Forecasting

[veekay]

Hi All

I am trying to figure out the sales forecast for the next 26 weeks /6 months from the given data. I have tried both ARIMA and ETS but both at 80% sample give straight line as a forecast. The TS plot does show some seasonality in the data. My questions are:

1) Is my choice of forecasting method for this data incorrect ?

2) If I change the estimate sample percentage to 99% and validation sample percentage to 1% , ARIMA states showing some dynamic forecast trends. But I dot think this is correct way. I would like to keep it at 70-30%

3) in the end the TSCompare tool (interactive browse output) , why doesn't it show the future 26 week forecast post Dec 2016 whereas it shows it in individual interactive output for ETS and ARIMA ?

Thanks in advance

VK

[pedrodrfaria]

Hi @veekay 

Could you please redo the Alteryx Package to include the data input? Or at least a sample dataset? This way we will be able to run the workflow and see what you are seeing on your end.

[veekay]

Sure @pedrodrfaria , apologies, I ddnt realise the package wasn't working. I have shared the workflow as well as the input file now. 

[AngelosPachis]

Hi @veekay ,

Concerning your third question, the TS compare tool can be used to compare the time series and decide which one suits for your instance based on different errors (shown in the R output anchor).

So the model that has errors closest to zero is the better option, in this instance, that would be ARIMA.

To create the forecast for 26 weeks, you should use a TS Forecast tool and connect it to the O output anchor of your ARIMA model. That will return you the different data points for the 26 weeks in the future, with the confidence interval bands

[AngelosPachis · Accepted solution]

@veekay  For questions 1 and 2, I don't consider that you have used the wrong forecasting method (after all there are only two that you can use).

Your ETS forecast is a straight line because Alteryx has identified that your model has multiplicative error terms, additive trend terms (hence that's why your forecast is pointing upwards) and no seasonal terms which explains why you are seeing a straight line. Alteryx has not identified any seasonality in your dataset. That's what ETS(M,A,N) stands for.

The same goes for the ARIMA tool, where in the second set of brackets that describe the models seasonality, everything is set to 0, hence likewise with the ETS tool, no seasonality can be identified.

Probably what happened when you increased the estimation sample % was that you fed more data to train your model and that made it easier to find a pattern in seasonality. But as you pointed out, it would be a poor approach to set the Validation sample % to 1% just to get the model to work.

Hope that helps,

Regards

Angelos

[veekay]

@AngelosPachis : thank you for such a clear explanation.