topic Re: 3-periods ahead forecasting with ARIMA/Transfer Model function in Discussions

3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Thu, 23 Aug 2018 09:25:11 GMT

Hi everybody,

I am trying to solve a problem where a 3-period ahead forecast is needed. So for example, if the frequency of my data is one day, what I would like to achieve is a model that when runned monday gives me bak a prediction for thursday.

I am accostumed with the ARIMA platform in JMP but I can not find an option to return the 3rd prediction instead of the 1st.

Is there a way to achieve this?

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Wed, 29 Aug 2018 10:44:44 GMT

Maybe my queestion was not clear, I would like to save the prediction (via the save prediction tab) and instead of getting the 1-period prediction I would like to get the 3rd. Is this possible? anyone?

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mark_Bailey — Wed, 29 Aug 2018 12:38:13 GMT

See Help > Books > Predictive and Specialized Modeling > Time Series chapter.

The launch dialog includes a number edit box for Forecast Periods. Enter 3 here for your case. Click OK, select the time series model, and then click the red triangle next to this model. Select Save Columns.

I illustrate these steps with the (inappropriate) Fitness data. I select Oxy for the Time Series role and enter 3 for Forecast Periods. I fit a first-order ARIMA model and select Save Columns. I receive these results:

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Wed, 29 Aug 2018 13:38:02 GMT

OK, thanks. Is there a way to save one prediction per period in the specified forecast?

So for example if I specify 3, I can save prediction for period 1, prediction for period 2 and prediction for period 3?

Many thanks in advance.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mark_Bailey — Wed, 29 Aug 2018 14:03:30 GMT

As you can see from the example that I posted above, the answer is yes.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Wed, 29 Aug 2018 14:20:12 GMT

Maybe I am missing something but in your example you have one table with one predicted value (in this case the 3rd period prediction).

What I am asking is the following. is there a way to have one column with the 1rd period prediction, one column with the 2nd period predicition and one column with the 3 period prediction for the same timestamp in one table as output of the ARIMA platform interface.

You can of course do it by running the arima 3 times with three different forecasts periods and then merge them, but is there another less cumbersome way to do it?

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mark_Bailey — Wed, 29 Aug 2018 15:25:02 GMT

I received a prediction column with the forecast for the next period (1), the period after that (2), and one more period (3). All three periods (rows 32-34) beyond the last observed time (row 31).

Three numbers as provided by JMP Time Series would require trivial effort for manual data entry into separate data columns. Otherwise, you could use the Tables > Split command to do it.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Thu, 30 Aug 2018 14:00:05 GMT

Uhm, I am not so sure we understand eachother. It is for sure a problem with my understunding of the ARIMA platform in JMP.

Let's try to pick it up step by step.

1) If I select the forecast period as 3, is the column Predicted Oxy in your example filled with 3rd period prediction for the related timestamp? i.e. row 3 contains the 3rd period prediction for the same row in Actual Oxy and so on?

2) If this is the case, is it possible to obtain a column showing a predicted value for 1st forecast period and one for the 2nd forecasting period as well for all timestamp points in the table and not only for the last one?

Sorry for the confusion and thanks for your patience.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mark_Bailey — Thu, 30 Aug 2018 14:34:52 GMT

Rows 1-31 are the actual (observed) responses used to fit the ARIMA model along with the predicted response. Rows 32-34 are the forecasts (predicted responses) for the 1st, 2nd, and 3rd periods beyond the observed data. So row 31 is last observation on Monday (in your example) and row 34 is the forecast for Thursday.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Tue, 11 Sep 2018 13:59:21 GMT

Sorry for radio silence, I had a week vaction.

Ok now I see the problem, I would like to run the ARIMA model on a validation dataset separated from the training and for all points/observation I would like to see the 1st, 2nd en 3rd period ahead prediction. Is this possible?

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mark_Bailey — Tue, 11 Sep 2018 14:25:02 GMT

You can click the red triangle in the outline for the fitted model ( e.g., ARIMA(1,1,1) ) and select Save Prediction Formula. You can use this formula, like any other, with any data you like.

Please read Help > Books > Specialized Modeling > Time Series.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Tue, 11 Sep 2018 14:36:04 GMT

No because there you get the predicted value, which is the predicted for the next and not the i-th period a-head. At least this is how I understand the saved ARIMA/Transfer Function formula works, but maybe I am missing something.

What I would like to have is a saved ARIMA/Transfer Function which can also return any i-th prediction a-head for all rows in a dataset. What I would like to compare is how worse is the third period ahead wrt the first period.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Di_Michelson — Tue, 11 Sep 2018 20:28:14 GMT

Hi Mattia,

I think you are looking for the three-step ahead forecast, is that correct? I think you can apply the Lag() function to the prediction formula and it will do what you want. There is probably an easier way, but try this. Save your prediction formula. Then right-click at the top of the prediction formula column and select New Formula Column > Row > Lag. This will give you the predictions going the wrong way. Right-click at the top of the lagged column and select Formula. In the second argument, enter –1. (You will probably have to enter a 1, then click the +/- button to make it negative one.) Then click OK.

This is the one-step ahead forecast. Note that row 2 of the prediction formula is row 1 of the lagged prediction formula. For the three-step ahead forecast, you would lag by –3.

Please let me know if any of this is not clear or if I have answered a different question than you were asking!

--Di

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Wed, 12 Sep 2018 08:12:17 GMT

Hi Di,

Thanks for your input. I understand what you are suggesting but this just shifts the value by one but you are still using one-step ahead predictions just shifted, this predictions are made with more information than what was effectively available at that timestamp in time.

What I would like to do is compare the 3rd step a-head prediction with the 1st step a-head for any observed value in my validation dataset but made with the same amount of info, i.e. at that timestamp). In other words, I would like to really see the forecast for i-th period for any oberved value(row). Or using JMP terms and quoating from Help > Books > Specialized Modeling > Time Series:

I would like to save the right part of the forecast plot for all the rows in my validation datasets. Assuming my forecast period is 3 I would then like to store the three periods in three different column next to the observed value.

Hope this makes it a bit more clear.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Di_Michelson — Wed, 12 Sep 2018 18:50:39 GMT

I don't think there is a way to get just the three-step ahead prediction in a column from the time series platform. I also checked with the developer. I think it's possible to do with a script that excludes rows, fits the model, saves the predictions, writes them to the table, then loops by unexcluding one row at a time.

Re: 3-periods ahead forecasting with ARIMA/Transfer Model function

Mattia — Thu, 13 Sep 2018 07:10:34 GMT

Hi Di,

Thanks for taking the time to check this. It sounds quite combersume to do it in JMP at the moment. I might have a more in depth chat with one of the developers during the next summit to address this.