Thanks for evaluating Moirai. There are some issues with the way inference is being done for Moirai:

1. Using the mean of sample forecasts instead of median. In GluonTS, metrics involving an absolute error typically use the median of sample forecasts rather than the mean.
2. Incorrect default parameters were being used. Our paper uses context_length 1000 and patch size 32 as default hyperparameters for the Monash experiments.
3. Extremely small batch size is being used, slowing down the model significantly. Differing batch sizes are being used across the various models - with the new default parameters, a larger batch size should be possible.

This PR fixes these issues, and below contain the updated results. The results for the fix was run on an A100-40G GPU, but the appropriate batch size can be updated when running on your hardware.

We note a significant improvement in accuracy across all metrics, as well as inference time.