Issue with forecast range predictions

[anie_p]

Could you help us with an issue regarding forecast range predictions when dealing with multi-series projects? And more generally, is it expected for forecast range predictions to produce forecasts for future (unknown) periods?

On multiple occasions (different data and tasks) we have noticed a limited functionality in the results of forecast range predictions.

With regards to test data setup – we have followed the standard requirements, including:

• sufficient history – to accommodate the FDW span
• forecast rows – with no values in unknown periods (excluding FKA-s)
• features known in advance – populated with respective values for future periods

We observe normal functionality when running forecast point predictions – e.g., when using the latest historical forecast point, predictions are successfully generated for the “n” unknown future periods (FW).

When we attempt to produce forecast range predictions our expectation is to see sliding window predictions with a constant length, regardless of forecast point (including future ones as per a preselected date range). However, the results we do observe do not include any predictions for future (unknown) periods. The latest prediction we get has a “Timestamp” matching the latest available historical date.

Would you please confirm if this is indeed the expected functionality – for forecast range predictions? If not, can you elaborate on what may be causing the issue?

Illustrative materials:

1. Test data structure

2. Forecast range predictions

[Abdul.J]

The last line in the UI explains this behaviour. 

[Abdul.J]

Hi, 

This clarifies the prediction setup. The behaviour you're seeing is by design. Prediction Range is used mostly for validation purposes as the expectation is that you would have values available for Feature Derivation and then validation with actuals. For production purposes you will use Forecast Point prediction as you are looking for future predictions. Hope this helps.

[anie_p]

Hi, 

With regards to the target variable in the prediction data of the example:

• From 2022-02-27 to 2022-06-27: target values are included (as historical records are available)
• From 2022-06-27 to 2022-10-24: target values are represented as missing values / NA (future period part of the prediction range)

Here is an example of the prediction dataset structure:

[Abdul.J]

Hi,

Thanks for the additional information. In the example you mentioned of prediction range from 2022-02-27 and 2022-10-24, do you have nulls in target column after 2022-02-27 when you upload the prediction dataset? 

[anie_p]

Here is some additional information for one of our time series projects:

• Historical rows: -26 to 0 weeks
• Forecast rows: +1 to +17 weeks
• Number of features known in advance: 2

Forecast settings for forecast range predictions:

As illustrated on the screenshot above part of the prediction range (depicted in yellow) includes future dates - with unknow predictor values (apart from those for features known in advance.) Is it perhaps such intersection that you are referring to in your question?

To reiterate the issue - despite having a prediction range, which in this example - starts on 2022-02-27 and ends on 2022-10-24,  predictions generated from the "Forecast Range Predictions" setup are not available beyond 2022-06-27 (the latest historical date in this instance).

We were wondering if this is the extent of the intended functionality for "Forecast Range Predictions" or if we are encountering some technical issue.

[Abdul.J]

Hi Anie, quick sanity check question, Do the dates in Forecast Range intersect with any of the dates in the Forecast Window? This might limit predictions because the model gets Nulls in the feature derivation window.