Yep, Dixon-Coles is way out now. Generally, Poisson-based models don't work that well. They model some aspects of the game correctly but not others. And when it is bad, it is terrible (their paper was genius though, and changed the industry). You also wouldn't do the one-hot/regression stuff on teams, as there is so much player-level data...some kind of online, off/def model is useful though.

You also have the issue of modelling a bivariate as two independent univariates. The adjustments made by the authors were intended to get around this problem but there has been a lot more work on this aspect of the problem since then, and a few new solutions.

And you are right, expected goals was cutting-edge back in ~2011-12 but even some clubs have caught on now. Most gambling syndicates, afaik, actually produce their own datasets (the largest ones employ a huge number of people, I know one that started up five years ago and already has ~10 PHds on staff + traders + data people, etc.).

I will say though, it is surprising how inefficient betting markets were until recently. In the early part of the 2010s, I imagine that large syndicates were making, conservatively, 300-400%/year in the largest markets that were, apparently, "unbeatable". I am aware of very simple models, similar to DC, that could beat retail bookies on Brier for EPL as recently as 2017. And, ofc, smaller markets are far softer (or other sports). At the cutting edge, yes...it probably is complex (I have no idea). But a surprising amount is actually just doing the simple things well.