This message was deleted.

Thanks for the tip @Manuel! Indeed there are several strategies in the HPO literature like the one you mention (succesive halving, hyperband, etc). I wonder if your current configuration of the TFT might be stuck in a suboptimal region given that the random seed has a large effect on the performance (for example too large lr)

@Cristian (Nixtla) Yes, it is exactly inspired by Hyperband but focuses on random seed. Unfortunately, even when decreasing the learning rate, the behavior remains similar. In particular, with some seeds, I cannot predict certain seasonal peaks in the time series. Instead, I obtain flatter predictions, even when the number of heads or the size of the hidden layers increases.

Very interesting, thanks for sharing 🙂. This is why it is also very common to create ensembles based on the forecasts produced by multiple random seeds. In the NBEATS paper they ensamble more than 100 models 😅

It makes sense, although in my case the model with the best seed would risk being made worse by an ensemble with models with bad seeds