Skip to Content

Intermittent demand prediction - Croston method, alpha value

Hi PA team,

I am trying to predict out the next purchase date / quantity from a vendor and want to build an automatic model around it.

I have reviewed that we can use CROSTON's method to predict the same, however as part of its SAP online documentation, the default value of alpha = 0.1, with no option available for automatic alpha calculation.

Hence I want to confirm two things:

1. Whether the CROSTON's method in HANA uses SES( Single exponential smoothing ) as done in R, so that I can go ahead and call "Forecast smoothing" with model type = SESM to compute alpha prior to calling CROSTONa's method ? I believe it is SES only because there are no options to pass BETA and GAMMA to CROSTON.

2. If it is SES, is it preferable to call "Forecast smoothing" on the exactly same data set as the one to be passed to CROSTON's method( with zero values ) or should I pass to "Forecast smoothing" method only nonzero values because CROSTON method itself applies "Exponential smoothing" both for non-zero demand and zero periods separately ? Please suggest?

Thanks,

Hasan

Add comment
10|10000 characters needed characters exceeded

  • Get RSS Feed

3 Answers

  • Best Answer
    avatar image
    Former Member
    Jul 07, 2017 at 07:15 AM

    Hi Hasan,

    Sorry for the late response. I think it's reasonable to choose the one with the lowest MAPE. However, to automatically choose optimal alpha is not in the scope of PAL.

    Best Regards

    Jerome

    Add comment
    10|10000 characters needed characters exceeded

  • avatar image
    Former Member
    Mar 09, 2017 at 07:10 AM

    Hi Hasan,

    The Croston's method doesn't use SES. Therefore we cannot say that the alpha calculated by Forecast Smoothing is the optimal one for Croston's method.

    Best Regards

    Jerome

    Add comment
    10|10000 characters needed characters exceeded

  • Mar 10, 2017 at 11:53 AM

    Hi @Jerome,

    Thanks a lot for your response !

    Now I am more confused, if it is not SES then I believe Croston may not be applicable for an automated model. Currently I am looping over different values of alpha incremented iteratively and choosing the one with lowest MAPE.

    However, how should one calculate an optimal value of alpha for Croston ?

    Hasan

    Add comment
    10|10000 characters needed characters exceeded