Results for the week: Swing Trades: +$8,652 Daytrades: +$3,898 Total: $12,550 Pretty routine week... All trades this week were posted live on Twitter. Hopefully seeing an automated system posting live for the week was helpful to you in some way. I don't see much value in continuing to post all my trades - so will turn Twitter back off.
Congrats, nice profits! But after you posted on Thursday 'I am sitting on over $30k in unrealized gains from the swing trades opened earlier this week', I was expecting your Swing Trading profits to be much bigger, especially following Fridays continued bullish action..
The unrealized swing trade profits is sitting at roughly $34k at the moment. The bot took some off the table today, but vast majority is still unrealized. The numbers given for the week are only the realized profits.
Thanks for posting this thread. I think the mistake that most people make with machine learning trading systems is trying to model returns over some period t. I get the sense that you made the critical insight that your base classifier simply predicts long or short with some probability, and then you can connect that probability to the backend system on the base time fractal. As you mentioned, managing the trade is not about stop losses or profit targets; it's about the ML system giving you a confidence level at every time point. The most important factor in trading ML models is riding the calibration curve, e.g., if my evaluation metric is accuracy, and the accuracy shows only a small edge, that's still okay if the calibration curve is solid -- you can still be selective with probability thresholds at the tails of the curve. Put all of these together with an ensemble of systems (as you've achieved), and you're good to go. By the way, there is an interview with Jim Simons somewhere on YouTube (he's done a lot of talks) where he actually said "it's just machine learning". So apparently, you have your own microcosm of systems that accomplish a similar objective. Congratulations! Regards, PTR P.S. If you want to play around with some of these concepts, you can check out my Github repo: https://github.com/ScottfreeLLC/AlphaPy
That's fair and understandable you don't want to give up the specifics of your edge regarding ML. Are you willing to touch and further expand on your thoughts regarding your other two major edges, being both position sizing and time? I think most people agree and understand most intra-day traders lose money. If I had to guess I would say a big part of that is that they over size and generally cannot use time as an edge. What are larger players using against intra-day traders? Being able to push the market to extremes that doesn't effect their end goal, but where intra-day traders can't hold either for literal margin reasons or because the loss would be too extreme for their strategy or account size. If that doesn't work they also have the ability to hold price over time and make a move after hours or before market opens(which often intra-day traders want to be flat before close). (A side note would be that Time is actually an edge that longer term "news" players benefit from as well, where they mistakenly think that their news research is their biggest edge, I am sure it is in some cases, but for the majority time is their true edge). Just doing some simple math let's say you're long 1 MES and it drops to 0, you would lose $22767.50. 1 MNQ drops to 0, you would lose $28870.50. So, just doing some rough numbers you could be in 28 total contracts and have them all go to 0 and would still have roughly $77k left in your account. So, seems like you're essentially trading on no leverage? Which allows you to put forth a very wide ranging amount of strategies, since you are not contained and forced into an arbitrary stop loss due to position sizing. I also noticed though that your worst swing results, corresponded with the large down move we had in January. Just wondering your over all thoughts on that? Do you personally believe your edges rely more on position sizing and time, or in your opinion your biggest edge is the ML? Thanks if you take the time to expand on these other two edges. Maybe the point of the thread and everyone else just wants to focus on the ML. But I would think to be successful in the ML you would need to understand the other edges as well, as to not have a false sense of security or rely too much purely on the ML, without understanding what the ML needs to obtain results.
When I look at my system, I believe the top contributors to my edge are: #1 The dataset/data manipulation #2 The unique way I am training my strategies #3 Strategy diversification (trading 450 different strategies). Creating a smoother equity curve which helps with #5. #4 Position sizing. Ability to withstand drawdowns without any psychological impact or desire to "stop the system". This ends a lot of would be automated traders. They lose faith in their system after hitting their first draw down period. A lot of my earlier attempts at automated trading ended due to this as well. #5 Strong belief that my system is "solid". This allows me to give the system the "time" it needs to be right. What you identified as an edge in position sizing really affects #3, #4 and #5 in my list. If you are over leveraged you can't take advantage of those 3. With that said, I don't consider position sizing more important than #1 or #2. Without getting those right, I could have a losing system and while I could survive a LONG time due to position sizing, it wouldn't be fruitful in the long run.
Since all those trades weren't marked as being day trading or position trading, it made it difficult to really get a sense of what was going on. It seemed like trades were being opened up at random times. I get that they were different strategies, but sometimes at the same time that a position would close, another would open at the same price, and sometimes it was in the same direction, and sometimes in the opposite direction. So this leads me to believe that with enough trades open, a bunch end up being winners. Sometimes the day trades are in total opposite to the swing trades, but the idea, as has been pointed out via the mean reversion question, they eventually come back. Of course I'm sure there is an exit point, so there is risk management in there, but seeing short taken in on market as another market you're long in just screams a bit of randomness to me. When I'm sitting here watching the market highly bullish, I was thinking why you had any short positions still open. So its almost like you take some shorts, and hope for the mean reversion, but you've got some longs as a swing, to cover the day trade shorts just in case they bust. I don't know what went into designing these systems, but I get the sense that any one system on its own isn't really that good, but as a whole, they back up for eachother nicely. I'm probably highly uninformed about all of this, but as I watched the trades in real time, some just didn't make sense at all based on what was happening at the time. I of course can't argue with your final PnL, and that is the only metric that really matters, but I almost think that your systems aren't really that smart, and the biggest edge is lots of trades with the key being to keep the winners going, so risk management in a way.
Why so many strategies? How do you tell them apart and how do you know that you won't get many identical signals that would increase open position size one directionally way above the normal threshold? What kind of edge would each of those strategies even pursue? In 30 years I have probably not come across more than potentially 8 to 10 edges I can clearly define in a single asset, non-spread, futures product. I assume some strategies use different asset pricing series as features in your ML data sets than others. Perhaps that is what makes different strategies.
I don't think so. Some strategies take advantage of defined trends and momentum, with the assumption that such momentum holds post entry. Mean reversion would get you into the exact opposing position.