cancel
Showing results for 
Search instead for 
Did you mean: 

Seasons per period

0 Kudos

People,

I saw question form Kelly W, about Season per Period, answer by James S.A., and I would like know that you could help me. In my case I have the next situation: I plan in weeks, but my historical data it is in days(now I have two years of historical) repeat each month, like I notice, then which numbers of periods can I use in this case?

Regards

Nestor

Accepted Solutions (1)

Accepted Solutions (1)

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,

Any domain knowledge or knowledge about your data or the industry ( knowledge about the business for which you are trying to forecast) helps in deciding on the number of periods per season.


I suggest you to consult your Demand Planning Manager who knows your historical data.

As you may know, Demand patterns can be found when data points are plotted. The patterns can give you an idea on demand. Also it will tell you about 'Periodicity' - duration or length or time period over which pattern of your demand or signals repeats over itself.

Typical data history patterns are constant(Level), Trend , Seasonal , Seasonal Trend and sporadic history(Randomness)


To understand your data, you can try to plotting data and/or summarizing it. This will help you get answer to your question.

It is better to Forecast at more aggregate level  instead in Day buckets may be in Weekly buckets or instead of Weeks may be in Monthly Buckets. Aggregation produces a more accurate forecast as there will less variation or distortion of data.

Please see if this helps.


Thank you

Satish Waghmare

0 Kudos

Satish,

Are we using Demand Planning as to predict values in your ATM(Automatic Teller Machine), in the context at CMO(Currency Monetary Optamization) SAP solution, day by day, then it is not possible for us make any aggreation as you mention it.

We know that our historical data has certain behavior, increase values at end and begin of each month, and during month we have something like a constant.

We saw int the http://scn.sap.com/message/14643900, some suggestions by James SA, that help Kelly W., and because that We decide start this discussion.

Thank you for try to help us

Nestor

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,

Yes, it is correctly being answered by James SA in the thread which you mentioned.

I will assume - You are very sure that your Historical data is Seasonal, you want to use Seasonal Model for forecasting the demand.

As per my understanding,  In your case History repeat itself every month or ( I should say) on monthly basis and you do the planning in WEEKLY buckets.

Typically we have 4 or 5 Weeks in a Month. (52 Weeks in Year, 26 Weeks in 6 Mnths, 13 Weeks in Quarter, 4 or 5 in a Month)


Your Season Per Period would be either 4 or 5. (Depends on how many weeks you have in a month).

Hope this will clarify your doubt.

If you need any more information, please do let me know or else you can mark this question as "Answered".

Thank you

Satish Waghmare

Former Member
0 Kudos

Hi,

it all depends on your data:

- if for ex. data from first week of august 2013 is closer to first week of august 2012 than to first week of july 2013, then you will use 52 periods +/-something, depending of the year (however 2 years will probably not be enough)

- if july 2013 is closer, then use 4-5 as suggested by Satish.

A more practical approach woul be to try both, compare errors (ex. MAPE) and choose whatever suits best. With some DP models you have the possibility of letting the system explore periods around a given value, give it a try if it is also available to you.

Anyway... for two years of data, my bet is on 4-5.

Regards,

J.

0 Kudos

James,

Thanks for help!

Let me see if I understood clearly how DM makes calculation: Our historical data is from November, 14, 2011, then if We use 52 periods, DM, compare this WEEK, November, 14, 2011, with WEEK November, 14, 2012, and WEEK November, 14, 2013, for bring us results to November, 14, 2014. Are We correct?

Now if I use Satish`s suggestion 4-5 periods, let me work on with 5 periods, then DM compare WEEK November, 14, 2011, with WEEK November, 21, 2011, with WEEK November, 28, 2011, with WEEK December, 05, 2012, with WEEK December, 12, 2012, and with WEEK December, 19, 2012, for bring me results to a WEEK period that We can choice as our convenience, like WEEK January, 06, 2013, takes in account the same MONTHLY basis, like Satish mention it above, ahead in my historical data. Are We correct?

If We correct in both understands, my doubt that it would be remain is about bucket, mention it by Satish too. Our bucket is in DAY that We use in general profile, SAPAPOM96B transaction, additional configuration, period code. When I did our prediction is always for next WEEK, but DM brings to me this next WEEK in DAYS, relate our daily withdraw money historical data. Then I ask if would We not have problem in our case to use 52 periods or 4-5 periods that look like indicate for WEEK and not for DAY?

Regards

Nestor 

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,


I have a suggestion, when you are unsure about Historical data, Season and not seasonal and which forecast model will yield you better forecast. You can rely on Statistical Forecasting using Composite Forecasting profile.

Develop a Composite forecast profile (/SAPAPO/MC96B)) comprising multiple Time Series(Univariate) forecasting models like - Constant, Trend, Seasonal, Seasonal Trend and Croston Models. You can even include Automatic Model Selection. All of them will use same historical data but Forecast strategy/model and forecast parameters( like for example : alpha/beta/gamma/sigma/season per period etc)  can be different. You can keep the weightage for forecasting model as 100% in Composite Forecast profile. So whichever Forecast model has least forecast error will get adopted. This should results into large reduction in Forecasting Errors and hence more accurate forecast.

In your composite profile, please do include seasonal models (strategy 30/31) and Strategy 35 (Seasonal + LR),  but please be aware that they use fixed smoothing parameters (alpha, gamma). In all the above try it with outlier (sigma of 2 or 2.5) and without outlier correction.

SAP Help Link :

Composite Forecasting - Creating a Master Forecast Profile - SAP Library

Please give a try this should solve your dilemma and let me know if able to answer your query.

Thank you

Satish Waghmare

0 Kudos

Satish,

We have a certain knowledge about our historical data.

We beleive that our doubt is how DM calculation about season per period is working, because this I asked James about that, after him answer.

Thank for your suggestion about it, but we are using Automatic Model Selection, and strategy 50, that are indicating for us: season test is positive in our DM test. Then We beleive that if We would understand this calculation, can We use "tip" that You and James offered for us about season per period (4-5) for reach our goal.

Regards

Nestor 

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,

Yes, you can try setting up season per period (4-5) and do a trial run by generating statistical forecast.  Analyze the Stat forecast to see if it is satisfactory and then decide on the next action.

Note, changing the value in 'Season Per Period' parameter is not a big deal, it can done pretty quickly.  You can always try various values/option before you finalize one which suits the best.

Hope this will help.

Thank you

Satish Waghmare

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,

As a good practice (According to The Rules of Engagement defined by SCN), If your problem is resolved then please mark question as "Answered". or Else let us know if you need any additional information, will try to help.

Thank you

Satish Waghmare


0 Kudos

James,

Could answer to me question that I sent to you about this stuff in jan, 04, 9:09PM?

Thank you

Regards

Nestor

0 Kudos

Satish,

We tried to use 4 or 5 periods and results are not so good.

We are waiting for James`answer to us about our question about how DM makes calculation! We believe that we would understand this calculation, We can use DM parameter properly.

Than we can not assign answered for our question untill there, We suppose.

Regards

Nestor

Former Member
0 Kudos

Hi José,

about the calculation, you can find it in the SAP DP online documentation (look for Holt-Winters or Linear Seasonal Model, 45/31) or even Wikipedia (for the first).

http://help.sap.com/saphelp_SCM700_ehp02/helpdata/en/4e/f9029644ee1b01e10000000a421bc1/content.htm?f...

Unfortunately for you, those models will probably not work well for your data, they are better suited for monthly data and no calendar/week effects. Meaningful forecasting of daily data could be done with some development effort, but for this you need somebody how knows a bit of forecasting, just ABAP won't do.

Good luck,

J.

satish_waghmare3
Active Contributor
0 Kudos

Hello Nestor,

Since you are using Forecast strategy 50(Auto Model Selection-1),  Below link will provide you some very good insight.

Automatic Model Selection Procedure 1 - Creating a Master Forecast Profile - SAP Library

Hope this will help.

Thank you

Satish Waghmare

0 Kudos

James,

Thank you for your reply!

We already know documentation from link that you mention it!

We are beginnng understand of our huge problem, because historical in days, and forecast in days too, and it is not look like a best practice for season, that more adequate for monthly basis when we have season like summer, winter, or more else. Unhappy we do not selling ice cream, instead we need to put Money day-by-day in our ATM(Automatic Teller Machine).

I think that you use best words for our problem: Good Luck!

Thank you

Regards

Nestor

0 Kudos

Satish,

We already know this link!

Thank for try to help

Regards

Nestor

0 Kudos

James,

After long time We see light in the end of tunnel! We discovered how SCM made calculation to compare period in day by day basis. For example if You have three historical years in Your database, SCM compare the same day of first year with second year and third year, and this is the "problem". For example day 25 august in 2014, it was monday; 25 august in 2015, it was tuesday; and 25 august in 2016, it is thrusday; then SCM compare different kind of ATM´s withdrawal behavior; what would be more complicate in finished week days, that has different ATM´s withdrawal behavior when compare with normal day of the week. Because this, We could not achieve Our goal to reflect any kind of statistical model, one univariable profile, just because it was compare diffente kind of behavior.

We had certain about Our discovery, When received a visit from CMO (Current Monetary Optimization) consultant solution, that It was sold for us by SAP, that has SCM in central point, and He confirmed that We was right in Your descovery. He offered for Us another way, that present in transacion: sapapo/mc96b: MLR profile, based in statistical model MLR (Multiple Linear Regression) that use a combination of depent variable (Our historical) and indepent variables, like things that reflect behaviors change in ATM withdrawal like pay day, before pay day, or after pay day. Now Your are adjusting profiles that We created for a Our ATM park, and We are thinking that You We would success in Our goal to predict ATM´s withdrawal.

I Would like to say thank You for You and the other that tryng to help me in my stuff, and I wait that this solution that We are adpting now could be useful for the others.

Regards

Answers (0)