We present the promised second part on the use of AI tools in trading. In it, we show how the new tools can be used in an example fund. It consists of discrete traders and the head is a trader who has practically been trading oil for several decades. Part I can be found in the first issue of our magazine
The Fastest Path To Cash: Support The Best And Improve What Already Works
Let’s go back to the fund and ask yourself – what tools to use and where. This is the question of where investments can bring the fastest, measurable benefits: stop the erosion of profits, reduce risk and perhaps increase returns.
What I am describing now is not a universal prescription, because each situation may be different. From my perspective, I provide topics to think about. I would say this – the general rule is that we should support what works best, brings the best profits and what carries the most risk. In the case of a fund, or discrete traders – this is to support the best traders.
Who in our fund is the best?
In the described fund, the best trader is the holder. Support for him is obvious as long as he wishes, of course. However, AI support in this case can fail, as we are about to demonstrate. We get the best practice from implementing AI tools when we support, improve, automate what already works well. In general, we will get the best results by supporting what works best.
If we have a group of top traders then they are a fairly obvious choice.
However, one more condition must be met – they must be open to new methods. It happens, for example, that the statistical approach is so far from the trader’s experience that he is unable to switch. He is unable to combine the insight provided by the new knowledge with what he already knows. The result is unnecessary tension, and investment and time are lost.
The best usually don’t have a problem with this – that’s one of their key characteristics. They are open to change, they know that the market is constantly changing and you have to adapt. As an example, let me mention Larry Williams, who I interviewed for a book a few years back. He is 77 years old today and you can find his current lectures on the Internet about the new indicators he is creating. Age is no barrier to taking advantage of new things. Our trader, the head of the fund is a man in his 70s and yes, I think he is the one who should be supported if he wants to.
The second question is – where is the greatest risk and how can new tools reduce it?
The third area is (and here I think there is very great potential) the implementation of new MM (money management), I will write about this in future articles.
And some more tips at the end
If you are thinking of a larger AI department (using only new tools) then create a separate team. In the next issue I will describe good and bad practices in this area, i.e. how not to lose tens of millions using what others have already burned on.
New Strategies For Managing Position Size For Different Classes Of Signals
I often refer to what the best discrete traders do because it is their practices that are worth emulating and improving (through automation).
The best ones adjust the size of the position according to the “goodness of the signal,” in the sense that they are willing to increase the position when the signal is exceptionally good (according to their evaluation criteria).
If we start from the point that some signals are better and some are worse then we can move on to the creation of a seamless MM, based, for example, on statistical evaluation, assessment based on the results of the application of Machine Learning.
The starting point can be a simple observation that the best inputs may have something in common. What is it, what are the factors – this is the task for the ML tool (classification of inputs due to various parameters).
If we find this, then we can count new dynamic rules for allocating position size to each signal ( depending on the parameters that characterize the input – whether it is “better” or “worse – to which class it belongs).
This approach opens the door for us to create very liquid MM strategies – the best signals get the most and are pyramided and the weakest (but still “valuable”) – get less or only “control inputs”.
In the last ones, for example, we can try “control entries” and if the position holds and is secured – we add so much that if the market goes against us – we will come out at zero (this is one of the simpler strategies for “weak” signals).
Why is this important? These signals are not useful enough to be overlooked by other traders and other systems. In case the bigger ones with very strong AI systems seize larger and larger portions of the market these very signals may remain the longest. One of the most efficient billion-dollar quant funds has specialized in just this approach for many years.
More complex MM strategies are practically out of reach for discrete traders, requiring too much work, beyond the capabilities of a single person or even a team.
Add to this the fact that a single position can be broken into thousands of smaller ones (to minimize adverse moves and dynamically adjust to pending and incoming orders) and this absolutely requires the use of intelligent automation.
Hiding Entries And Exits
If I wrote that brokers and market makers copy the positions of the best then many might be offended. So, I will not write this (oh wait!). This is one of the oldest tricks in the book, the simplest and very profitable, used since the beginning of the stock markets.
The best ones do it in several creative ways, from what I’ve observed in the life cycle of a fund there comes a time when entries are hidden to preserve their profitability. Too many eyes are watching, many smart people are able to “crack” the strategy and it doesn’t take intelligence at all to copy.
Some of the best go to the extent of not only hiding the inputs but also hiding the intent, the system, the strategy behind the inputs to prevent reverse-engineering it. This is understandable, if they have something that works well (for which they usually paid a lot, sometimes tens of millions) they want to use it for years.
Here I want to talk about two strategies I’ve observed, so that if someone hasn’t dealt with it yet they can form an opinion.
The first strategy is to split the inputs between two, three or more brokers. The system has a certain number of profitable inputs and a certain number of losing ones, and the split itself can cause fluctuations – perhaps more losing ones will flow to one of the brokers?
It is possible to consciously (if such a practice is suspected) direct inferior orders to any of the brokers. It is also possible to do such things routinely, changing from time to time to whom the better and worse orders are directed. This is bound to reflect on the copyists, as the system will not be as good as expected.
Another strategy requires having or running its own HFT trading department. One large fund does several hundred thousand entries a day (between two hundred and four hundred thousand entries), and among these HFT hides its entries from other systems.
This fund is operated by two very large brokers.
If someone had access to one of the brokers and tried to crack its strategies, he would first have to filter out the large-scale signals (5 minutes and above) from the tens of thousands of HFT signals. This is of course possible but … the other problem remains – neither path gets the system fully. Second – the fund uses not one strategy but dozens. And breaking this is already a non-trivial matter (but still probably possible for a patient statistician).
But then there is another matter – the fund employs a lot of excellent mathematicians who, over time, will surely realize that one path has worse results than another and ask themselves why (the fund has already had to deal with the use of its signals and knows the subject from experience).
They won’t necessarily be able to prove that someone is sneaking their signals or trying to break their strategies, but they might just want to change brokers.