)
5) “Processing method” of the trend / pattern characteristic determination result / applicable model correspondence master (monthly) (FIG. 25), and “data count” of the trend / pattern characteristic analysis master (FIG. 12)
This is the process of calculating the logical value in the past for the part (the part with the actual value) and storing it in the theoretical value file for evaluation (FIG. 26).

Step 2815 will be described with a specific example. The theoretical value calculation method of “model 2-3” first refers to the annual trend parameters from the trend prediction model parameter transaction file (FIG. 18).
Calculate the theoretical value of annual gross. Then, referring to the pattern ratio from the “pattern index calculation transaction file”, the theoretical value is calculated on a monthly basis.

The method of calculating the theoretical value of “model 1-3” is as follows: “Annual index smoothing demand actual result file” (FIG. 16)
Calculates the theoretical value of annual gross by the exponential smoothing method based on Then, referring to the pattern ratio from the “pattern index calculation transaction file”, the theoretical value is calculated on a monthly basis.

The method of calculating the theoretical value of “model 0-2” refers to the monthly trend parameter from the trend prediction model parameter transaction file (FIG. 18).
Calculate the theoretical value.

The method of calculating the theoretical value of “model 0-1” is as follows.
The theoretical value of annual gross is calculated by the index smoothing method based on the “monthly index smoothing demand actual performance file” (FIG. 17).

The “calculation of logical value” (step 2817)
Based on the theoretical value file for evaluation (FIG. 26) and the actual demand file as shown in FIG. 14, the evaluation value = {Σ | logic value for evaluation (i) | ÷ | logic value for evaluation (i) −actual demand value ( i) |} ÷ This is a process of calculating the predicted data number t and storing it in the evaluation value file (FIG. 27).

Here, if it is not desired that the prediction deviates greatly at a certain point in time, [{| logical value for evaluation (i) | ² {predicted value (i) −actual demand value (i)} ² ]} The predicted data number t is set.

"Extraction of applied model" (step 2819)
Is based on demand actual information as shown in FIG. 13, a theoretical value file for actual evaluation (FIG. 26), and an evaluation value file (FIG. 27).
A prediction model selection screen (FIG. 29) is displayed, the end user is allowed to select an optimal prediction model, and a selection result is input.
This is a process of registering in the “selection model result” of the evaluation value file (FIG. 27). Here, when it is desired that the computer automatically selects a model, a model having the highest evaluation value is automatically selected.

FIG. 10 "Calculate future predicted value" is a process of calculating a predicted value for the future using the selected prediction model.

The “calculation of the predicted value by the applied model” (step 2901) extracts the predicted period from the “predicted period” of the predicted product master (FIG. 11), and executes the “evaluation value file” (FIG. 2).
This is a process of calculating a prediction value for a prediction period using the prediction model selected from the “selection model result” of 7).

The “predicted value output” (step 2903)
This is a process of displaying a prediction result output screen (FIG. 30) based on the calculated prediction value.

[0088]

As described above, according to the present invention,
Forecasting can be performed using a forecasting model that meets the demand trend in a timely manner for each product, and the prediction accuracy can be improved.

[Brief description of the drawings]

FIG. 1 is a system configuration diagram for performing the present invention.

FIG. 2 is a main flowchart.

FIG. 3 is a flowchart of a trend / pattern property determination.

FIG. 4 is a flowchart of a trend presence / absence analysis.

FIG. 5 is a flowchart of trend presence / absence determination registration.

FIG. 6 is a flowchart of a pattern presence / absence analysis.

FIG. 7 is a flowchart of a pattern index calculation.

FIG. 8 is a flowchart of a pattern property presence / absence result registration.

FIG. 9 is a flowchart of prediction model selection.

FIG. 10 is a flowchart of future predicted value calculation.

FIG. 11 is a table diagram of a predicted product master.

FIG. 12 is a table diagram of trend / patternability / exponential smoothing analysis.

FIG. 13 is a table showing an annual trend responsive demand record.

FIG. 14 is a table showing monthly trend-response demand results.

FIG. 15 is a table showing a demand result corresponding to a monthly pattern within the year.

FIG. 16 is a table of annual index smoothing corresponding demand results.

FIG. 17 is a table of monthly index smoothing corresponding demand results.

FIG. 18 is a table diagram of a trend prediction model parameter transaction.

FIG. 19 is a table diagram of a result of trend / pattern property determination.

FIG. 20 is a table diagram of a correlation test table master.

FIG. 21 is a diagram showing an example of a pattern index transaction file.

FIG. 22 is a diagram showing an example of a pattern index transaction file.

FIG. 23 is a table diagram of an F analysis table master.

FIG. 24 shows a trend / pattern property determination result. It is a table figure of an application model correspondence master (year unit).

FIG. 25: Trend / pattern property determination result. It is a table figure of an application model correspondence master (month unit).

FIG. 26 is a table of theoretical values for evaluation.

FIG. 27 is a table diagram of evaluation values.

FIG. 28 is a table of future predicted values.

FIG. 29 is an image diagram of a prediction model selection screen.

FIG. 30 is a screen diagram showing a prediction result output.

[Explanation of symbols]

1001 ... control program, 1003 ... work area,
1005: Prediction model module, 1007: Trend / pattern presence / absence judgment module, 1009: Prediction model selection module, 1011: Future prediction value calculation module, 1013: Prediction product master, 1015: Trend /
Pattern / exponential smoothing analysis master, 1017 ... Actual trend-response file, 1019 ... Monthly trend-response file, 1021 ... Monthly monthly pattern-response result file, 1022 ... Annual index smoothing-response file, 1
022: Annual index smoothing-response result file, 1023: Monthly index smoothing-response file, 1024 ... Trend prediction model parameter transaction file, 1025
... Trend / pattern property determination result transaction file, 1027 ... Correlation coefficient test table master, 1029
… Pattern index calculation transaction file, 1
031: Pattern index calculation transaction file, 1033: F distribution table master, 1035: Trend /
Pattern judgment result / applied model correspondence master (year unit),
1036: trend / pattern determination result / applied model correspondence master (monthly basis), 1037: theoretical value file for evaluation, 1039: evaluation value file, 1041: future predicted value file.

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Takao Sato 890 Kashimada, Saiwai-ku, Kawasaki-shi, Kanagawa Prefecture Inside the System Development Division, Hitachi, Ltd. No. 1 Hitachi Tohoku Software Co., Ltd. F term (reference) 5B049 BB11 CC27 EE03

Claims (7)

[Claims] Claims: 1. Analyze the actual demand of a product and determine the trend /
The step of determining the presence or absence of the pattern property, the step of selecting a demand prediction model candidate from the result and calculating the evaluation value in the candidate model, the step of selecting an appropriate model, and the step of A model selection type demand forecasting method for each forecast purpose, comprising a calculating step. 2. The method according to claim 1, further comprising the step of performing all of a trend presence / absence analysis and a pattern presence / absence analysis in accordance with a prediction unit. Type demand forecasting method. 3. The trend presence / absence analysis according to claim 2, wherein the demand actual information of the product is regression-analyzed, a correlation coefficient is calculated, and the trend is determined by comparing with a value of a correlation coefficient test table. A model selection type demand forecasting method for each forecast purpose, characterized by: 4. The method according to claim 2, wherein:
A model selection type demand forecasting method for each prediction purpose, wherein a pattern index is calculated from actual demand information of a product, an F value is calculated, and the presence or absence of a pattern characteristic is determined by comparing the F value with a value in an F distribution table. 5. The method according to claim 1, wherein a predicted value of the candidate model is displayed in a graph or an evaluation value is displayed so that a user can easily select an appropriate model. Demand forecasting method by model selection for each forecast purpose. 6. A demand forecasting method according to a forecast purpose, wherein the method is capable of not only selecting a model by a user but also automatically selecting a model having a high evaluation value. . 7. In claim 5 and claim 6, the evaluation value is expressed as ｛Σ | Demand actual value (i) | ÷ | Predicted value (i) −Demand actual value (i) |｝ ÷ Predicted data number t or [ Σ | Demand actual value (i) | ² ÷ ｛Predicted value (i) −Demand actual value (i)｝ ² ] ÷ Model selection type demand for each prediction purpose, which is a method of calculating by the predicted data number t Forecasting method.