Predictive Scenario generating data outside of the planning time range defined

[kennedydb]

Hi colleagues,

We are running a predictive scenario with SAC, and for some reason the system is generating forecast data out of the time range defined in the settings.

We are running a 14 month forecast, based on a YYYY-MM date, for a 10+2 + 12 scenario, and for some entities it's using a forecasting gap of 45 days rather than 30 days, so although it's creating 14 months, it's generating data for two additional months, and when we try to save it back to the forecast version, we get the following error:

"The forecast can’t be saved to version "Forecast" because either data is written outside of the planning time range, or a validation rule was broken, or some of the data cells to be saved are locked."

This is an example of the correct forecast periods, as we are using monthly granularity:

And this is what it's generating for a few of the entities:

We have the following settings for the time granularity:

I'll appreciate your feedback if you have had a similar issue.

Thanks,

David

[achab]

see the similar answer here: https://blogs.sap.com/2021/02/19/sap-analytics-cloud-predictive-planning-frequently-asked-questions/

[kennedydb]

Thanks Antoine, yes it makes sense, in fact we were able to bypass the error by aggregating the data. What is confusing, is that in some cases you get an error, when there is not enough data to predict, and the entity is excluded from the forecast results, or you just get a warning, but you are still able to export the data to a forecast version. In these rare cases, it causes an error only when you try to export the data, and because of the error, not all the data is copied, so even some of the "good" data forecasted is missing in your forecast version, and you have to go back and check the entities one by one, to find the ones that are expanding over the time limit. Is there another way to prevent this from happening? Thanks, David

[achab]

Hi David, the only ways I know are:

- either to check the data beforehand and filter out the data that does not make sense to predict (when staying at very granular level) This can be done using data actions for instance, you can code a logic that removes from actuals the sparse data. it can be time-consuming to put in place though and depends on the inner complexity of the data models. Another question to tackle is to complete the predictive forecasts so that they can ultimately be compared to actuals. So typically there is a data pre-treatment but also a data post-treatment. it's data pre-treatment then predictive engine then data post-treatment.

- or to aggregate the data at higher level (middle-out forecasting / top-down forecasting).

- or any combination of the two approaches. You might want to approach some of the data at very granular data because it's not sparse (good for predictive) and tackle some of the data at aggregated level.

considering all this the most sophisticated approach you can think of would be:

1. data pre-treatment to separate the granular data from the data that needs to be aggregated

2. run the predictive engine on the granular data and in parallel on the aggregated data

3. combine everything into a single private version

4. possibly use data actions & allocations along the way

Not all cases carry the maximum complexity. Luckily we have nice improvements coming to facilitate different aspects on the end to end handling.

in 2021.Q4 QRC one is now able to filter out the entities using dimensions, hierarchies & properties which facilitates the separation between the good, the bad and the ugly entities.

somewhere in H1 2022 (planned) it will be possible to combine multi actions & predictive planning which means once the end to end logic data & predictive logic has been created & proved, it can be automated.

Sorry for the long answer but thought it was worth detailing my perspective on this. Hope this is useful, I wrote this fast so please ask if anything is unclear.

Best regards,

Antoine