negative forecasts

[peterdudfield]

Detailed Description

Would be good to make sure we cant return negative forecasts

Context

• we used to have this, but it got removed, as it was causing the API to error (correct thing to do)
• This makes sure we dont return negative values, of course -0.0001 is fine, but if its bigger, its really obvious and should not be shown to the user

Possible Implementation

• clip forcasts in forecaster to 0.0
• in api, loop over forecasts and adjust to >=0, or perhaps we can do this in sql code?

[simlmx]

I feel the model should be responsible to not return negative values. We could also filter or warn in other places but I'd be afraid we don't notice the eventual problems with models. To me this falls under bad predictions which we should be able to detect anyway. In other words for now I would simply fix it in the model and that's it.

[peterdudfield]

I perhaps see it more in a risk way. We dont want the user to see negative forecasts, how do we do this.

• I agree the model should at first be responsible for this.
• What about if there is bug, and the model misses something, should we have another mitigation against it? This could be in the datamodel
• We could also something in the read function SUM(GREATEST(ordered_item.amount, 0)) as purchases or something like that

[simlmx]

• What about if there is bug, and the model misses something, should we have another mitigation against it? This could be in the datamodel

In any case we need some sort of validation step before we put a new model in production. This is when those could be found.

• We could also something in the read function SUM(GREATEST(ordered_item.amount, 0)) as purchases or something like that

This could slow things down but especially this could hide a bug.

The question is which is worse: having the user see a negative value and tell us that there is an issue, or returning a negative value that gets silently changed into 0 and we don't know that there is an issue.

---

We can always put the check back in the pydantic model, and, after we have fixed the model. If a model were to return 0 ever again, it will crash there and be found in the validation step I describe anyway.

[peterdudfield]

Ok, is it it fixed in the model?

[simlmx]

Not yet, it still returns -0.001 kind of values once in a while.

[peterdudfield]

ok, probably worth clipping somewhere

[peterdudfield]

@simlmx did you add this to the forecast?

[simlmx]

Not yet. Seemed low priority as the negative values are pretty much invisible in the UI (since they are so close to 0).