Nate Silver of fivethirtyeight.com, who made his reputation with statistical models accurately predicting baseball outcomes and absolutely nailed the 2008 presidential election results, has two provocative posts exploring what the best possible outcome would look like for Republicans and what the best possible outcome would look like for Democrats. He's actually predicting Democrats hold the Senate with 52 seats and Republicans gain 54-55 seats in the House, but says it is equally possible that Republicans could take over the Senate and gain 77 or more House seats if polls underestimate their support OR Democrats could limit losses to around 30 seats in the House and five in the Senate if the likely-voter model is underestimating likely Democratic turnout and polls are just plain missing voters who do not use landlines.
Here is an interesting piece from The Daily Beast compiling how far off experts were in 1994. And below is Silver talking about the ways in which predictions missed in 1994 and 1998:
In ... 1998 ... Democrats overperformed their polls by about four points in a great number of races around the country. What was supposed to be an echo to the Republican boom year of 1994 basically flopped, eventually costing Newt Gingirch his job as majority leader.
Consensus expectations also considerably underestimated the Republican wave year of 1994, although a few indicators (like Gallup's generic ballot poll) got it about right.
If we wanted to be generous to Democrats (which is, of course, the purpose of this article), we could say that the consensus basically failed in two out of the last four midterm elections. Of course, that the consensus view could fail does not mean that it will fail in the Democrats' direction: instead Republican gains could be much larger than expected.