Why is first training slower, but subsequent trainings faster

[ngupta23]

I am curious by the first training for AutoARIMA using statsforecast is slower compared to subsequent trainings (even though the models are instantiated again before the subsequent trainings - i.e. new objects). Is some information being reused from the first training?

https://gist.github.com/ngupta23/59cc0ce155048f72b80a0431c57b7d17

Initial Training

Subsequent Training (independent object)

[hrzn]

Hi @ngupta23 , I haven't checked in details, but I strongly suspect that this is due to JIT compilation by Numba. The code is compiled to LLVM on the first run, which takes time, and subsequent runs re-use the compiled code.

[kdgutier]

That is the case @hrzn and @ngupta23 , the first Numba compilation takes time then the code is reused.
We are pondering the possibility of creating pre-compiled wheels, to avoid this issue.

[ngupta23]

Thanks for confirming this. We can close this as the question has been answered.