Say I have two variables, y_1t and y_2t, which follow a VAR(p) process. To forecast the n-period-ahead value of y_1, I can estimate this VAR(p) model, then do one-period-ahead forecast recursively to get the n-period-ahead value.

Alternatively,since the n-period-ahead forecast of the above method is ultimately a function of y_1t, y_1(t-1), …, y_1(t-p), I can just regress y_1(t+n) on these lags directly.

My intuition is that the second method is more efficient, but why do people usually do n-period-ahead forecast in the first way?