Even if they have a massive tensor file with five years of captured data, past performance isn't a guarantee for the future, especially not in a network like LN that has been rapidly evolving over these same five years. Are patterns from early LN still valid? I don't think so, but happy to be proven wrong.

I think that there could be useful scenarios, but remember that LN still has known weaknesses in the channel closure process - which is why you run watchtowers - so you need to make a judgement call every time you open a channel. I'd not leave using judgement to something that literally has no capacity for judgement ¹