Betting sites tend to be more accurate than polls because the bettors have access to the polls and they can handicap the poll estimates when deciding to place a wager. They use specific knowledge and seasoned judgment in gauging the bias of certain polls. Election polls have a pretty dismal track record and have gotten too many big races wrong. It’s not that they mis-predict close races. It’s that they are wrong about races they say are not that close. Trump vs Clinton is one. Truman vs Dewey is another. Phone polling is vulnerable to increasing degrees of bias. Who answers their phone these days? Some of the surveys are done over landlines. How much selection bias does that introduce? The sample size needed to pin down a percentage to within +/- 5 points with 80% confidence is about 328. That does not sound like a lot, but you need over 16,400 responses to get estimates for 50 states. It gets worse: given the 80% confidence level there will be on average 20%*50 =10 of the states will have empirical observed percentages off by over 5 points. If we focus on sampling the battleground states, and under-sample the others, we can end up telling Hillary there is no need to go to Wisconsin. Stratified sampling can theoretically get us better estimates with smaller sample sizes, but we need to guess the weights for the cells. That leaves a significant amount of uncertainty. What if Hispanics don’t turn out for Hillary and working class Mid-westerners do show up in record numbers for Trump? Polling provides some information, but it is far from omniscient. Give me the people willing to bet their own money on it, every day.