Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

is there a way to "disable" the autoregressive nature of TFT? i dont want historical y values to be an input to the model at inference time

Hello! I don't think that's possible. Also, I'm a bit confused by the "historical y values". Do you mean that you only want to use exogenous features to make predictions?

i was able to alter the source code and actually achieve this but the results were poor.  i guess autoregressive behavior of `y` is a strong influence on the model

You could also try models that support the `exclude_insample_y` parameter, like NHITS or the MLP