https://github.com/nixtla logo
k

Kaustav Chaudhury

04/04/2023, 3:49 PM
Could any one suggest loss and scaler type for TFT for negative values in data
k

Kin Gtz. Olivares

04/04/2023, 3:50 PM
Most point prediction losses can work: MAE, MSE, sMAPE And other distribution prediction losses: MQL, DistributionLoss-Gaussian, Gaussian Mixture
k

Kaustav Chaudhury

04/04/2023, 3:51 PM
Also we have multiple features as input
k

Kin Gtz. Olivares

04/04/2023, 3:52 PM
You can feed your features into the model through hist_exog, futr_exog and stat_exog lists
k

Kaustav Chaudhury

04/04/2023, 3:53 PM
Yes but the model doesn't improve on Validation loss
k

Kin Gtz. Olivares

04/04/2023, 4:06 PM
Not all features need to have predictive power Some can also have redundant information that makes learning hard NeuralForecast does not have automated feature selection capabilities yet
2 Views