forecasting models

[Former Member]

we are thinking of using the auto model procedure 1 as the approach that we would be following as the initial forecasting strategy for the selection ids when we go live. auto model procedure 2 is more accurate then the procedure 1. would it be a good idea to use the 56 strategy rather then the (50 or 53) strategy. If some one can explain the pros and cons of the two in go live.

Thanks in advance

[Former Member]

Hi

Auto model 1 takes the standard values for alpha, beta, gamma etc and only uses those when it steps through each model type for it's test of best fit.

Auto model 2 uses steps of 0.1 for each of the alpha, beta and gamma values for each model.

Therefore auto model 2 should in theory find a better model fit as it varies the factors as well as the model, the downside is obviously performance as it takes a lot longer to run this forecasting model.

It depends on the number of CVC's you intend to forecast, the buckets you use and the forecast horizon (as well as the size of the box you are running on).

Hope that helps.

Ian

[Former Member]

There are approximately 1 million CVC and the time buckets are 40. I understand the theoretical difference between the two, I wanted to have your opinion on the effect on the performance , which the forecast strategy (56) compared to (50 or 53) might have in Production.

I have some cases where I have seen MAPE for a particular forecast to be less in the case of the auto model 1 then in auto model 2. Is this an exception?

Thanks in Advance

[Former Member]

Here are some more generic details on AutoModel Selction1 and 2

<b>Automodel Selection 2</b>Used when no knowledge of pattern in historical data - Highly detailed

tests carried out.

Tests for Constant, Trend, Seasonal & Seasonal Trend - More precise.

Uses all possible combinations of Alpha, Beta & Gamma smoothing factors. - Longer time to run.

Chooses model based on lowest MAD – Not recommended for use in Mass Processing.

<b>Auto Model Selection 1</b>

Used when there is no knowledge of patterns in historical data - Quicker than

Auto Model Selection 2

Tests for Constant, Trend, Seasonal & Seasonal Trend - Not as precise

as Auto Model Selection 2

If no pattern detected, system uses Constant model -Shorter time to

run in comparison with Auto Model Selection 2

Auto Selection Model 1 is used when data pattern is not obvious.

Automodel Selections are never precise as a Model Selections arrived at based on data analysis. The best way to select statistical models is by trial and error

A fair deal of effort goes in before arriving at a good and accurate forecasting model.

A good way of checking the accuracy of the forecasts is through Ex-post forecasts and setting up alerts based on deviations.

A few sample comparisons of the various forecasts can also be run in the forecasting workbench using Model Comparison and checking the errors in each model

[Former Member]

Thanks for the detailed analysis, I would appreciate if I could have a comment on the comparative effect of the auto model 1 and 2 on the performance in production when we have 1 million CVC and 40 time buckets.

Please let me know ur perspective on the following :

"Start with the Auto model selection 2 as the strategy in the start of the go live and then let the user do the trial and error for understanding the historical patterns"

Thanks in advance

[Former Member]

To comment on the statement: If you are looking at performance at Go-Live you might be better off with Auto 1. Auto 2 takes more time

some more ideas if you want to explore them

your final solution will depend on the weightage of the performance of the forecast run vs the accuracy of it

If you are focussing on the performance use Auto 1 during go live. It will definitely be faster.

If you are looking at accuracy go for Auto 2

You also dont have to use a one-size- fits-all solution for the million CVCs.

Also your forecasting level will make a difference. You might notice dont have to forecast at the most detailed level or for all the million CVCs all the time. Check if all the characteristic levels are pertinent to your forecasting.

You can also forecast at a aggregate level depending on the product characteristics. epending on the business you should look whether a top-down forecasting approach can be adopted vs a bottom-up approach that you are looking at. You can use KF disaggregation to let the forecast take its proportions

You need to be able to group your products at least into 2. You might also use the other characteristics to group these

examples of Group 1:

Critical Products were forecast accuracy matters

Products with Trend, Seasonality

Products which will have an immediate visible effect post go-live

High value products

Regions where not going out-of-stock might be important

use Auto 2 on Group 1

Group 2:

Rest of the products

Use Auto 1

[Former Member]

Thanks a lot for the prompt reply,

I think it would be a good idea to club together the accuracy intensive products.

[Former Member]

If you havent read this paper on <a href="https://www.sdn.sap.comhttp://www.sdn.sap.comhttp://www.sdn.sap.com/irj/servlet/prt/portal/prtroot/docs/library/uuid/2bb941f8-0a01-0010-88a6-ee870ea61a13">Forecasting Model Selection</a> yet... you might want to take a look at it... its in the BPX community

[Former Member]

Hi,

Please check out the following link:

https://www.sdn.sap.com/irj/servlet/prt/portal/prtroot/docs/library/uuid/7a4025f8-0a01-0010-0fb2-b7a...

Hope this should help you.

Best Regards,

Ganesh