Holt Winter - how to use damped trends , boxcox tranfration and bias adjustments

anildas

question 1.

I am exploring Holt winter forecast model in one of our use case, would like to use the hyper parameters for

1) damped trends , 2)boxcox transformation and 3)bias adjustments

how to set these parameter in Rapidminer process?.

question 2.

Also I have used a python script to optimize the parameters for Holt Winters, and the result of optimized parameters are

Trend Damped    Seasonal      Seasonal Period            Box Cox     Remove bias Additive TRUE      Additive 12         FALSE TRUE

 How these parameters are equivalently mapped in Rapidminer Holt winter process with parameters alpha ,beta , gamma and length of period & seasonal model

Kindly advise.

anildas

Hi,: 

For easy readability , I have re-framed the second question as given below : 

Question 2.

I have used a python script to optimize the parameters(automate) for Holt Winters, and the result of optimized parameters gave me more accurate prediction (Python built model) than RM Holt-Winter model.

So please help me to map the equivalent hyper parameters to achieve the same prediction result in RM.

Python:

from statsmodels.tsa.holtwinters import ExponentialSmoothing

 model = ExponentialSmoothing(df, trend=Additive, damped=True, seasonal=Additive, seasonal_periods=12)

 model_fit = model.fit(optimized=True, use_boxcox=FALSE, remove_bias=TRUE)

RM

How these parameters are equivalently mapped in Rapidminer Holt winter process & forecasting 

(alpha, beta,gamma,length of period, seasonal model)

David_A

Hi @anildas,

the following parameters are equivalent between python and RapidMiner

alpha = smoothing_level (python, model.fit)

beta = smoothing_slope (python, model.fit)

gamma = smoothing_seasonal (python, model.fit)

The Box-Cox transformation is not yet implemented in RapidMiner, but might come as an additional operator in some future release (there is already a backlog ticket for).

Remove bias is currently not implemented, but could be an improvement as well.

The main difference is, that the RapidMiner implementation does not have a build-in parameter optimization (but we have operators and building blocks for that.

You can also take a look at the new extension from @mschmitz that is an optimized auto-time series forecast, that also includes Holt-Winters models:

https://community.rapidminer.com/discussion/57657/new-extension-how-to-do-time-series-forecasting-with-just-1-operator#latest

Best,
David