I'd like to better understand how the lag transforms work. In particular, how do they get applied when predicting future values when they are unknown for predictions but algorithms require them in the future dataframe? To help me understand, let's say I was not using Nixlta but trying to replicate MLForecast using lightgbm. I could calculate rolling stats using pandas similar to what is now natively built into MLForecast. But, how would the future values be created?

By default we predict one step at a time and update the features with the forecasts, we have this very old blog post that goes through it manually: https://nixtla.github.io/blog/machine%20learning/forecasting/2021/06/10/Intro-mlforecast.html#Example

Thank you for this information, @José Morales. It is very helpful to me.