Is there any research in this area?
High-frequency trading is where you see it being used, essentially, decision-making algorithms analyzing and making transactions in microseconds. It accounts for a significant percentage of market activity, and has been considered a source of greater market volatility (See: Flash Crashes).
You can bet that hedge funds are evaluating every form of AI for trading and predicting market trends in general, but, unlike academics (and a segment of the tech sector) I doubt they consider it in their interests to publish methods which competitors could then utilize.
It's difficult to source information on exactly what is being looked into and utilized in the financial sector because there is a lot of marketing noise (exaggerated claims, unreliable sources) but I did find these articles:
My gut tells me that Machine Learning methods will surpass humans in these kinds of decisions before too long (educated guess), and only increase in utility as the dataset these algorithms draw from grows, because it's ultimately a statistical problem.
AI is related to Game Theory, the study of economic decision-making, in the context of utility. Game Theory might be said to have had its greatest success in computing in general, via minimax, but has traditionally been much harder to apply in real world economics. Minimax can be utilized in machine learning, and it's actually hard to think of any economic decision-making that wouldn't utilize it in some form.
Machine learning based hedge funds are currently the worst performing slice through the industry. Large quant funds also claim to use ML but their performance is also very poor and worsening over time. Mostly what they say they are dong is not really what they are actually doing. I know of cases where funds claimed to be doing AI but actually its just simple technical analysis. Successful applications are more in the area of data collation of sentiment using NLP or feature extraction using dimensional reduction but these are not actually part of the trading strategies themselves. The only ML influenced trading I know is HFT trading of the order book which is a well define problem divorced from the actual price series itself. Apart from that 99% of ML use in the hedge fund industry is solely as snake oil for their marketing departments. Out of 10,000 or so hedge funds world wide only 0.25% generate alpha according to academic studies. That's 25 out of 10,000 ... and many of those will be cheating like SAC.
The fundamental problem is that ML requires data that is high dimensional, highly structured and is low noise and has stationary dynamics and statistical moments. Unfortunately financial price series are low dimensional, unstructured and noise dominated (negative SNR) and exhibit multi-level non-stationarity of the underlying processes and statistical moments in a manner highly related to multi fractals. It is hard to imagine a time series less suitable for machine learning.
The industrial age in the 19th century was dominated by mechanical machines who were able to do useful work. For example a knitting machine can produce clothes, while a steam engine can generate power. The shared feature of machines for automation is, that they doesn't require a decision by the human. In the 20th century the information age was triggered by machines who can't be used for production but for games. A roulette wheel is a typical example, but a slot machine too. The idea here is that these machines are requiring a decision process from a human. He has to think about how to interact with the device.
The logical consequence of a roulette wheel are more advanced machines which are supporting stockmarket speculation. By definition, the betting with money has nothing to do with work but with decision making. This is researched by newly founded disciplines like cybernetics, game theory and the science of economy. Such a development was taken place in the 20th century. The idea of managing a hedge fund with Artificial Intelligence is maybe profitable or maybe not. In each case it will teach a lot about game theory, psychology and about the inner working of financial markets. It is a typical problem for the 20th century, not for the industrial age.
The topic is very complex. .A good starting point would be first to realize a money management system which is equal to the doubling strategy, known as Martingale system. The player has to increase the betting sum, if he loses something. This allows him to win money back. The second problem to answer is to predict the stockmarket itself. This can be done with the ID3 learning algorithm which generates a decision tree. The input for the ID3 algorithm is an Excel sheet with financial data from the past which contains feature. A feature could be if the share has fallen or climb, if an indicator of another share was positive or not. The principle of decision trees is discussed also for weather forecasting.
The classical decision tree is formulated in propositional logic which is not turing complete. The more powerful way for storing the decision rules is a boolean cellular automaton. The physical representation is called vibrator circuit.