Yes this is accurate and put very well. This is so much the case that if you have a strong background understanding of the field, the ML part can actually be picked up quite quickly or contributed by someone else. There are a few notable users who are both domain and ML experts and they tend to absolutely clean up in their field. I'm thinking of a couple of med students in particular who are formidable in every medical imaging competition.
I recently read 'The man who solved the market', about Jim Simons and Renaissance Capital. The way the book tells it, looking for patterns without seeking domain expertise (e.g. ignoring fundamental valuation of equities) is exactly what Renaissance did, and it worked out very well.
But all of this has to be utilized in the context of the data. The reality is that you're not going to develop a sophisticated options trading strategy without a strong understanding of what an option (and more generally, a derivative) is. You can't develop a viable statistical arbitrage strategy just by treating market microstructure as a blackbox signal to be solved with e.g. Fourier analysis. You can certainly find an edge in using fundamentally superior methods of analysis, but you still need to know what that data represents in the context of the market.
Don't be fooled: people working at firms like RenTech have a strong understanding of the underlying finance. It's just that they learned it on the job, because the ethos at these firms is that learning fundamental theory in math and statistics is harder than learning fundamental theory in finance. You don't have to take my word for it though. Read about one of the few strategies of RenTech's which has been publicized: https://www.bloomberg.com/opinion/articles/2014-07-22/senate.... Deutsche and RenTech didn't team up on this strategy (to fantastic success) by treating basket options as some kind of blackbox abstraction devoid of delta, gamma, theta and vega.
(Also, even if I agree it was purely intended for tax avoidance, I don't understand why you think that would obviate having to understand how the options work intimately well).
https://www.rentec.com/Careers.action?computerProgrammer=tru...
They look for programmers with knowledge of Tax and Risk Management.
Kaggle encourages a domain agnostic approach to modeling, in the sense that participants use sophisticated machine learning and statistical methods but typically have no domain expertise in the underlying data. This kind of approach to finance has historically performed poorly. [1]
Good quantitative trading is usually backed by a strong fundamental thesis and an interpretable model, which is obtained by cross-pollinating sophisticated math and statistics with domain expertise in some part of finance. That domain expertise might be in different kinds of assets, liquidity or market microstructure, but it's there.
$100k is cheap for Jane Street. If nothing else they have a new recruiting pipeline of people with demonstrable machine learning skills.
______________
1. I would also say this is a poor way to approach statistical analysis in most domains, and usually leads to spurious or overfit results. But the idea that you can just run a model and find patterns in pricing data is especially attractive and insidious.