Here's the latest post I was planning to publish this week or next week: https://www.datasciencewithmarco.com/blog/chronos-the-latest-time-series-forecasting-foundation-model-by-amazon It's on Chronos. I first explain the framework, how it works (basically rehashing the paper and trying to make it more interesting and less dry) and then do a small test of Chronos. Full source code is on GitHub and I link to the benchmark done by Nixtla. Let me know what you think and when I have the green light to publish it!

Hi Marco! We read your post, its really good! Glad that we found similar conclusions independently 🙂. The only comment is that you are currently importing the losses from nf:

from neuralforecast.losses.numpy import mae, smape

Can you instead load them from

utilsforecast

? We are trying to increase the usage of this library. They are also in the losses module (https://github.com/Nixtla/utilsforecast/blob/main/utilsforecast/losses.py). Also, if you can write a very short paragraph just mentioning that the library contains several functions commonly used in time series analysis (plotting, evaluation, preprocessing, etc.) it would be great. Again, amazing work!

Thanks for the feedback! Will change that 🙂