The training PackedNLLoss is negative.

[gwyanCN]

Hello, I have recently been fine-tuning my own dataset using Moirai. However, I noticed that my training loss has become negative, and the prediction loss first decreases and then increases. It seems that the model training cannot converge. Therefore, I would like to ask why my training loss has become negative and how to deal with it?

Here is my train/val loss:

[zqiao11]

Hi, the training loss is the negative log-likelihood, which can naturally take negative values. So there’s no need to worry.

Your training converges, but the issue might be overfitting. You could consider tuning hyperparameters such as the learning rate.

[gwyanCN]

Hi, the training loss is the negative log-likelihood, which can naturally take negative values. So there’s no need to worry.

Your training converges, but the issue might be overfitting. You could consider tuning hyperparameters such as the learning rate.

Thank you for your help. I have adjusted the learning rate and other parameters, but currently, my validation NLL loss still seems abnormal, and the trend graph of the prediction doesn't look right either. Could you please tell me what might be causing this?

Here is the loss and prediction:

[zqiao11]

Hi, the loss curves look fine to me. The validation performance is evaluated with three different metrics, and it is normal that one may increase while the others decrease. If you are concerned about PackedNLLLoss, you can add an early stopping criterion that monitors it.

As for the visualization, the forecasting results may not perfectly match the ground truth, which is normal and can be due to multiple factors. For example, it is strongly affected by your time series pattern and context length. Although your time series may appear to follow a relatively simple trend, it is important to consider whether it can be forecasted using only historical data, without being influenced by external factors.