For people who develop and test their strategies, do you try to understand why it works? Fo example, your strategy shows great results and when you dig down you realize it was buying dips. So you say to yourself - great I have a good strategy to deploy in a strong bull market, but keep it on a shelf otherwise. What I don't understand is how people use ML etc to get seemingly random strategies. How will they trade it without understanding how/why it works? How will they know when to stop? Are they just trading a basket of uncorrelated strategies? Would be interested in how people on ET approach this.
Every strategie is a 'random' strategie at the exact moment you enter the trade. You do not know the outcome of that trade, no matter how you come to the idea why you have entered that trade. That said, it does help if you know 'why' you have entered and why you will win in the long run with your strategie. So the strategie can not be random... A random strategie is an oxymoron.
Sure, but a sequence of trades should realize your edge if you have one. HFT strategies, for example, have clear edge which can be understood(but not necessary realized). So when they stop working you can quickly tell that something has changed. If you don't understand your strategy, how do you know when it brakes? How much trust do you need to put in your ability to test, code, and quality of the data!
A strategie not known, is a strategie not worth trading. Even a strategie that make sense to you because of testing or other 'magic', is still vulnerable to the things you do not know. History is filled with people that believe in their strategies for what ever reasons and go bankrupt doing so. It al comes down to how you would handle your strategie with the money you have. That is the real issue in any trading strategie, not 'knowing' or 'understanding' why you have a winning strategie. I hope it makes sense...
Let's say you flip a coin. Do you need to understand that why it is heads this time? Or do you need to understand that in the long run you have edge or no edge? The coin is known for two sides, but trading strategie is not known for two sides, it has unlimited variables. The only thing you really know, if you trade an unleverage product, is how much you are going to lose when wrong. If you understand that part of any trading strategie, you are going to win.
i worked with a cray for a couple of decades running zillions of automated systems for many big firms. with a cray it took five minutes to produce a years worth of figuring out what it found. i kept seeing the same methods produced over and over again but no one including myself wanted to believe it. it was not important to understand the edge, but it was essential that you use it. i say looking back, use it and then in your spare time figure it out.
Do you need to understand where the edge comes from? -Yes, so when you lose it you will know why and how to get it back
Two thoughts: ☼ Not knowing the details of an algorithm is a trading disaster waiting to happen. The market over which that algo was formed simply does *not* last forever, and if you're not monitoring, and *knowing*, when your expectancy goes to Suck! level, then you're going to lose not one or two trades, but a string a mile long. ☼ Machine Learning, though, is not a set 'thing' -- so a ML-derived trading system is itself ever-re-inventing, ever monitoring performance and seeking to modify things according to the programing precepts. The responsible trader needs to know that set-up in detail, sure -- but to know (or *need* to know) the particulars on a specific day? Would not help the overall results. When the pooh hits the oscillator, those who know in detail do not need to waste time trying to figure out what's going on -- they already know that part, and *should* already have a responsive plan. When time matters, they are many steps ahead of those who don't.
Personally, I would not trade an algo that was generated from some deep learning routine unless I knew the mechanics of how the deep learning routine worked. I use supervised machine learning where I am in control of the features that could make up an algo so it is easily understood what the algo is doing and what sort of market dynamic it is exploiting.