The art of forecasting in the SAP F & R system for inventory management
SAP F & R (Forecasting & Replenishment) is a system for scheduling orders and forecasting demand for generating order projects at the supplier shop level. The system is part of the SAP SCM (Supply Chain Management) solution and is implemented in two variations:
In this post, we consider the possibility of calculating the average forecast in the SAP F & R system.
The average forecast in the terminology of SAP F & R is such a value of the volume of goods on location, which with a probability of 50% will satisfy customer demand in the store, or, in other words, will ensure the level of customer service = 50%.
In order to avoid lost sales, the target level of customer service, as a rule, is planned to be not lower than 95%. This means that in 95 cases out of 100 the customer will buy what he planned in the store. Ensuring a high level of service in SAP F & R is made by means of an insurance premium to the average forecast, which depends not only on the target level of service, but also on the variability of past values of sales of goods. Thus, a maximum sales forecast is formed in the system, the volume of which will be sufficient to minimize stock in a warehouse or in a store (and, therefore, withdraw frozen capital in stocks) and meet the target level of customer service.
')
To build an average forecast in the SAP F & R system, sales forecasting models are used, taking into account not only static consumption data, but also the influence of external factors, such as calendar events or promotions. The effect of such factors can be set either manually or automatically detected by the system in the past. Thus, when forming a predictive model of a time series, both in the past and in the future, SAP F & R uses data on a possible change in the predicted value and imposes the effect of an external factor on the smoothed series.
As can be seen in the figure, when forming a forecasting model in the past, SAP F & R clearly identified seasonal sales fluctuations and peak spikes in the pre-New Year period. However, not every historical data changes naturally, and very often when forecasting, you can face the situation below:
How in such a case should the average forecast be considered so that its reliability is high enough?
Such time series are a random Poisson distribution, so the probability that the sales will be m pieces is calculated using the formula:
Despite the fact that the Poisson series is very easy to see - they are quite difficult to identify. In the above example, the value of m is the best predicted value (forecast A). Zero predicted values (or one for the last case) are the median of this time series, i.e. 50% chance of meeting customer demand (forecast B).
To determine the most relevant forecast value, we will check the forecast error when using the Poisson and median distribution to calculate average sales using the wMAPE formula - weighted average absolute error in percent:
Thus, we get the forecast error:
In each case, the forecast with the wMAPE error is lower than expected. Therefore, the use of wMAPE in this case almost always leads to the wrong result.
Then we will resort to the calculation of the error using the MSE (mean square error) and sRMSE (root of the mean square error) formulas. To calculate the mean square error, we use the formula:
The results are as follows:
In this case, it is more preferable to use the probability of obtaining a value from the Poisson distribution for prediction. Now we calculate the error by the formula sRMSE, using the formula:
The results are as follows:
Thus, it is obvious that the systematic use of the calculation of the error using the wMAPE method leads to a decrease in the accuracy of the forecast, and when detecting the Poisson distribution, it is impractical to estimate the error using methods such as wMAPE, MAE (mean absolute error), MAD (mean absolute deviation), MASE ( mean absolute scaled error).
This effect is often manifested when forecasting sales of goods with a low turnover rate (less than 0.2 units per day). With a similar turnover, MSE and sRMSE are the best methods for estimating sales forecasting errors. At the same time, it is necessary to remember about the constant monitoring of the level of inventory and the assessment of the shortage of goods, because the accuracy of forecasting is only a means.
- SAP F & R SCM - implementation with seamless integration with SAP systems;
- SAP F & R OI - system for integration with non-SAP systems.
In this post, we consider the possibility of calculating the average forecast in the SAP F & R system.
The average forecast in the terminology of SAP F & R is such a value of the volume of goods on location, which with a probability of 50% will satisfy customer demand in the store, or, in other words, will ensure the level of customer service = 50%.
In order to avoid lost sales, the target level of customer service, as a rule, is planned to be not lower than 95%. This means that in 95 cases out of 100 the customer will buy what he planned in the store. Ensuring a high level of service in SAP F & R is made by means of an insurance premium to the average forecast, which depends not only on the target level of service, but also on the variability of past values of sales of goods. Thus, a maximum sales forecast is formed in the system, the volume of which will be sufficient to minimize stock in a warehouse or in a store (and, therefore, withdraw frozen capital in stocks) and meet the target level of customer service.
')
To build an average forecast in the SAP F & R system, sales forecasting models are used, taking into account not only static consumption data, but also the influence of external factors, such as calendar events or promotions. The effect of such factors can be set either manually or automatically detected by the system in the past. Thus, when forming a predictive model of a time series, both in the past and in the future, SAP F & R uses data on a possible change in the predicted value and imposes the effect of an external factor on the smoothed series.
As can be seen in the figure, when forming a forecasting model in the past, SAP F & R clearly identified seasonal sales fluctuations and peak spikes in the pre-New Year period. However, not every historical data changes naturally, and very often when forecasting, you can face the situation below:
How in such a case should the average forecast be considered so that its reliability is high enough?
Such time series are a random Poisson distribution, so the probability that the sales will be m pieces is calculated using the formula:
Despite the fact that the Poisson series is very easy to see - they are quite difficult to identify. In the above example, the value of m is the best predicted value (forecast A). Zero predicted values (or one for the last case) are the median of this time series, i.e. 50% chance of meeting customer demand (forecast B).
To determine the most relevant forecast value, we will check the forecast error when using the Poisson and median distribution to calculate average sales using the wMAPE formula - weighted average absolute error in percent:
Thus, we get the forecast error:
In each case, the forecast with the wMAPE error is lower than expected. Therefore, the use of wMAPE in this case almost always leads to the wrong result.
Then we will resort to the calculation of the error using the MSE (mean square error) and sRMSE (root of the mean square error) formulas. To calculate the mean square error, we use the formula:
The results are as follows:
In this case, it is more preferable to use the probability of obtaining a value from the Poisson distribution for prediction. Now we calculate the error by the formula sRMSE, using the formula:
The results are as follows:
Thus, it is obvious that the systematic use of the calculation of the error using the wMAPE method leads to a decrease in the accuracy of the forecast, and when detecting the Poisson distribution, it is impractical to estimate the error using methods such as wMAPE, MAE (mean absolute error), MAD (mean absolute deviation), MASE ( mean absolute scaled error).
Conclusion
This effect is often manifested when forecasting sales of goods with a low turnover rate (less than 0.2 units per day). With a similar turnover, MSE and sRMSE are the best methods for estimating sales forecasting errors. At the same time, it is necessary to remember about the constant monitoring of the level of inventory and the assessment of the shortage of goods, because the accuracy of forecasting is only a means.
Source: https://habr.com/ru/post/307842/
All Articles