My "Holy Grail" comment was in response to the above. You can find and edge in anything, but you can't say one has better edge than the other, imho. When I say "Holy Grail" I mean something that is printing money and as far as I know, it's either inside info or some of the HFT strategies. Everything else is subject to the normal probabilities. Because I know what it costs to build consolidated limit order book (CLOB) in U.S. equities (tradable one, not hypothetical one based on canned data). If you know what your system needs, you can easily estimate the costs, right? Anything that requires low latency is very expensive.
From my experiments so far, I can tell you order book flow has a definitive advantage over analyzing price alone. From what I've seen the best edge is to analyze: order book flow+price+volume+correlated instruments. Speaking about algo-trading in short term time frame.
%% I take your word for that.....correlation. We also saw a derivative trader post, elitetrader ,post many , many pages of [profit ] daytrading.He made a funny but more practical [profitable ] than accurate comment;''Its correlated to it self.'' LOL; but he made plenty of public daytrade profits over many , many days.
From your posts I inferred you were trading on order-book-derived features on a single asian crypto exchange using some sort of LSTM or CNN ensemble, and were looking to try the same thing in US listed instruments. Has that changed? Are you now trading order book dynamics in US futures or equity names? If so what broker or prop firm did you end up with? This type of LOB feature extraction has been widely publicly known at least since the Penn-Lehman competition in 2005. Definitely easy edges existed then, even for automated retail traders. What makes you so sure those edges haven't been traded down to the risk-adjusted FF rate in the 13 years since?
yes, that's pretty much correct, although it's not plain LSTM or CNN technology, it's something more sophisticated that works better I'm in the process of testing it with bunch of Futures CME contracts. Still early to say if my system can find an edge or not on these instruments. Collecting data, and building the right features and running it on supercomputer takes lots of time. I'm sure in some markets the most obvious edges are gone. But lets say there's 40k instruments in the world, let's say 1000 is liquid enough, I'm pretty sure my system can find edges on some of them. Although I'm not sure about one specific instrument, from what I saw so far in my experiments, I'm very sure that my approach and technology will be able to find these edges given the right market, it's a matter of testing and lots and lots of CPU power.