About the training loss function

[Xiao-congxi]

During the pre-training, is the used loss function a combination of the MSE loss calculated on the first [mean] output head and the quantile loss calculated on other 9 [quantile] output heads?

[rajatsen91]

Yes that is what we have been using.

[Xiao-congxi]

Thank you very much for your immediate reply. I have another question about the experiment results in the paper. Why are the MAE of the fine-tuned TimesFM (Table 2) is higher than the zero-shot results (Table 5) , on ETTm1 and ETTm2?

[ani0135]

@Xiao-congxi Can you please provide the loss function used here which is the combination of both heads?