TSG you do that all by yourself? goodness, i knew we were lazy, but that is ... something. we stopped our pairs trading late 2004. edge gone. a horizontal line for more than a year. pity. was good stuff. and much work ... (for the lazy us ...) ... my perspective on human pattern recognition is similar. but my conclusion is different. i stick to systematic trading. but i try to use the human intelligence to decide what and how to develop. brute force backtesting does not do it any longer with the machines being so fast that they constantly pop out statistical fluke systems. in the nineties you needed quite an idea before you did a backtest. machines were not fast enough then. now they are. we have 64 processor cluster that is a beast in this respect. if you just take random numbers on daily data, you will have 3 out of 100 random systems with a sharpe above 1. frightening. so we need better understanding of what to test. means we need better choices of subjects to begin with. our developers all produce a short forecast of the sp future for every day before the opening. thus, with all their other work, they at least once a day focus their attention on current price action. so we are training the choicemaking process so to speak. one other thing is that matlab can compare images and calculate theirs similarities. i want to use that for throwing a blurred photoshop file of a certain price movement into matlab and have it compared with recent action. whenever the similarity is matched sufficiently ... bom. so i use my own recognition and just have to transfer it on a piece of paper ...
Hi Richard, I think I can answer your question in more detail. We use TWSLink ( http://www.trade-commander.org/twslink/twslink.htm ) for the connection to IB. We have just started with this project some days ago, so far the connection to IB works (receive market data, order placement, ...), but we are not at the point of testing systems yet. I'm pretty sure that for real high frequency trading Matlab will be too slow. But for now we are going to test systems based on 1min bars, so this should be no major problem yet. Do you use Matlab and or TWSLink for automated trading, can you share your experiences with us?
Flip, Thanks for your info. I know of TWSLink; I tried it some time ago but it had a limitation that may be gone at this time (the author had assured me of that). I'm a heavy Matlab user, but alas not for trading, so I probably can't help you there. It would be nice to use Matlab with sub-minute bars. I may try TWSLink again to see if it's possible when I get some time. Richard
Hrmmm.... Well, I guess I fancied myself by writing AI. It's only my discretionary brain. There's a few things: 1. My system development process has changed from the time I fully automated my models to now. For any system trader/developer, we hit the problem with robustness. No one knows and will know whether the system will work in the "future". What we do know is whether the system has worked in the past, so we backtest to confirm our ideas along with setting the quantitative contingencies for risk management purposes. Another thing we can do is, to decide whether the system is working "currently". Just like any outright discretionary trader, you trade what is happening in the market right now. I simply do that with the systems and models I have. Still, it's a bit tricky to do decide whether the systems/models are supported by the current market. I follow the market just like any discretionary trader, but instead of following whether the market is strong/weak in which ever direction, I view it from a tendency standpoint. To do so, I need to understand my systems very well. So... as part of my system selection (implementation) criteria, I added my own intelligence. I place full risks only with the systems that I am able to manage. I've tried trading a model that fully relies on Neural Net / Walk-forward Optimization but I tend to lose control over those models because I am not able to monitor what is going on inside the Neural AIs. Anyways, I trade systems that I understand and can see in the market. 2. So... I only trade systems that I understand well. Simply, I need to understand: a. Why the system works? b. How the system makes money? CLARITY of why I am trading a system, (or taking a position) is what I am looking for here. That is why we run tests and TEST... TEST... and TEST. Running tests is not about developing a profitable system. It's about understanding the market, a profitable system should come from the understanding you get from the research. Anyways, it's not "exactly" the case but as an example, for my mid-long terms systems, I need to be able to quick view the market (eg. charts) and say, "This system should be long." I may not know ALL the aspects of the reason why the system works but I have a very good idea about it. Once you have some idea about the systems, it's a simple progression to understand, what aspects of the markets the system is vulnerable towards. 3. So I monitor... Interestingly, regardless of news or other factors, markets are always changing. So I understand how the systems work and follow whether the market supports it. As a simple example (not "exactly" what I do but to give you an idea): Let's say I have a short-term trading model. The system is a single leg delta neutral arbitrage between a cash basket and ES. The model relies on the order flow of the market to be "normal" statistically. I wouldn't include a stock in the cash basket this model when there's an earnings report or during Fed meetings as a whole because the volatility is not "normal" statistically. Also, I would reduce the trading for the system if the execution efficiency gets skewed. I would be following the Bid/Ask price/volume, along with slippage/liquidity to see if the order flow is changing. As mentioned, the system relies on a specific tendency of the order flow. If the tendency changed for any reason, I would reduce the weight of the risk. 4. Agent-based. I have different programs working as a whole. I am only part of it. I delegate myself as part of the system.
I think we are doing something very similar. Just a bit differently. One note is, I'm talking from an experienced trader's perspective. I don't recommend newbies doing what I do. Newbies should be going out picking up mathematical and computational finance skills, even more... trading, trading and trading. and... TEST EVERYTHING
my bet is that i can run many more systems parallel, that i can actively monitor. and that the robustness coming out of that outweighs the lack of individual monitoring of each of them. and, sure, by "monitoring" i mean your way of deciding whether to trade a model at any one point of time. nevertheless we are playing in two leagues below you. we are not even close to basket trading at the moment. yet it is not where i am heading to. i think this game is already horribly efficient (but then, which is not ..). i read once that in the early 70s you needed bout 80 stocks to replicate the sp500 with a given tracking error, while it took you near to 400 to achieve the same error in the 90s. anyways, that is not my game. how many stockPairs do you trade? with what frequency? if that is not giving out too much ...
I think we all agree that system traders don't give out the details. Giving out a few hints though... key to the matter is not about duplicating the index with your basket. It's about setting up the basket so that it's arb.-able. Make better sense?