JP2001022729A - Method for selecting prediction model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve prediction precision by eliminating an operation man-hour for selecting prediction models to be applied to respective conventional articles. SOLUTION: An optimum threshold determination part 5 actuates a prediction execution part 6, an absolute error rate calculation part 7, and an autocorrelation coefficient calculation part 8 to calculate prediction data of respective periods wherein prediction models A and B are applied as to time- series actual result data 11, calculate absolute error rates of the respective periods by the models, and calculate autocorrelation coefficients of the respective periods, and then select the models to be applied as to thresholds and decides as an optimum threshold the threshold value having the least mean of the absolute error rates for an object period. A prediction model selection part 9 selects a prediction model to be applied according to the autocorrelation coefficient and optimum threshold of a prediction period.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a demand forecasting method using a computer, and more particularly to selecting one of a plurality of forecasting models based on past time-series actual data and applying the same to forecasting in a forecasting period. On how to do it.

[0002]

2. Description of the Related Art When a product is manufactured, stored and transported in a planned manner, it is necessary to predict the demand for the product with high accuracy. Conventionally, several forecasting models have been used for demand forecasting. Representative prediction models can be classified into a prediction model that considers seasonality or periodicity, and a prediction model that does not consider seasonality or periodicity.

[0003]

While demand trends of individual products show seasonal fluctuations, some products show little seasonality. Therefore, in order to perform demand forecast with high accuracy, it is necessary to determine which forecast model is applied to each product. Conventionally, a prediction model is selected based on human judgment, so that a large number of types of products to be handled requires a large number of man-hours. Therefore, for products that once had a seasonality tendency, the same prediction model is continued to be applied even if the seasonality is lost, resulting in a decrease in prediction accuracy. Conversely, for products that applied a prediction model that did not consider seasonality, the same prediction model was continued to be applied even if seasonality appeared later, resulting in a similar decrease in prediction accuracy.

An object of the present invention is to solve the above problems, and to provide a method of automatically selecting a prediction model.

[0005]

SUMMARY OF THE INVENTION The present invention is a method for calculating prediction data in a prediction period from past time-series actual data using a computer, and is a method applied when the actual data has periodicity. A method of selecting one of the first prediction model and the second prediction model applied when the periodicity of the actual data is denied, and applying the selected prediction model to the prediction of the prediction period. For the actual data over the period, the prediction data of each period when the first prediction model and the second prediction model are applied are calculated. (2) The actual data and the prediction data of (1) are calculated for each model and for each period. And (3) calculate the autocorrelation coefficient from the past time-series actual data for each period of (1), and (4)
A plurality of thresholds are set, and a model to be applied is selected based on a comparison between each threshold and the autocorrelation coefficient of each period in (1), and the absolute error rate when the selected model is applied is calculated. The absolute error rates extracted for each threshold value of (5) and (4) are averaged to calculate an average absolute error rate over a plurality of periods, and the average absolute error rates of (6) and (5) are minimum. The threshold is extracted as the optimal threshold, and (7) the autocorrelation coefficient is calculated from the past time-series actual data for the prediction period,
(8) A method of selecting a prediction model to select a prediction model to be applied to the prediction period based on a comparison between the autocorrelation coefficient of (7) and the optimum threshold value of (6).

[0006]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a diagram showing a configuration of a computer according to the present embodiment. The computer includes a processing device 1 and a storage device connected to the processing device 1, and the storage device stores time-series actual data 11 and prediction data 12. The time series performance data 11 is past demand performance data arranged in time series for each product. The prediction data 12 is data predicted by the prediction model for the next period for each product.

The main memory of the processing device 1 stores a prediction model A (2), a prediction model B (3), and a prediction model selection program 4, and is executed by the processing device 1.
The prediction model A (2) is a prediction model applied when demand is considered to fluctuate periodically or seasonally. The prediction model B (3) is a prediction model applied when demand does not have periodicity or seasonality. For example, as the prediction model B (3), there is a prediction model in which a numerical value obtained by performing a weighted average of the actual data of several periods immediately before the prediction period is used as the prediction data.

The prediction model selection program 4 has respective program means of an optimum threshold value determination unit 5, a prediction execution unit 6, an absolute error rate calculation unit 7, an autocorrelation coefficient calculation unit 8, and a prediction model selection unit 9. The optimum threshold value determination unit 5 reads the time-series actual data 11 for the specified product, activates the prediction execution unit 6, the absolute error rate calculation unit 7, and the autocorrelation coefficient calculation unit 8 to optimize the prediction period. Is calculated. The prediction execution unit 6 adds the prediction model A (2) to the actual data.
Alternatively, prediction data is calculated by applying the prediction model B (3). The absolute error rate calculation unit 7 calculates an absolute error rate when predictive data in a period in which actual data is known is calculated. The autocorrelation coefficient calculation unit 8 calculates an autocorrelation coefficient in a prediction period from past actual data. Prediction model selector 9
Selects and activates either the prediction model A (2) or the prediction model B (3) to be applied to the prediction period from the calculated optimal threshold value and the autocorrelation coefficient of the prediction period, Is calculated, and the result is stored in the prediction data 12.

FIG. 2 is a diagram showing a data structure of the time-series performance data 11. The time-series performance data 11 is provided for each product indicated by the product code, and stores demand performance data of the product over a plurality of periods. In this example, it is assumed that the actual data of each period from the first period to the 29th period has been obtained, and if the actual data has periodicity or seasonality, the cycle is regarded as the 12th period. . Second
From the fifth period to the 29th period, it is also a prediction period to be obtained when obtaining the optimum threshold value.

FIG. 3 is a flowchart showing the flow of the processing of the prediction model selection program 4 for calculating the optimum threshold value. The optimal threshold value determining unit 5 calculates the time-series actual data 1
The result data of the designated product is read from 1 (step 21). Next, the prediction execution unit 6 is started, and the prediction data is calculated by applying the prediction model A (2) for each period from the 25th period to the 29th period (step 22). That is, the forecast data of the 25th term is calculated using the actual data from the first term to the 24th term, and the forecast data of the 26th term is calculated using the actual data from the second term to the 25th term.・ Calculate the forecast data for each period as shown in. Next, the prediction execution unit 6
Is started, and the prediction data is calculated by applying the prediction model B (3) for each period from the 25th period to the 29th period (step 23). That is, the forecast data of the 25th term is calculated using the actual data of the several terms up to the 24th term, the predicted data of the 26th term is calculated using the actual data of the several terms until the 25th term, and so on. The prediction data for each period is calculated as shown in the above.
Next, the absolute error rate calculation unit 7 is started, and from the 25th period to the 29th period.
The absolute error rate when the prediction model A (2) is applied to each period up to the period is calculated (step 24). The absolute error rate is calculated by | actual data−predicted data | / actual data. Next, the absolute error rate calculation unit 7 is started, and the second
Forecast model B for each period from the 5th period to the 29th period
The absolute error rate when (3) is applied is calculated (step 25). Next, the autocorrelation coefficient calculation unit 8 is activated, and the 25th
An autocorrelation coefficient is calculated for each period from the period to the 29th period (step 26). That is, the autocorrelation coefficient for the 25th term is calculated using the actual data from the first term to the twelfth term and from the thirteenth term to the twenty-fourth term, and from the second term to the thirteenth term, , The autocorrelation coefficient for the 26th period is calculated using the actual data from the 25th period to the 25th period, and the autocorrelation coefficient for each period is calculated as in.

Next, the optimum threshold value determining section 5 sets an initial value (lower limit value) of the threshold value (step 27). The range of the threshold value to be tried is 0 to 1, for example, the initial value is 0.1.
1, increment 0.1, upper limit 0.9, or initial value 0.
2. Parameters such as step size 0.01 and upper limit value 0.99 can be set. Next, it is determined whether or not the current threshold value has exceeded the set upper limit value (step 2).
8). If not, a prediction model is selected for each period from the 25th period to the 29th period (step 29).
That is, if the autocorrelation coefficient of the period> the current threshold value, the prediction model A (2) is selected, and the absolute error rate when the prediction model A (2) is applied is adopted. Conversely, if the autocorrelation coefficient of the period ≦ the current threshold value, the prediction model B
(3) is selected, and the absolute error rate when the prediction model B (3) is applied is adopted. Next, the optimal threshold value determining unit 5
An average absolute error rate is calculated for the target period (step 30). That is, for each period from the 25th period to the 29th period, the average absolute error rate is calculated by averaging the adopted absolute error rates for the 5th period. Next, the current threshold value is incremented by increments (step 31), and step 28 is executed.
Return to When the current threshold value exceeds the upper limit value (step 28 YES), an optimum threshold value is calculated (step 3).
2). The threshold at which the average absolute error rate is minimized is the optimal threshold.

FIG. 4 is a flowchart showing the flow of processing of the prediction model selection program 4 for calculating prediction data in the prediction period using the calculated optimum threshold value. The prediction model selection unit 9 starts the autocorrelation coefficient calculation unit 8 and
The autocorrelation coefficient of the period is calculated (step 41). That is, the autocorrelation coefficient for the 30th term is calculated using the actual data from the 6th term to the 17th term and from the 18th term to the 29th term. Next, the prediction model selection unit 9 compares the autocorrelation coefficient calculated in step 41 with the optimum threshold calculated in step 32, and determines whether or not autocorrelation coefficient> optimum threshold (step S32). Step 42). If the autocorrelation coefficient> the optimum threshold, the prediction execution unit 6 is started and the prediction model A (2)
Is applied to calculate the prediction data of the 30th term (step 43). If the autocorrelation coefficient ≦ the optimal threshold, the prediction execution unit 6 is started, and the prediction model B (3) is
The period prediction data is calculated (step 44). Finally, the prediction result of the 30th term of the product calculated is referred to as prediction data 1
2 (step 45).

FIG. 5 is a diagram showing an example of data during the calculation up to the calculation of the optimum threshold value. Given the actual data of each period from the 1st to the 24th period and the actual data of each period from the 25th to the 29th period, a prediction model for each period from the 25th to the 29th period A (2) prediction result, prediction model B (3) prediction result, absolute error rate in the case of prediction model A (2), prediction model B (3)
In this case, the absolute error rate and the autocorrelation coefficient can be calculated. The threshold is set to an initial value of 0.1, a step of 0.1, and an upper limit of 1.0, a prediction model of each period is selected for each threshold, and the absolute error rate is extracted. Calculate the average absolute error rate for the threshold. The threshold at which the average absolute error rate is minimum is the optimal threshold, and in this example, 0.7 is the optimal threshold.

The absolute error rate is a numerical value indicating the prediction accuracy for each period, and the average absolute error ratio is a numerical value indicating the prediction accuracy for the target period. In the above embodiment, if the threshold value at which the prediction accuracy was the best for the target period is regarded as the optimal threshold value and is applied to the prediction period, as a result, the prediction model A or the prediction model B is selected. Regardless of whether it is done, it is based on the assumption that the best forecast data will be obtained. The reason that the target period is a period composed of a plurality of periods is to avoid a period that temporarily shows good prediction accuracy and to calculate a threshold value that shows stable prediction accuracy over a plurality of periods. .

[0016]

As described above, according to the present invention, a prediction model can be automatically selected from past time-series actual data, so that the number of man-hours for manually selecting a prediction model for each product is eliminated. In addition, the prediction accuracy of the demand prediction can be improved.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a computer according to an embodiment.

FIG. 2 is a diagram illustrating a data configuration of time-series result data 11 according to the embodiment.

FIG. 3 is a flowchart illustrating a flow of a process of calculating an optimum threshold according to the embodiment.

FIG. 4 is a flowchart illustrating a flow of a process of calculating prediction data using an optimal threshold according to the embodiment.

FIG. 5 is a diagram illustrating an example of data in the middle of calculation until an optimum threshold value is calculated.

[Explanation of symbols]

2: prediction model A, 3: prediction model B, 4: prediction model selection program, 5: optimum threshold value determination unit, 6: prediction execution unit, 7: absolute error rate calculation unit, 8: autocorrelation coefficient calculation unit , 9: prediction model selection unit, 11: time-series actual data,
12: Predicted data

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Kazutaka Matsumoto 890 Kashimada, Saiwai-ku, Kawasaki-shi, Kanagawa Prefecture, Japan Information Systems Division, Hitachi, Ltd. (72) Inventor Seiji Nakahara 3-2-1-16 Omorikita Ota-ku, Tokyo Inside System Engineering Co., Ltd. (72) Inventor Hideo Ikeda 3-2-1-16 Omorikita, Ota-ku, Tokyo Hitachi System Engineering Co., Ltd. (72) Miyoshi Nagasaki 3-2-1-16 Omorikita, Ota-ku, Tokyo Hitachi System Engineering Incorporated (72) Inventor Hiroshi Morita 3-2-1-16 Omorikita, Ota-ku, Tokyo F-term in Hitachi System Engineering Co., Ltd. 5B049 AA02 CC11 EE01 EE31

Claims (1)

[Claims] 1. A method of calculating prediction data of a prediction period from past time-series actual data using a computer,
One of a first prediction model applied when the actual data has periodicity and a second prediction model applied when the periodicity of the actual data is denied is selected and the prediction period is selected. (1) calculating the prediction data of each period when the first prediction model and the second prediction model are applied to the actual data over a plurality of periods in the past, and (2) ) Calculate the absolute error rate for each model and period for the actual data and forecast data of (1),
(3) The autocorrelation coefficient is calculated from the past time-series actual data for each period of (1), (4) a plurality of thresholds are set, and the thresholds and the self of each period of (1) are set. Selecting a model to be applied from a comparison with the correlation coefficient, extracting the absolute error rate when the selected model is applied, (5)
The absolute error rates extracted for each of the threshold values in (4) are averaged to calculate an average absolute error rate over a plurality of periods, and (6)
(5) The threshold with the smallest average absolute error rate is extracted as the optimal threshold, (7) The autocorrelation coefficient is calculated from the past time-series actual data for the prediction period, and (8) (7) A method for selecting a prediction model to be applied to the prediction period based on a comparison between the autocorrelation coefficient of (1) and the optimum threshold value of (6).