Maybe I can shed some light on this issue given that I have spent the last two years mostly working on forecasting and economic modelling using large datasets.
What we usually do is we define a recession as does the NBER: we have a monthly variable available through the Federal Reserve Economic Database (FRED) which assigns a 1 to periods the NBER called a recession and a 0 to periods the NBER called expansions. Of course, we could as well come up with a different method for defining recessions and use this classification instead, but we'd be quibbling over a matter of one or two months around turning points. When you're mostly interested in forecasting 12, 18 or even 24 months ahead, that's not exactly interesting. Once you have a classification, you can compare different methods for assigning a probability of recession ahead of time. The way we do it is essentially recursive: we use data, say, up to 1990 first quarter, we train our models on it, and then use the models to forecast 1990Q2, for example. Then we add 1990Q2 to our sample, we retrain the models and then forecast 1990Q3. This way, you imitate the conditions in which you would have used the models back in the day and you can get an idea of how it will perform when the future really is unknown -- i.e. when you'll use it for real.
I know a guy who did this in a very serious manner: he used vintage datasets (economic data get revised over time, so you're cheating a bit if you use the latest release), he looked for when the NBER declared recession dates (even if you're in a recession in 2001, he does as if he didn't know because the NBER calls them with delay), etc. His models use the yield curve and he produces an entire term structure of forecasts (i.e., he forecasts 1 quarter to 8 quarters ahead). For reference, this is a paper by
Kotchoni and Stevanovic.
One interesting point he and his coauthor makes is that the shape of this term structure is generally like a bowl (it's convex), but that since the 1980s it always has inverted (like a dome) prior to recessions. On his website, you can see updates in day by day for the probability of a recession:
DS_forecasts
If they are right, there should be a slowdown in 2020.