Best method of forecasting

[Former Member]

I have configured my system the option of selecting 56 different methods of calculation of the forecast, but in some cases this provision does not fit reality. Articles especially true trend changing at any given time. Items that sold weekly 1 or 2 units go to sell 500 units, when I took several periods selling 500 units per week. The method that gets or is prone thereby anticipating the hit 1000 units per week because he believes it will continue to increase. Or does a very low forecast because it considers the previous periods. In this case, if only lame periods of change from the forecast will be correct, but that would force me to review items daily. It was a way to parameterize the system to have an accurate forecast of these items. Thank You

[Former Member · Accepted solution]

Hi Raul

You need a forecasting lesson!

I understand that you are using model 56 for forecast run and not happy with what it pics.

While model 56 is a good way to guide you in your "initial" run, it is not always the best model to use for following reasons:

1. By nature of the model, system runs through every model and select the best fit. This kills performance , if you run in background and regularly.

2. Not all products have the same lifecycle stage and so it is practically incorrect to use same model.

I would use the model 56 to see what the system suggests, start with that recommendation, then understand where you product is in the product life cycle, launch, growth, maturity or decline. Maintain a lifecycle profile for the products so that the system considers these in forecast run.

Some times it is quite obvious what method has to be used, if you know a growing product or declining or a cyclical ( where you can clearly see the cycle periods), then use those directly.

One other good model to use is moving average model but be careful to specify your moving average periods.

If there are too many gaps, you need to use something like a croston method.

If you have a new product with no history but can correlate to an existing product, you may use a like profile

If you have a new product that does not correlate to any existing product with history, that is the most complicated to work with!  APO can help less in that.

Also make use of the error comparison in forecasting models, usually MAPE or WMAPE is regarded a good metric. Lowest error is the better model.

[Former Member · Accepted solution]

Thanks for the reply. However, a question arises me. I am applying the method 56 to each item individually, which generates a different model per item and choose the one minor error has. Whereupon are supposed to choose the correct model on each item. True, there is a variable that is the cycle of life that makes the behavior is different, but items that are in their life cycle maturity, this method should give the forecast with minor error, or you may get more models adjusted using other parameters. That would be the most important parameters when parameterized expectations. Thank you very much and greetings

[Former Member · Accepted solution]

Hi Raul

Unfortunately statistical forecast is just a projection of past.So the results in maturity decline may not show the right numbers, that is why we have lifecycle settings to tell the system.