The more I look into systematic trading the more I am convinced that alpha is to be found on the turn, where one regime falls and another rises. In between, there is such a thing as 'common sense' which won't generate much but which can at least keep the lights on

Why look at multifractals? I think they may be able to cope, mathematically, with the idea of meta-modelling across regimes

Why look at agent based models such as Bouchaud is pushing? Well I'm not 100% sure of his reasons but for me, again, it's a way to model the turn.

And, yes, calibration is a trap - as @_sbr1 reminded me a while back. So you make sure know what you know, ensemble what you can model, list everything that you can't - and never confuse the three lists.

Coming at these techniques with an options background is useful, it's natural to me to say uncertainty has a price. It's natural to differentiate risks:

Parameter risk: your model is sensible and has value but your wazzou parameter is wrong by 3x and so you lose

Method risk: your model is sensible but your implementation only works in certain situations(timescales, vol regimes, asset specifics)

Model risk: your mental model of the system is just wrong, not applicable. game over player 1

All these 3 risks are very different. If you have a loss the hardest job is to decide which risk to focus on.

Ensembles exist in each case, at least theoretically. Parameter ensembles - obvious.
Method ensembles - somewhat obvious, I think this is what is meant by 'domain expertise'

Model ensembles - not at all clear (to me), that this is useful. It's like having 10 experts in a room. So maybe you don't ensemble these.

But you still don't get tied to one model - i see lots of business struggles as this mistake. French exotic derivatives spring to mind

So: one model, strongly held, until it fails - then pick another. Inside that model: all parameters sets in all implementation methods are valid, weighted in ensembles.

And the language spoken has to hold across regimes: there's more to this than a backtest, especially when everyone has a backtesting tool.

Tail risk - useful: either you are short tail and don't know, or you are short and do know (institutions), you can safely leverage once you own it (Bhansali), or you can dynamically allocate to it. Make it part of any model, implement by ensemble.

Yes this is all generalities but hey it's late and this is only twitter

Something more specific:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2951348