Hello. I just stumbled upon Nixtla and think it is the coolest thing ever! I do have a quick question: Do any of the Nixtla models allow you to customize the loss function? In particular, I am thinking about time series predictions where there is an unbalanced costs, ie it costs much more to overshoot the true times series value than to undershoot it. For example, in the famous data-driven newsvendor problem it costs c1 to overstock newspapers but costs c2 to understock newspapers and generally c1 >> c2. How do the Nixtla models handle such problems? Has their performance been tested on time series predictions with 'imbalanced' loss functions?

Hi @Eitan Turok, In the NeuralForecast library we are prioritizing modularity. During instantiation our models allow you to customize the loss functions that you can choose from a catalog, or in its absence you can use your own. In particular for the under and overshooting loss (inbalanced losses), we have

QuantileLos

and

MQLoss

that are based on the NewsVendor problem. Here is the NeuralForecast losses catalog: https://nixtla.github.io/neuralforecast/losses.pytorch.html Here is an MQLoss example: https://nixtla.github.io/neuralforecast/examples/uncertaintyintervals.html