Question:
Hi folks,
I am a novel to demand planning of APO. I wanted some conceptual help regarding statistical techniques. Any forecasting model calcutates lot of forecast error parameters like Mean, MAD, Variance, ET, Mean percent error, Mean absolute percent error, Mean square error, Root Mean Square Error etc.. I wanted to know what is the significance of each of these parameter relative to every other in understanding the sanctity of historical data? What immediate inference can we draw about the population from the sample based on these parameters and can take judgement about whether the given forecast how best suits the historical pattern? Why we have to calculate so many parameters? Why not any one?
Any pointers on this will be appreciated. If anyone knows good website on this place send me the URL.
thanks and regards,
mnk
Answer:
Mnk,
I suggets that you get a user with a stat forecasting background to help you!
The Mean Error (ME) measures the average error in fit or forecast. However, note that averages are insufficient in choosing one model over another. A model that has a lower Mean Error is desirable if it also has
a lower MSE or RSE. However, a mean error of zero may result from extraordinarily high positve and negative errors that cancel each other out.
The Mean Squared Error (MSE) is a measure of the variance of the errors. Errors are squared in order to avoid the high positives errors cancelling out "high" negative errors. Some find the use of the variance less meaningful than the use of the square root of the variance which is a standard-deviation measure.
Std. Dev. (Residual Standard Error, RSE) is a measure that approximates the scatter or dispersion in forecast errors much as the standard devition does in classical statistics. For example: the following formula can be used to generate prediction intervals (confidence intervals) assuming errors are
approximately normally distributed with a mean of zero:
Actual Future Value = Forecast +/- Z*RSE
where Z is the usual standard normal deviate used in generating confidence intervals. A Z of 1.96 yielding a 95% prediction interval when the sample size is greater than 30.
The Mean Percent Error (MPE) is the mean of several Percent Errors (PE = Error/Actual) Mean Percent Error has some of the properties of the Mean Error, however it is expressed in percentages. The MPE can vary dramactically if the denominator (the actual values) are very low numbers relative to the numerator.
Percent Error = PE = Error/Actual
If the actual is extremely low relative to the Error, then MPE approaches infinity, either negative or positive infinity depending on whether the Error is positive or negative. The MPE is unreliable when extreme values of the actual remain in its calculations. However, it can provide useful information.
The Mean Absolute Percent Error (MAPE) is the mean of several Absolute Percent Errors (Absolute Value of Error/Actual) Mean Absolute Percent Error is a very good relative measure of forecasting accuracy because high and low forecast errors do not cancel each other out. It is a useful statistics in that if errors are relatively normally distributed then we can interpret the MAPE as the value about which 50% of the errors are above and below.
Absolute Percent Error = APE = ABS(Error)/Actual
If the actual is extremely low relative to the Error, then MPE approaches infinity, either negative or positive depending on whether the Error is positive or negative. The MPE is unreliable when extreme values of the actual remain in its calculations.
Hope this helps
Haribo
Answer:
Hi haribo,
It was really thought provoking information from you. Can you also suggest some good books or a website where i can get such information.
thanks and regards,
mnk
Answer:
Do a search on google, for example;
http://www.stat-help.com/links.html[/url]
Answer:
The links Haribo56 provided should cover most things. The online help is pretty good on the individual methods for calculating in DP. The book that SAP actually took the forecast calculations from is "Forecasting Methods and Applications" by Makridakis, Wheelwright & Hyndman. This is a relatively ok introduction but unfortunately with this area of DP practice is everything.
Get someone from the business who really knows the demands well and select a really fundamental product with a well known demand profile to begin on. Run through a series of forecasts to find one that comes closest to your demand profile (try to avoid causal or multivariate forecasts and stick to stad univariate at this early stage). try to use graphical comparison to start then look at the individual numbers in detail. A "feature" of interactive forecasting under 3.0a (and 3.1 I think) to watch out for is that it will save your forecast profile settings as soon as you exit the interactive screen and overwrite your defaults. Look in OSS for a fix based on your user account settings to stop this as it really annoys.
Answer:
I can only back up James with regards to getting someone from the business who really knows the demands well.