AI: The Holy Grail of Trading?
Let me start off by saying “maybe” to this question and give some context to AI and trading. In 1956 a group of academics decided that it was about high time that they would formulate a definition for “intelligence”, they all picked a summer vacation of 6 weeks, sat together at the now famous Dartmouth workshop. 61 years later we still don’t have a solid definition for intelligence. Of course it’s no secret that when it comes to definitions, academia always finds it hard to agree on anything, but this specific case has been proven to be very interesting.
Before I dive deeper into the subject, I’d like to ask you to spend a moment to “define” intelligence for yourself…. If you are like me you’ll start off thinking of things like “self awareness”, “dealing with abstract concepts”, “ability to plan the future”, etc, etc. The problem with these definitions is that you would need to start defining the concepts in the definition and most likely the concepts in those definitions as well. Quickly you’ll run into an exponentially growing web of definitions. Like any computer scientist will tell you, any problem with an exponentially growing solution space is probably not the way to go.
The closest that we get is a sort of meta-definition by Alan Turing who 10 years before the Dartmouth workshop started thinking about machine intelligence. He was asking the question “Can machines do what we [as thinking entities] can do?”. By formulating the question like this he by-passes a lot of the initial problems we had with defining intelligence. He bluntly stated “humans are intelligent, so any machine that can behave like a human is intelligent”. The conclusion of his article was that if a machine’s behaviour is indistinguishable from human intelligence than the machine can be regarded as intelligent. There are some great articles online on this topic if you want to read into it further. [1][2][3]
From the 1956 workshop till now there has been massive progress in the field of AI. We’ve seen iterative algorithms like hill climbing algorithm[4], neural networks [5], NLP[6] and deep learning [7]. Each adding a new layer of complexity to the field of AI. We see AI’s that work in medicine, play AlphaGo and participate in stock picking contests, but are they really intelligent? By the definition of Alan Turing, no. Yet in all these instances the AI outperforms human intelligence by a mile, which is also part of the problem. Firstly they can only work in that specific field(or space) and secondly they don’t behave as humans would, but better. So here comes the crux of the whole AI space probably of all of modern science. Trade off.
What do we mean by this? Science has become sufficiently advanced that in order to gain progress on one point you lose somewhere else. The most famous trade off is probably in computers where we trade off processing speed with processor size.The faster you want to process the bigger the CPU. The trade off in AI is that if you want it to give you the optimal solution it will take a lot of time to process, because it has to go through each possible solution.Let us take the chess game example. Chess has roughly 7.7x1045states.Taking the current 8 Ghz processors it would take roughly 60 eons to calculate one move. So how do we deal with this? Mostly just minimizing the space (trade off) that the AI has to consider. Like for instance to only look 6 moves ahead it will take a split second. This method of looking at a problem is called Lagrange Relaxation, which is a fancy term for reducing the complexity of a problem in order to get to a result faster. In the study of artificial intelligence as a whole this means that we are currently not building an AI that can mimic all of the human activities, rather an AI that is really really good at one thing and hopefully applying that to new fields of study.
Taking our new found knowledge, how do we apply this to AI trading algorithms? Firstly, we need to define the space that the AI is working in and the goal that it’s trying to attain, secondly, we need to define the constraints that the AI has to consider. Thirdly, we need to define the trade offs we are willing to make. The space in our case is the cryptocurrency space. Let’s say that the goal is the maximization of return of investments. An actual investor would probably scan through various sources of structured and unstructured data, analyze sentiment, price movement, risk factors and other highly specialized models.
Trying something like that is a hopeless task for various reasons, so we have to relax our constraints. For instance only use price and volume data like in traditional investing. Trading can happen at high frequency but unfortunately also lose at the same speed. We can try a more risk averse strategy but this means that our returns are significantly reduced. Put simply we need to make trade-offs in investing as well.
Which trade-offs we are willing to make depends on the individual investor, and creating an AI that fits everyone’s needs won’t be built easily. But knowing that these trade-offs are being made in the background gives you the tools to ask the right questions when it comes to investing your money with trading bots.