That's not a matter of semantics as ML and deep learning are vastly different concepts and calling an approach ML when it's actually specifically and clearly deep learning is.... I am perhaps the only one here but I am hanging myself out of the window in saying that your claim of hundreds of strategies cannot be accurate. Whether you intentionally mislead (to obfuscate) or whether you are perhaps a bit clueless about what you are doing I am not sure. But nobody in this space would make the statements in regards to deep learning and machine learning that you have made. It's simply a reflection of you not being well versed in this space. So, I am not sure what else you are perhaps misunderstanding and confusing or whether you purposely make up a story about hundreds of strategies out of fear of others trying to imitate your approach. What I know for sure is that you are not running hundreds of different strategies without knowing the slightest about their intent and purpose and how they behave. I have simply worked way too many years in this space to know that this cannot be the case, especially not in the deep learning domain. It makes me question your pnl claims altogether, and I get it, everyone can say whatever they want and if I don't like what you have to say then I don't have to participate in this thread. Correct, but I can equally question the claims, made, and why not. At least I gave some rational reasons of why I doubt the claims of your setup. Only in reinforcement learning would you end up with a single optimization equation (q function) that optimizes towards a certain metric without, however, knowing how you got there. In most other cases you need to a-priori define the target which the system peruse to fit towards and optimize on. That requires knowledge of the approach how an algorithm ought to behave. In RL you can potentially tweak the optimizing function and it's related constraints and rewards and penalties but you would certainly not end up with hundreds of different algorithms, each of which behaving completely differently without knowing what any of them are doing and why. Strong doubter here. Can I be wrong and potentially make a complete ass out of myself? Sure, but I estimate with a confidence level of 98 percent to be correct on this one.
Another thing, I doubt a one man band can achieve doing this, it is better suited to a team imo. Way too many balls in the air for a single trader to be able to manage.
The fact you claim to be in the field of Machine Learning but your hung on semantics of ML vs deep learning is troubling... Yes, Deep Learning is a subset of ML. Would you consider IBM to be a reputable resource on ML? Quote from their site: "As we explain in our Learn Hub article on Deep Learning, deep learning is merely a subset of machine learning." link: https://www.ibm.com/cloud/blog/ai-vs-machine-learning-vs-deep-learning-vs-neural-networks Simply google: "deep learning subset of machine learning" you will find hundreds of scholarly articles and sites making that claim. Maybe in your field you "look down" on ML and don't consider deep learning to be a subset. Thats fine, but it doesn't change the fact that for many in the space it is in fact a sub set of ML and is normal to be expressed as such. I know "a lot" about my 450 strategies. I know what matters. Common performance metrics, expectancy, profit factor, win%, draw down etc.. I know how each one performs relative to the other in different time periods. Which ones are likely to open positions in the same or opposite directions (yes some may be going long while others are going short.. but with different goals). I know what matters about these strategies. What I don't know is the pattern or the 'why' that it has discovered and uses for its decisions. I have some generalizations about which are trend following vs mean reversion, momentum based etc... some are frankly random in appearance.. but have passed strict performance criteria needed for me to trade it. As much as your attempting to draw me out, I am not going to go into details about how I pre process the data or how I actually train the models. I believe I am doing some fairly innovative work in that regard (even more so after seeing your comments) and is part of what I consider my secret sauce. As I stated from the start, it an area I will not dive in.
I am aware a lot of sources define DL as a subset of ML, hence why I edited part of my post. Strictly speaking, though, ML has a very specific meaning in the larger context of artificial intelligence. And deep learning does not share commonalities with machine learning in this stricter context. And no, I do not consider IBM an authority in this space whatsoever. Am happy to stand corrected but most anything in deep learning over the past 10 or so years did not originate from any research at IBM. In fact several papers that were coauthored by IBM researchers regarding deep learning were even rejected by some leading journals. If you are truly curious I can name you the papers in a PM. Google and FB, Amazon, and certain researchers in academia play in a completely different ballpark than IBM. Having said that, I acknowledge that there are quite a number of websites that would define dl as a sub space of ml hence why I edited my post. But it does not change my opinion on the rest of your post(s). You said "I know what matters. Common performance metrics, expectancy, profit factor, win%, draw down etc.. I know how each one performs relative to the other in different time periods." That does not corroborate with your earlier statements at all. If you understand the expectancy, draw downs and other metrics then you definitely must have an intricate knowledge of what each and every strategy does. That is not what you claimed earlier, in fact you stated the opposite which is that you have no knowledge of what most strategies do. What also makes no sense is having hundreds of strategies that seemingly do something different. There are only this many metrics how would one end up with that many strategies that seemingly all do something differently. The states that the market can be in are not even that numerous, so what additional states could a much larger number of algorithms possibly exploit? Even if you permuted over different meaningful holding periods or observation compressions of time and possible states of market dynamics you would still not end up with a count as high as your number of strategies in your repository. You were asked several times by others but could not exclude the possibility that multiple strategies might generate same directional signals at the same time as others. It lead to my getting the impression that in fact you are not really sure what is going on. Draw you out? What word or sentence gave you the slightest idea that I want to draw you out? I work in this space professionally, one thing every new hire must abide by from day one is that one only works on strategies and ideas that can be explained in very simple terms and that resonate with market dynamics and market behavior. Spurious correlation is the greatest of all sins in this space. There must be meaningful and reasonable causal relationships. Why would I be interested in your strategies or data when by your own admission you have not the slightest clue what goal and intent most of your strategies pursue? I merely showed curiosity what deep learning approach you generically take. You refused to answer which is fine. But you clearly attempted to mislead all of us with your ML mentioning. You would have given nothing away by having stated that your algorithms were derived from deep learning methodologies. Yet you intentionally used the term ML. That leads me to doubt I can believe anything else you would claim after this. I doubt and think multiple of your statements are untrustworthy. Yet in the spirit of camaraderie I nonetheless wish you success and well being.
My interpretation of what the OP was saying: If you have a black box algo (neural net or otherwise) that generates signals from data, you can back test it and get all those metrics for it. But you still have little to no idea whats going on inside the black box. How it generates the signals it does. Personally i could never trade such a system over the long run, but maybe others could. The exception would be a 'holy grail' system that hardly ever lost or had only small drawdowns. But those don't exist. Although if combine lots of mediocre systems together you could in theory build something approaching a 'holy grail' system
Before the "black box" outputs any predictions or generates signals it must be trained. The training only occurs by broadly two approaches (there are more but broadly there are two). By learning from seeing what is the correct answer and by experimentation. Only through the latter can a model be trained that makes decisions that are potentially difficult to understand. For example, through experimentation a model can be trained for a car to change lanes on a road in a proper and safe way when overtaking another car. The goal is clearly defined and through penalties and rewards a reward optimizing function can be trained to complete the task in the desired manner. But that entire algorithm describes one single model. You would never end up with hundreds of models to accomplish such task. The only time you train different models is when you apply supervised learning and through a target variable define toward which metric the model should optimize towards. This target can only be properly derived through an algorithm or mapping function that reflects a very specific intent and approach as to how the fit model should optimally behave. That requires knowledge a-priori of the specific strategy that the researcher intends to equip the model with. He/she does not know how it will accomplish it but he/she prescribes from the very beginning the end goal of the to be trained model.
Wow... if this is the type of "rigid' thinking that goes into "professional" shops. I may have under estimated another edge my system has.
Most spurious correlation and nonsensical causal relationships evaporate over time into thin air. Most importantly, you as quant trader will have nothing to have faith in during times of drawdowns. You mentioned time and again to trust in the system. When the system, however, is built on algorithms that exploit non understood relationship then there is nothing to base any trust in such system. That's at least the hard lessons I learned from many years in this space. Anyway, wish you nonetheless good luck and more importantly a good learning experience in your endeavor.
The way I view it is like this. Price moves towards the previous day low, so you have a trigger to go long. But perhaps the bounce is met with selling, so yet another system based on order flow says to short. Now maybe the previous day low breaks, so your swing trade system scales in, since its designed to go through a drawdown. At the same time, the poke lower hits a channel bottom, so yet another reason to add to the long. This happens to me all day long where different trades set up for different reasons, but since I'm only entering positions in one account, I have to choose an ultimate direction. But its absolutely understandable why you would want to be both long and short at the same time for different reasons and both can end up being winners. To be honest, its all pretty clever, and smoothing out a PnL curve with more strategies I think is a good way forward. Imagine if all of a sudden, all hell breaks lose, and while some traders might be stunned and wondering if they should close long positions, you know that your system is already short for some trades so you're covered either way. Isn't this also how some people structure their options plays? You know its gonna make a big move, but you don't know which direction, and you set it up in such a way that you make money either way. The only way to lose would be for price to not move very much. But rally or plunge will both end up profitable.
Since you don't use stop loss i am going to assume that all strategies are trained to take profit only. I think range expantion is your weak spot, this is what happened in January and ended in a drowdown.Mentioned win rate is 67% from that i am interpreting that in January losing trades were larger than winners and some of them may even come from trades initiated in late December. I would run backtest of all my strategies used in December and January across data going back as far as i could,i know you say deep learning trains your strategies,but i don't think you could come up with more than 450 strategies that concentrate on exiting trade with the profit. You use deep learning to stay emotionally detached and of course with no stop loss capitalization is of great importance,well thought out. I am curious if that backtest of same strategies would show any 2 months period ending in a drawdown and how often this happened and actual numbers in % terms.If each year was profitable and thats the case then it holds true that its impossible to come up with more than 450 strategies that concentrate on taking profit only,more would be the same as they are going to show increased correlation. It is interesting that you chose 21 days and you trade futures and day has 24 hours and day trades are closed in 24 hours.i don't think you train the system on minute bars,more likely volume or tick bars are used as data input, i would say 450 strategies come from less than 5% of best scenarios.Most of them must repeat themselves over and over again logic wise-by this i mean from month to next month.Do this and that does not change,targets change all the time even if by small amounts. In system development most strategies work until they don't.It requires deeper look into why January ended with a loss.Closed out positions with a profit and subsequent new profitable trades did not compensate for those trades that ended up closed in 21 days with a loss.They were there at all times right ,until time came to close them with a loss. BTW Congratulations nice going and thanks for opening this thread