cancel
Showing results for 
Search instead for 
Did you mean: 

Statistical model decision

Former Member
0 Kudos

Hai,

How do you decide which statistical model to go with for a demand planning scenario? Are there any general pratices for different industries with rspect to models? I mean chemical industries generally use constant model, autotive use sporadic models(80) etc....

Thanks.

Accepted Solutions (1)

Accepted Solutions (1)

Former Member
0 Kudos

Hi visu

I dont think there will be a single statistical model you decide to use even in one organization let alone an industry. I would say it would depend on a lot of things - just a few of which are listed below

what kind of product - mature, new....

what level the forecast is done at - SKU, aggregated at packsize, state/region

what kind of history exists - accurate, trimmed for promotions and

external factors - other indicators

seasonality- time range to be considered

Typically forecasting models are a business analysts domain and is something that is arrived at after a lot of trial and error and still might depend on a lot of things beyond a planners control

You can generalise a few things out of the exhaustive list in order to narrow down the search for the models from whole range available to use

for eg Spare parts usually have sporadic demand patterns

Automobiles might be one of the trend models

but within the industry and within an organization you might have to give special attention to each set of forecastable products

In some cases even no forecast and a manual judgement of demand might be a good way to go

One way of seeing how good your model might be is to use ex post forecast

A good forecast accuracy tracking system might help in getting the right model.

Former Member
0 Kudos

Hai Harish,

Thanks a lot. If you don't mind, can you explain me the business process flow in working with forecasting model?

Can I say this:

You create a couple of forecast profiles and give the user. He runs the interactive planning with both of them and change the forecast models and rerun(if it's trail and error only one model is enough I guess). Comparing the final MAD or MAPE, the user would attain a conclusion as to which model to be used. So is this run initially and then assigned to it forever? or as the situation may change, is the model determined for every time(once in a month) a interactive run with forecasting is done?

Thank you.

Former Member
0 Kudos

edefining any existing forecasting models would depend on the scope of your implementation

A sample Prcoess flow -

1. Define the products and the model based on questions like: -

--> define the procducts - What level or characteristics, single or Group or

-->Check for history at these levels - quantity(periods) as well as quality of the forecasts are important

-->check - does the present situation justify using the history good or do you manually change something

-->If there is no history - can you use something else that has history and is close to the product?

-->check -Do you know if you can pick a forecasting model or do you want the system to pick it up for you automatically

-->check - Do you want to correct history for Outliers, promotions

--> does the product need composite forecasting

--> phase in / phase out needed??

Based on the above you should be able to bundle a lot of products into grooups or products and create selection variants for them

2. test the model for a product group you have decided by checking out a sample in interactive planning

- you can manually record the forecasts for various you choose models and use "forecast comparison" to

check the error

- you can run the automatic forecast and check the model suggested by the system and the error due to it

You can also use diagnostic groups

3. Create a profile(univariate, MLR or composite) and a master profile based on the model

4. Maintain the assignment of forecast profiles to selection IDs (selection variants) by choosing Goto -->Assignment in the master forecast profile. For example, you might have a selection ID for each product family you forecast and different forecast profiles for different product families

5. Decide what to do with outliers and Create alert profiles etc based on need

After you set this up and use it for some time set a time table to revisit the groups in detail and more elaborately.

Use some kind of ABC analysis to decide on the priorities - products with most error or products with most volume etc. you might also want ot regroup the products based on how they perform.

0 Kudos

Visu,

Take a look at this document in https://www.sdn.sap.com/irj/sdn/developerareas/bpx-scm?rid=/webcontent/uuid/240738f8-0a01-0010-4889-bd254c570cd4">scm-bpx [original link is broken] [original link is broken] under Demand Planning which talks about Data Analysis and Statistical Model Selection. I am sure you will find this document useful to answer your question.

Thanks,

Somnath

Former Member
0 Kudos

To learn about forecasting, here is a good book you can read "Forecasting: Methods and Applications" by Makridakis, Wheelwright & Hyndman.

It's available from <a href="http://www.amazon.com/Forecasting-Applications-Spyros-G-Makridakis/dp/0471532339/ref=pd_bbs_sr_1/104-9869827-4274318?ie=UTF8&s=books&qid=1174301538&sr=1-1">Amazon</a>

Answers (0)