Getting a constant forecast from ETS tool, Forecast factory tools

[Jack6627]

Hi,

I am using the ETS and Forecast factory tools and I noticed that I am getting a constant value for forecasts from both of them. I am using 3 years of historical data to predict 6 months into the future on a monthly basis.

ETS Tool:

TS Forecast tool:

Subperiod 1, 2, etc. indicates forecast for January 2021, February 2021 etc.

Following are the data values from January 2018 to December 2020 arranged chronologically.

15
9
6
12
14
4
27
4
6
13
14
12
24
33
21
41
32
35
4
22
2
14
21
0
5
8

4
6
2
1
10
2
5
45
11
0

It is very unexpected when you consider that there is quite a bit of variation in my original data. I am observing a similar behavior for forecasts of other items from the TS Forecast Factory tool.

I am using the default settings provided by the ETS tool. Could somebody please verify if this is correct and tell me why the model could be doing this?

[AngelosPachis]

Hi @Jack6627,

The number of periods in your historical data doesn't suffice to estimate the seasonality element of your model, this is why you prediction is a straight line with big confidence intervals.

Do you have more historical data that you can use to train the model?

[AngelosPachis]

@Jack6627 Also, if you look at the output of your ETS tool, you can see it says ETS (A,N,N). The final two Ns in those brackets stand for trend and seasonality and currently your model couldn't identify any of them so those are set to none (N).

If you try to change the setting of the model to include a Seasonality Element? Have you also tried using the ARIMA time series, and if yes maybe the output from that tool was more promising?

[Jack6627]

Thank you for your response @AngelosPachis . I am a little surprised that 3 years of data is not sufficient for the model to be able to forecast 6 months into the future. I created a bunch of dummy data for the year 2017 and switched the Seasonal setting from Auto to Additive and now I do get some pattern in the forecasted values by the ETS tool.

Now my concern is that, I am not just forecasting for one product, but for 300 of them and for that I will need to use the TS Model Factory and TS Forecast Factory tools. Unfortunately, for these tools (as far as I know), I can't manually toggle the seasonality to Additive like for the ETS tool. Any ideas on how to tackle such scenarios?

[AngelosPachis]

@Jack6627 my suggestion would be to create a batch macro that will process each product individually. That will allow you to either tweak the configuration settings of the ETS tool within the batch macro if you want to (and set the seasonality component to additive for example) or leave it to auto and see what's the results for each product.

[leozhang2work]

You should always do a decomposition plot before ETS, so you can tell which one to go for. As the default went for ANN (i.e. constant model). In the attached screenshot, you can see addictive for seasonality and reminder, a downward trend, therefore I tried AAdA to give the output for the second picture.

For the trend, you don't have enough data to show the increase in the middle is outlier behavior or cycle behavior, and ETS doesn't have option to really handle cycle, so prediction is not the best. You do need to care about those spikes though, I would suggest to try Box Cox transformation (or just log it, pic 3).

If you are really going for accuracy rather than explainability, then as @AngelosPachis suggested, build a macro for grid search to run through all combinations. 

[Jack6627]

Thank you for your responses @leozhang2work and @AngelosPachis. I decided to give up doing it on Alteryx and doing it within excel now. This way, with excel macros I can test multiple configurations of the ETS models and compare the MAPE values. 🙂