In races involving incumbents, most "undecided" voters usually break strongly in favor of the challenger. This is because most incumbents begin the election campaign with nearly universal name recognition, while challengers lag behind primarily because voters simply don't know who they are or what they stand for. Additionally, if a high negative rating is present, that is a clear sign that the incumbent has limited upward mobility. Therefore, the "undecided" column is generally comprised of voters who have reservations about the incumbent, but don't have enough information about the challenger to comfortably commit their support. An incumbent who is running under 50% is always vulnerable, even if he holds a 12-point lead.
Most sophisticated political observers understand this dynamic, but there are still plenty of hacks in the media who don't. Rather than giving poll results their proper interpretation, too many reporters and news outlets continue to report them like football scores -- placing all the emphasis on the size of the "point-spread." This is particularly true among media outlets that do not subscribe to polls and do not have access to a professional analysis and interpretation of survey results. Too cheap to pay for the information, these news organizations rely on second-hand sources (condensed wire service accounts and campaign operatives) to acquire partial top-line numbers and report the "score." In a situation where an incumbent leads 48%-36% a week before an election, it is not uncommon for the challenger to eventually win 51%-49%. Still, the headline or lead story will scream "Governor Jones leads by 12-points in re-election bid."
There are a few exceptions to this rule, most notably presidential elections. Point-spreads in presidential races tend to hold up better. This is because the challenger usually achieves universal name recognition early in the campaign, and voters form their opinions of the candidates at a relatively early stage of the race. Support tends to "lock-in" more quickly, and late voter shifts are far less likely to occur. This pattern is also true when two high profile politicians square off, such as when a sitting governor decides to challenge an incumbent U.S. Senator. Point-spreads are more relevant because both candidates enter the race with similar name recognition.
Another interesting dynamic that may not be commonly understood is how "undecided" voters perform in ballot initiative and referendum races. In almost all cases, "undecided" voters break heavily into the "NO" column. Since most ballot language is often complex, voters can easily be confused about the intent of a measure. In such cases, the safe vote is "NO", as it maintains the status quo. A ballot measure that leads 43%-35% in a late poll will probably lose 55%-45% on election day.
The major exception to this rule seems to occur only when referendum proponents have a very significant financial advantage over those who are in opposition. It literally takes a barrage of advertising to convince voters to back any controversial measure, and that can only happen when supporters have deep pockets. For example, in two recent local referendums on constructing new football stadiums (San Francisco and Seattle), well-financed proponents were able to win slim victories by heavily outspending stadium opponents. In both cases, stadium support was well under 50% early in the campaign, rose into the mid-to-high 40's late in the race, and ended up at 51% on election day.
Polls can't always predict the future, as public opinion is always volatile and often fleeting. A candidate's comfortable lead can instantly evaporate with one stupid quip or poorly advised policy position. Additionally, survey research is still as much art as it is science, and there is no set of statistical rules that holds firm in every situation. Still, a close examination will show that most polling firms have pretty remarkable track records -- particularly when the habits of "undecided" voters are more clearly understood.