How to set num_samples, top_k, top_p and temperature?

[abdulfatir]

Opening this as a FAQ.

• num_samples specifies the number of sample forecast paths that will be generated to construct the probabilistic forecast. We used 20 samples in the paper for all models. Using a larger value will improve the estimation of the quantiles, likely resulting in better probabilistic metrics.
• top_k, top_p and temperature mean the same thing as in the case of LLMs. Check out transformers documentation for details. We used the defaults from the transformers library and did not tune these parameters. However, one might improve the accuracy of the forecasts further by carefully selecting these parameters (Let us know if you do that and have some insights!). Particularly, selecting a larger value for top_k may lead to a better coverage for certain time series. In our qualitative analysis (e.g., Fig. 12 in the paper), we set top_k = vocab_size which led to better prediction intervals than the default value of 50.