[united]

The unfortunate thing about high danger chances is that generally while "high danger" sounds good, and intuitively the concept makes sense, it under performs almost all other metrics as it relates to correlation to future goal differential and repeatability.

I've looked at Natural Stattrick's high danger metric, which DOES NOT use passing data, and in every single season it lagged every other "advanced" statistic fairly significantly.

Early in the season, high danger chances is significantly less predictive and significantly less repeatable than both corsi and expected goals. The gap continues toward the 50th game of the season, at which point the spread between the three slowly shrinks until the end of the season. Of course, historically around game 60 actual goal differential first equals then surpasses corsi and expected goals (and obviously high danger chances).

So in the end you have a metric - high danger chances - that is significantly less meaningful than corsi and expected goals from the start of the season to about two-thirds into the season, at which point it is less meaningful than actual goal differential. In other words, at no point in a season does high danger chances outperform other metrics so I'm not sure why it's referenced so commonly in its current form.

Of course a big part of that is the sample size high danger chance needs to stabilize which is the largest problem I think: people who don't understand the importance and impact of sample size, and that it isn't equal across all metrics, shotgun the metrics out, for instance quoting a player's or team's corsi, expected goals, and high danger chances across the last 10 games. A 10-game sample of corsi has a much different meaning than a 10-game sample of high danger chances and should be treated accordingly, but it isn't unfortunately. It isn't done intentionally just out of innocent ignorance.

Now, all the above was using a high danger metric without passing data. Bingo has mentioned above that his model does use passing data which would undoubtedly make it significantly more powerful than anything on Natural Stattrick's site and likely up there with private data.