I think what I am saying is that the short-term forecasts are good in terms of a *dead-reckoning* algorithm. They can see where the trend is headed, but they can't see the bend in the curve that lies too far ahead. In other words, the model for that is non-existent. And that is where the 2 to 4-year time frame is important because that gives the forcing and characteristic period of the physical mechanisms. Beyond this point, the data will inflect. These pseudo-cyclic periods may, or may not be, extracted by the machine learning.
Perhaps when the learning interval is too long, it simply averages out the cycles, leaving nothing to dead-reckon with. At least with a short interval, it can see where the data is going over the short-term.