They use numbers to come up with probabilities, but there are a lot of baked-in assumptions about things like turnout, various sampling errors or biases, etc.
They have reasons for their assumptions, such as using information from previous elections. But the current election may violate those assumptions, but it's hard to say how much and with what probability.
I trust 538 for the most part. But it's just models and those models have assumptions and it's quite possible the assumptions were just wrong in 2016. We only have one data point, so we'll never really know.
The worst part of figuring out the trustworthiness of a source is putting in the time and effort to comb over all of its output and comparing it to reality, an effort that can take a long time.
Which lends me to ask: how can you say that you trust 538 if you also say there is only one data point? Because there shouldn’t be: 538 made big news during the election of 2008. 538 has been around since 2008 so you should have a lot of polling predictions to comb over.
I actually don’t have an opinion of 538 as I haven’t put in that work.
"The p value and the base rate fallacy" https://www.statisticsdonewrong.com/p-value.html
"Misinterpretations of Significance: A Problem Students Share with Their Teachers?" "https://www.researchgate.net/profile/Heiko-Haller/publicatio..."
Or about how Omicron defies statistics. No, it doesn't.
People just don't fucking understand statistics.