Nils Kiele

In [the latest commit](https://github.com/awslabs/gluonts/pull/2267/commits/7830201bc496dfd4cefdc48980694d19e680d7c8), there now is yet another approach for the evaluation that is quite different from the one before. Thanks @jaheba for the inspiration! Most important differences are:...

Here's another slightly different take on specifying metrics. Without having implemented this, could this kind of usage be a compromise between hard-coding the `axis` argument (like it's done in my...

Changes in [latest commit](https://github.com/awslabs/gluonts/pull/2267/commits/9f9dd22dd75c763681125f58f5a75c4cadcad24e): - Differentiate between `BaseMetric`s (which have the same dimension as the input) and `AggregateMetric`s (which accept an `axis` argument to aggregate over - before, axis was...

I tried once more to make things easy and flexible and added a third approach to doing metrics in the `ev_v3` folder. 😄 [This commit](https://github.com/awslabs/gluonts/pull/2267/commits/888b890e95e1b4f44e4e613a417a4aca7c95563c) is taking a step back...

Status update: After taking a step back, it became clear that performance (in the form of caching dependencies for metrics) should **not** be a priority for the evaluation. That's because...

Separating the aggregation axis from metric parameters makes a lot of sense, I agree. While changing this as suggested by @jaheba, things got ugly for some metrics. One issue is...

> Can't we calculate the `SeasonalError` per time-step instead of on the level of time series? As far as I understand this metric, it does not make sense per time...

What's special about the seasonal error, though, is that because of |y_t| = |y_tm| < |past_data|, we get fewer values out that the context length is. So, they can't be...

That's how I understand it too, yes. The seasonal error is more of an attribute of a time series rather than the evaluation of a forecast. Maybe we could have...

One more comment about the seasonal error. It makes as much sense to have it as a `Metric` class as it does for the `MSISNumerator`. It's not so nice because...