He seems to have a combination of both, mean reversion and trending. With 450 different strategies he must surely has both styles covered
This might actually be the edge. Someone else smart early on said he thought the edge was a large account, and that may very well be true because it allows for all these strategies. The way I think about it is that mean reversion works easily 50% of the time. How much does the market spend trending after all? It goes from range, to breakout, to trend, and back to range. The breakouts are often fake. So you make a good bit of cash in via mean version. When the breakout if real and leads to a trend, you've already got some trades on hoping for that trend. They may have been losing before, but now they are winning. As long as the mean reversion losses are controlled, you're set.
I see there has been quite a few new posts. I'll attempt to address in one response: #1 It's rare in ML to really understand what the "black box" is doing. It's complex. In my mind, the point of machine learning is to identify patterns that I can't see myself. If I was able to "see" those patterns, I would argue the ML didn't do its job well enough. #2 Concerning Twitter. I gave it a week to provide some insight into what a production/profitable system looks like. Others questioned "why" even expose myself that way. While every trade was posted, I did make it purposely hard to reconcile by not providing a unique identifier on each trade. Even with 450 strategies operating, I do have a level of paranoia around people reverse engineering my system. Several of you made pretty good observations even with it being hard to reconcile. I have no doubt that with a unique identifier on the trades, some of you would be able to figure out more than I wanted to disclose. With that said, I did question what the continued value of posting to Twitter would be. Both for myself as well as the community. If there is enough value, I don't mind leaving it on. But I need to be given a reason... #3 I believe in an earlier post, I mentioned that from what I can tell some of the strategies are trend following, some are mean reversion, and some frankly look like randomness to me. I listed having 450 strategies as one of my top edges. With this many strategies my equity curve is much smoother than with just a handful. The market is constantly changing... the more strategies, the more likely I have a subset that is designed to work well in that situation. When the situation changes again, another subset will do well while others may just tread water for a while. It's also the reason I am constantly adding more strategies to the mix. When I opened this thread, I said I was now looking to add some strategies based on individual names vs the index futures. The first one I was training was AAPL - which completed this week and will now be running live starting tomorrow (28 strategies specifically for AAPL). My goal for adding new individual names is to further diversify. While having 450 strategies devoted to index futures is great, I need to venture out to individual names to further smooth my equity curve and avoid losing months like January. Ultimately, I want to eliminate all losing months.. then all losing weeks.. then all losing days. Thats going to be a tall order, but with enough strategies, I think it's feasible. An area I am still debating is what will be the best way to trade AAPL. Using shares or options. I am going to start with options. The AAPL strategies follow the daytrading process so they will be flat by end of day. Therefore, I will start with using weeklies ATM options. I have some concern that the drag from trading options will prove too much and i'll need to switch to trading the shares directly. I think focusing on these newer strategies and the challenges with the best way to execute the trades may be a good direction to take this thread going forward. The next individual name I am training is SBUX which will be ready for next week. At which point I am open to ideas on what I should train on next. The criteria I am looking for are names that have tight spreads, lots of volume, and help me to "diversify" a bit. Which is why I chose SBUX next instead of say MSFT. (I will ultimately add MSFT to the mix as well) Suggestions on a stock that meets that criteria?
OP, I made 150% over 3 years ending Dec 2021 and loads here found it hard to believe. So I believe you for the simple reason that I Layman question: why does a self learning system need training?
That is what some people are saying, not everyone just to be clear. I already stated I believe just about everything in his post. If you add up all the factors objective and discretionary, plus review some charts. His PnL makes sense, how he is making the PnL makes sense and the fact that he is bored makes sense. The only reason I keep harping on disagreeing with ML being the edge, is because the thread is based on ML. It's important not only for the OP, but for others who may be looking to follow in his path to truly understand all the edges of how the profit is really being generated. I've seen too many times where a profitable trader post good results somewhere and looking to help others, when the profitable trader themselves doesn't even fully understand where their edge is coming from. Why is that dangerous? Because the new trader or trader's that are looking to him for help are now going to be taught things and follow things that they genuinely believe is the reason the other person is making money, when it can be for a totally different set of factors, sometimes not even directly related to trading at all.
I don't disagree that being sufficiently capitalized for your style of trading is not important. It certainly is. For all the reasons already discussed such as not panicking over draw downs, to being able to diversify your strategies to be profitable under more conditions. However, there are tons of traders out their who are well capitalized. Most significantly more than I am. That doesn't produce results by itself. You need to have uncorrelated strategies with a significant positive expectancy capable of not only exceeding transaction costs but also produce "tradable" equity curves. Meaning drawdowns are manageable.
It's not self-learning. Very few ai driven approaches learn as they process new data in the arena of unsupervised learning. Most ML approaches are of a supervised learning nature, meaning, they require data to be trained on in order to fit a model.
Ok that is very fair statement and we're pretty much in agreement. It's quite possible that your ML strategies are a bigger factor in your results than I personally believe. Certainly don't have enough data to disprove that. Just simply wanted people to clearly see every edge that is being utilized, in case they decide to pursue ML based on your post. That seems to be accomplished. Was not trying to flood your thread or be so aggressive, just felt it was important for all parties to have a back and forth and make sure it was abundantly clear to everyone. Thank you for the thought out replies. Good luck in your trading, will continue to check in on the thread.
Would you mind elaborating on what ML approaches you take that generate models you know little to nothing about? I work in this space and that notion does not compute. Very few ai approaches work like this, foremost reinforcement learning approaches. Almost all supervised learning models require the clear definition of targets/labels that imply that you define precisely the strategy the model is to optimize on. Am not asking what data you use, just very generically which ML learning techniques you use because so far it makes little sense to me.
I know almost nothing about Machine Learning. Except watching a few short videos on youtube on ML for chess and image processing. But when someone says they are using ML for trading. It conjures up images in my mind of trading data (recent bars, volume etc) being fed into a neural network and a prediction comes out with probabilities for either long, short, or flat. So i was presuming he doesn't understand sometimes because a complex neural network is involved in making the predictions for some of his strategies?