DP - Forecasting Models

[Former Member]

hi

I need to work on forecasting models in my project. I read a lot of times on this subject various articles. I am still not understanding clearly the terms like Basic, trend, initialisation, Ex-forecast.. Is Ex Forecast generated by system automatically ? if so, what we should interpret from that ? its the forecast run in the past for comparison purposes. So if i run a forecast last month (oct 07) becomes ex forecast in Nov 07 ? What & How to do the Initialization ?

Hope someone can pass the light on this

thanks

venkatesh

[Former Member · Accepted solution]

Hi Venkat,

Ex-post forecast can also be called "back-fit forecast" and is important to show you how well your forecast profile fits your history. It shows you how the system is 'learning' the pattern of the demand history. The Ex-post forecast should be compared with the Corrected History KF (if you use it) or the History KF (if you don't have a Corrected History KF in your data view). The forecast error measurements are all generated from the comparison between Ex-post and History key figures, NOT the difference between history and forecast.

WRT Initialization, all models need a few periods of history to get started. Constant models need 3 periods to initialize, or 'learn' what the starting level (basic value) should be. 3 periods is the standard number of periods for this method. Seasonal models need those three periods plus at least one seasonal cycle. First three periods are needed to determine the starting level and the first seasonal period is used to set the seasonal indices. On our instance we use a 12-month season, so the system needs 15 periods to initialize. Now let's say we have 36 periods of history with a seasonal model using 12 periods per season. In the Ex-post forecast key figure, the first 15 periods should be blank, then values will begin to calculate from period 16 up to period 36 (i.e., last month). Note that you should clear the Ex-post and Corrected History KF between stat fcst runs.

You would have to save the fcst runs in a different KF or in Excel if you want to compare units to units. However, you can use the forecast compare tool from within "Forecast In Interactive Planning: Change Mode" to see how the forecast errors measure up from run to run.

Hope this helps

[Former Member]

Hi Srinivas, Visu,

Thanks a lot for your insight. But I still have some doubts. So now I am clear that only system creates the ex-forecast value. So how system will decide from which month onwards it should decide in the past, that it should calculate ex-forecast for which we already have the forecast as history. So is initialization comes into picture here. I am not clear about this. With which KF, we should compare the ex-forecast results ? How ex-f really helps us, I mean why should we use this?

Srinivas, I am getting confused in your 2 para.

Visu, in the attached document, I don’t see topics 6 & 7. Can you pls give me those last two topics as well?

I want to try running with different models and findout the best one. How can I save my different forecast runs and compare with any later runs? As it overwrites the previous one.

Thanks a lot, I put points.

venkat

[Former Member]

Hi Venkat,

The basic idea behind using a statistical model is to analyze the historical sales trend and find a analogous mathematical equation there y use statistics to solve the equation. This equation depending on how the trend is have few constants. Like an equation y=ax1bx2cx3+..... where y is your forecast and x1, x2 and so on are independent variables and a,b,c,d are all constants.

The statistical models here are using apha, beta, gamma and permo in place of these a,b,c,d. These indexes are used based on the curve.

alpha index for basic value

Beta index for Trend

Gamma index for seasonal

persmo index for outliers

Any curve always has basic value as it has to define where it starts from the base. If the curve is heading up or down, it has a trend (upward or downward) value. If the curve has a seasonal behavior like a sine or cosine curve, it also has a seasonal index associated with it.

Basic: This value decides the vertical placement of the curve with respect to base. Higher the alpha(basic), higher the curve with respect to horizontal axis.

Trend: Specifies the slope of the line

Seasonal: Specifies divergence from basic value.

Ex-post forecast: In the history available, you date back some time and consider part of the history as future. Then the history before this so considered forecast is used to forecast the future. You get a future curve and value. But you already have the future as it is history. This helps you compare the effectiveness of the forecasting model.

Read this

<a href="https://www.sdn.sap.comhttp://www.sdn.sap.comhttp://www.sdn.sap.com/irj/servlet/prt/portal/prtroot/docs/library/uuid/7a4025f8-0a01-0010-0fb2-b7ab22597675">document</a>

This is the best source.

If you select the ex-post forecast in univariate forecast profile, the system calculates the expost forecast and can be displayed below the history to compare.

Hope this helps.

[srinivas_krishnamoorthy]

Venkat, I am hoping you have gone through the links in SDN like

https://www.sdn.sap.com/irj/sdn/go/portal/prtroot/docs/library/uuid/2bb941f8-0a01-0010-88a6-ee870ea6...

you may find more documents in "articles" section in the menu bar of the webpage. Try to search with APO+Forecast

Coming to your Q, ex-post Forecast is not the archived previous month's forecast. It is the forecast in the historical horizon based on the preceding data present in the history.So Ex post-forecast in the month of Oct 07 using the forecast model that calculates with Nov07 planning start date would be the forecast using the very same model based on history upto Sep07.

Basic Level is some kind of general Average of the Historical pattern bereft of season and trend. Similarly trend is the average increase or decrease of a historical pattern bereft of season. Any forecast model takes a few time buckets to initialize, which essentially means derivation of a starting point of calculations.

Hope this helps.