A new paper shows Kolmogorov Arnold network outperforms transformers https://arxiv.org/abs/2408.11306

I know nothing about KAN. Does it work multiple time series or is this univariate?

@Vidar Ingason, neuralforecast has an implementation of the KAN model for time series, which also supports exogenous features. Feel free to try it out!

Awesome! Does it work for multiple time series or is it univariate?

It's univariate.

Does it actually support exogenous features? It isn't listed in the docs here that it can/is the doc out of date? https://nixtlaverse.nixtla.io/neuralforecast/docs/capabilities/overview.html

By univariate, we mean that it won't consider the inter-dependency of other time series (other unique_id). You're right, it does consider exogenous features, but if your dataset has more than one unique_id, it will consider each series separately. This is unlike multivariate models, like iTransformer or TSMixer that actually consider all series and their inter-dependency. Indeed, there's a mistake in the docs, the KAN model actually support exogenous features! Thanks for pointing it out, we're gonna fix it!