The reason is because the stuff is complicated and esoteric. It's going to be using digital signal processing techniques and advanced ideas from volatility pricing and/or market structure, synthetic instruments and so on. You need expensive data feeds, and co located servers and shit. Funding for the technology. AI is pretty serious. You would basically need teams of technicians monitoring the thing, and that costs money. I use a similar type of thing. But it's as simple as price transformations, differencing against estimators, and tracking functions of price or price/estimator differentials. nuclearphynance, quant stack exchange, wilmott, and so forth will be doing things that are related, but AI (as in trading robots or managing trading robots) could mean a lot of different things. How much do you know about this stuff?
As an algo trader (with academic background in Economics, Stats, Finance and Comp Science), I steer away from AI techniques. From what I gathered and heard, AI may work in HFT order book area but not in slow to medium term trading. Ernest Chan, who once worked in IBM Watsons team steered away from using AI in his own trading until recent years because he mentioned that it's very prone to overfitting. And based on my experience in non-trading work, I can attest to that. I once applied some conventional techniques to backtests and was concerned that mere changing of random seed numbers could have vastly different results (according to Ernest Chan in a podcast, he used 100 different seed numbers and create some sorta ensemble or voting mechanism). There're still many risk-premia to exploit in the financial world to exploit before I will consider using AI. e.g. volatility risk premium, trend-following, mean reversion, risk parity, etc.
The second you mentioned Erny you lost me. Yes, we are in full agreement. Pseudo science and mediocrity does not correlate well with the hard work and intellect required to put deep learning and AI in general to good work in financial trading. I won't elaborate further but one thing I can guarantee you does not work is AI and hft. Keyword: latencies. Both together are not compatible.
I must admit that I'm not well verse in HFT area. How about trained models which just read in data and carry out scoring only? There may be algorithms with high time and space complexity for training but not for the scoring portion.
Like anything, people repeat what they've heard, or if they've tried without success, they'll say that it doesn't work. AI is just another tool, one must know how to use it before making it worthwhile. To respond to OP's original question, there are many groups in LinkedIn who talk about it. But many of them are focusing on the academic aspect.
I do not know where, but I like to talk about it. Actually investing time and effort in it for trading, not so much. AI obviously is very powerful when it comes to pattern recognition in certain domains, but in its current form it lacks imagination.
It could, reinforcement learning with function approximation could be used. But I think it will be very hard to make it work due to changing markets.
Actually, pretty much the only place where I've seen complex ML used in finance is high frequency trading. In fact, people were using random forrest models to play order book games before the current data craze (like late 2000s) and these days they are probably the only ones who use any forms of reinforcement learning. For example, if you try to iceberg an order, it's usually an ML model that will figure out what you're doing and will try to tick you. PS. It obviously requires care to separate the learning and the actual forecasting so you don't slow yourself down, but people manage. E.g. they guys I knew who used a random forest, they would train the model in a separate thread and used a huge chunk of shared memory to write out a dumbed down deconvolution of the tree. This way, the latency sensitive thread would simply do an array lookup which is pretty fast.
I don't believe that. Hft firms invest many millions in FPGAs, not a single inference algorithm on trained ML or DNNs is fast enough to not defeat the entire purpose of those specialized hardware units. Where have you seen them employed if you don't mind me asking?