This is a nice and surprisingly important distinction that matters a lot in practice:<p>> Notice we get good out of sample performance. If we are judging on quality of out of sample predictions (and not on system identification or inference) then direct fitting k-step out predictors looks like a justifiable strategy.
Actual article: <a href="https://github.com/WinVector/Examples/blob/main/TS/TS_example.md">https://github.com/WinVector/Examples/blob/main/TS/TS_exampl...</a>