Fine tuning Forecast Parameters

[Former Member]

Hello,

Can someone please clarify on fine tuning forecast models:

I think the first step in forecasting is to find a suitable model for our history, once a model is identified using lets say, Auto model selection procedure or other methods, the next thing is to fine tune the selected model. For this, I believe we need to fine tune one more smoothing parameters Alpha, Gamma and Beta (whichever is applicable) depending upon the model chosen and then settle down on the combination of values for which we get the least values of error measures like MAD/ MAPE etc.

Please tell me if my understanding is correct or is there something more than just looking for lowest values of these error metrics to arrive at a best forecast model.

Tej

[Former Member]

Hi Tej,

You are absolutely correct about choosing forecasting model and fine tuning of the forecast models.

On top of it, I must add, that simply the historical data, as it is, might not be a good measure for forecasting. Many factors are not captured in raw sales data, like impact of externalities like competitors' performance and internal constraints like capacity. So at many times, the forecast with a greater error can actually be a better forecast. The broader perspective in which forecasting technique works is in making long term strategic decision like capacity expansion or more procurement from vendors.

You might want to consider these painful exercises, if the requirement is more for demand forecasting, and used as a strategic tool, rather than sales forecasting.

regards,

biplab

[Former Member]

Thanks biplabk for your clarification and also additional information.

Other experts are still welcome to provide any inputs from purely the best fitting statistical forecast generation perspective, if there any best approaches available to fine tune the statistical forecast or general procedures used in the APO world to find a best fit forecast model assuming that we already have a history that has been corrected.

Thanks

Tej

[Former Member]

Hi Tej,

Another method to get some good forecast values would be to use composite forecast technique, where in you can maintain multiple Univariate/MLR forecast methods and allow the system to pick up method with least error or you can even assign % to the models. I have used this model and works good.

Reg,Bopanna

[Former Member]

Thanks for your feedback