Skip to main content

View Diary: Rating the 2010 election forecasters (54 comments)

Comment Preferences

  •  Not all errors are created equal (3+ / 0-)

    I'm not sure how I'd do this rating differently from you, but it does seem that I care much more about the 1-point difference between a .5-point win and a .5-point loss than I do about a much more significant (by your methodology) difference between a 13-point win and an 18-point win.

    Again, I'm not sure how to fix it, but it might involve weighting predictions towards the "0" point of the scale (where the Dayton, McNerney, NY-1 etc. races cluster) much higher than those that are further out.  Maybe simply dividing the error by the summed absolute values of the prediction and the result would be a good first approximation of who did better "when it mattered."  Then my 1.0 difference would become (1.0 / (.5 + .5)) = 1.0 and the five-point difference would be (5.0 / 13 + 18) = 5/31 -- about 1/6 as important to the ranking -- which seems plausibly appropriate.

    •  Good point (1+ / 0-)
      Recommended by:
      Seneca Doane

      I don't really know how to solve that either, but you are right. Not all differences are created equal. I will think on that for a while.

      •  Well, if you're interested in trying out my above (0+ / 0-)

        suggestion as a first approximation to how weighting for "decisiveness" might look -- assuming that it would not be hard for your to do so given that you already have the datasets in workable form -- I'm very interested in how it would come out.

Subscribe or Donate to support Daily Kos.

Click here for the mobile view of the site