There has been a lot of posting about how the undecided break for the challenger, etc in a variety of threads, so I figured I'd toss up a bit of information, and hopefully shed a tiny bit of light on the topic...
A new thread seemed as good a place as any...
There have been two major studies that I have personally read on this topic, one from about 1990, and the other from about 2002 or so.
I looked for them on the internet, but could not find them, so I am going from memory here. If you do find the studies, and I am off a few percentage points.. don't shoot me..
I am confident I remember the "broad strokes" pretty accurately, but the percentages might be off a tad here and there...
There was a study done back in 1990 (?) or so which did a simple anaylsis of what the polls said before the election and what the actual results were.
This study concluded that about 80% (+/-) of the time the challeneger did indeed do better than the polls suggested, about 10% of the time the polls were right, and 10% of the time the incumbant did better than expected.
This study had a number of flaws however. Untill about 1990 or so, Republicans historically underpolled a bit - on average by about 4% or so. This combined with the fact that until the 1990s the Democrats held a crushing advantage in the number of state, local, and congressional seats badly distorted the data.
Because the Dems were overwhelmingly the incumbants, and the GOP systematically underpolled, if just looked like the "break" was mostly to the challenger.
When this bias was corrected for in later studies, the "break to the challenger effect" was dramatically diminished, but certainly still there.
After correcting the data for the systemic anti-GOP bias the ratio was that about 55% of the time the challenger did better than expected, about 25% of the time there was no break, and about 20% of the time the break was towards the incumbant.
It should be remembered that, due to random error in the polls, the "normal" result would be 33/33/33 if there was no "break" effect. (A challenger would, due to random chance, underpoll as often as he/she overpolled)
The study also went a bit further and in an utterly arbitrary way divided election races into two categories - Close (under 10% lead shown in the polls) and not close (Over 10% lead shown in the polls)
There is a fairly stong effect where in a one-sided race, the challenger does better than expected. It is fairly common to see somebody go into an election with a 30% lead and then win by "only" 19% as a bunch of his/her supporters stay home on election day. This appears to impact Democrat incumbants a bit more the GOP incumbants, but the gap is not huge.
This is what happened in 1996 - most polls had Clinton beating Dole by 12% or better when the actual result was about 8% - CBS News Polling for example missed this election rather badly having Clinton up by 18% when the actual result was about 8%
A lot of Democrats stayed home, while Republicans went to the polls to protect their hold on the House & Senate.
Within the "close" races, the "break to the challenger" effect was still there, but slightly reduced again. - About 45% of the time the Challenger did better than expected, about 30% of the time there was no break, and about 25% of the time the incumbant did a bit better than expected.
The average "break" at the end was about the same size for both incumbants and challengers and averaged a bit over 3%.
Summary:
A bit under half the time there is an (on average) 3+ % break to the challenger at the end.
About 1/4 of the time there is a 3+ % break to the incumbant.
About 1/4 of the time there is no break at all.
Hope this helps!