We have run and released polling for three of the special Congressional elections in 2017-2018. Most recently we had the Republican candidate winning Arizona’s 8th Congressional District by 6 percentage points: she won by 5.2 percentage points (i.e., we were off by 0.8 percentage points). Early in 2018 we had the Democratic candidate winning Pennsylvania’s 18th Congressional District by 1 percentage point: he won by 0.4 percentage points (i.e., we were off by 0.6 percentage points). In 2017 we jumped into the Montana race twice. We had an early big lead for the Republican candidate, but our same-day polling on Election Day showed it shrank to 5.5 percentage points after he assaulted a reporter: he won by 5.6 percentage points (i.e., were off by 0.1 percentage points).
Three elections and an average error of 0.5 percentage points. Three elections is a small sample, but this is a really good demonstration of what our faster, deeper, cost-effective polling can do.
Note on Methods: We collect the data via smartphone with Pollfish. This is a new and exciting way to collect data (and possibly the only useful mode going forward, as landlines become obsolete). Since we do not have a representative sample of Americans (no one does), we make use of the advent of Big Data and advances in machine learning and statistics to process the raw data and get representative estimates not only for Americans overall, but for more fine-grained demographic categories. This methodology is well validated by the academic community, our prediction of the 2016 general election, and a number of validation studies.