The following is from a fact-checking report published in The Guardian on October 21. The text in bold is what Trump said.
All available evidence shows that in-person voter fraud is exceedingly rare: you are more likely to be struck by lightning in the next year (a one in 1,042,000 chance, according to the National Oceanic and Atmospheric Administration) than to find a case of voter fraud by impersonation (31 in more than a billion ballots cast from 2000 to 2014, according to a study by Loyola law school).
...If Trump loses the presidential election, it will be because American voters do not want him in the White House, not because of a conspiracy involving Republicans and Democrats alike at state and city levels around the nation – a conspiracy for which Trump has provided no evidence.
“The following information comes straight from Pew Research, quote: ‘approximately 24 million people, one of every eight, voter registrations in the United States are no longer valid or significantly inaccurate.’” – 17 October, Green Bay, Wisconsin
Trump is quoting a study about voter registration, not about actual voting practice, as he misleadingly suggests. The Pew study points out that inefficiencies among state registrars are mostly the cause of innocent, if unfortunate, circumstances: one in eight Americans moved between 2008 and 2010, many pushed by the financial crisis, which also strained local budgets, and 51 million more Americans were not registered anywhere to vote at the time of the study.
“More than 1.8 million deceased individuals right now are listed as voters. Oh, that’s wonderful. Well, if they’re gonna vote for me we’ll think about it, right? But I have a feeling they’re not gonna vote for me. Of the 1.8 million, 1.8 million is voting for somebody else. Approximately 2.75 million people have registrations in more than one state.” – 17 October, Green Bay, Wisconsin
Again, Trump is misreading the Pew report to suggest sinister results. The report nowhere suggests that 1.8 million dead people have cast any votes, only that registrars have failed to catch up with deaths. Nor does it suggest that people who registered in more than one state have actually voted in more than one state. The report also highlights problems with registering in the first place: in 2008, 2.2m votes “were lost because of registration problems”, and another 5.7 million people “faced a registration problem that needed to be resolved before voting”.
“The following comes from a 2014 report from the Washington Post, [titled] ‘Could non-citizens decide the November election?’ … More than 14% of non-citizens in both 2008 and 2012 samples indicated that they were registered to vote.” – 17 October, Green Bay, Wisconsin
Trump omits important context for the article in question, which was written by two academics on the Washington Post’s Monkey Cage blog about their own study. The authors admitted their report was “fraught with substantial uncertainty” and relied on a small sample size of self-reporters: 339 non-citizens respondents in 2008 and 489 non-citizen respondents in 2010, about 1% of survey respondents overall. Other academics also noted that there is a high error rate in self-responding among non-citizens.