Too much of Forecasting Error due to input data

[nareshab]

Hi,

Can someone suggest me on how to improve the performance of the Forecasting model which is giving me MAPE error around 50+. and the distribution is attached.

What are all the best ways to improve the forecasting on a daily level.

Thanks.

[SnurreJensen]

By looking at the attached graph my first guess is that you have some periods where the absolute percentage (APE) error for the individual observations are extremely large due to very low (but not zero!) actual values. It's somewhat hard to tell from looking at the graph but what is the forecast and the actual value for 01jan2014 for instance? Try calculating the APE for that period for instance.

To counter this, either enable automatic outlier detection (I assume you use SAS Forecast Studio) or create your own model where you include events for those periods where the actuals are extremely low.

Thanks,

Snurre

[udo_sas]

Hello -

How well is your model doing on out-of-sample data?

Thanks,

Udo

[NareshAbburi]

Thanks for your reply.

I'm actually building model on whole 2 years data. And now I've got even more bigger problem with the sporadic nature of the data. please refer to the picture attached.

Please let me know what is the best way to handle this data and reduce the forecasting error.

The main problem with model is that it is not able to fit all the data points based on OLS.

Also, I'd like to know if there is any option in SAS to provide multiple seasonal cycles as input.

Any help would be appreciated.

Thanks.

Naresh

[SnurreJensen]

Hi,

Please bear in mind that when building a good forecast model the aim is not to find a model that fit all the existing data points. The aim is to find a model that is able to provide as accurate forecasts as possible given the exisitng conditions. For instance, if data is very noisy/erratic/stochastic etc. forecast accuracy will be lower than if the has nice stable patterns.

Having said that you can work with multiple seasonal cycles in FS by building your own model. For instance an ARIMA model with differences of (7,28) can be used to work with data where there are seasonal patterns each 7 and 28 days (assuming you have obs for all 7 days a week).

Thanks,

Snurre