Time Series Analysis - SAS Advanced Analytics Course - ARIMA model parameters using holdout sample

[eddieray01]

In Text Analytics, Time Series, Experimentation and Optimization module.  Forecasting a Holdout Sample Using the ARIMA Model Demo STSM02d05a.  The instructor points out that even if we specify back=6 periods (let's say series has 42 periods) for example, that proc ARIMA uses full sample of 42 periods for estimating parameters for the model versus just 36 (42-6 holdout).  Does this same logic  on using the full sample to estimate model parameters also apply to the proc esm and proc ucm procedures?  Just wanted to clarify or if it is just nuance with proc arima?  Thanks

[Cynthia_sas · Accepted solution]

Hi:

  One of the class instructors said:

"This is just a nuance of ARIMA. The ESM and UCM procedures do not react this way. That is why we also included, within the course, SAS code that separated and prepared the data for ARIMA so that it would do as we wished for honest assessment."

Hope this helps clarify the difference,

Cynthia

[Cynthia_sas]

Another instructor said that: "The BACK option is on the FORECAST statement in the ARIMA procedure, and will not impact the number of observations used to estimate parameters. This is not the way the similar options are implemented in ESM or UCM, and is unique to the ARIMA procedure."

Cynthia

[eddieray01]

Sorry another question for instructor:   I am confused on.  In Chapter 3 Exponential Smoothing Models 3-10 in the module notes  it says "Seasonal smoothing should be used when time series has no trend but has seasonality"  and "Winters additive or multiplicative ESM should be used when data has trend and seasonality".    However on 3-27 (3.2 Chapter Summary) it show on the flow chart in a case of No=Trend, Yes=Seasonality and says if seasonality stable = Yes then use "Additive Seasonal OR Winters"  if Seasonality not stable used "Multiplicative Seasonal or Winters"   My guess if no trend in series, then you should not use Winters.   Any clarification on this .  thanks.

[Cynthia_sas]

Hi,

  Here's feedback from 2 instructors

=== 1 ===

If no trend is present but seasonality is present, it is capable to use either a Seasonal smoothing model or Winters. You can think of Winters as the more general form of the Seasonal smoothing model. Winters will just use more degrees of freedom to try to estimate a trend when one may not truly exist.

========

=== 2 ===

There is an old saying: "Nothing succeeds like success."

 
If you do not have trend, then simple seasonal exponential smoothing might be appropriate. If instead of simple seasonal exponential smoothing you use additive winters or multiplicative winters, your forecasts will not necessarily be bad, but you will be estimating parameters that are effectively zero, and there are various statistical reasons for why you do not want to over-parameterize a model. Furthermore, there is no multiplicative option for simple seasonal exponential smoothing. Multiplicative models allow you to talk about percent change rather than absolute change, which can be appealing. Therefore, you might choose multiplicative winters over simple seasonal exponential smoothing even if no trend is present. If this works, and you get good forecasts, then "nothing succeeds like success."

========

  Hope this helps clarify your question.

Cynthia

[eddieray01]

Thanks.  That makes sense.  I appreciate the clarity.