I ask Nate Silver to estimate some parameters.

It seems to me that in many senatorial races the actual outcome was closer to the last few polls than to the smoothed average. That is it seems to me that less smothing (more sensitivity) would have been better for Senatorial races. Nevada is an exception, but this works for Penn and Wash (closer than smoothed average although they are still counting in Wash) Colorado (other sign than smoothed average) and Kentucky (wider than smoothed average).

My thought (which I believed before the election so I might just be seeing what I expected) is that optimal forecasts involve less smoothing for Senate electoins than House elections or Presidential elections. The logic is that people are learning about the candidates in October so population voting intentions change a lot.

In contrast people have been inundated by information about Presidential candidates since it seems like forever so they have learned all that they are willing to learn by October and made up their minds unless they are determined not to.

Also in contrast, most people never learn much about candidates for the House of Representatives ( I strongly suspect that many voters don't learn the names and just look for the D or the R).

I'd guess gubernatorials are like senatorials.

In any case, the question of which smoothing parameter works best is an empirical question and can be answered using your data set. I'd be interested in estimating optimal smoothing parameters and testing the null that the optimal smoothing parameter is the same for senatorial, presidential and house elections.