If you have several indicators, you may construct several systems (note, each system should be profitable alone). Each system is totally independent, if it has long signal you go long 1 contract. If the other system has short signal, you short 1 contract and you have zero total position. But this will reduce risks in comparison when you trade each single system with 2 contracts. You may combine two systems into one, but net profit will be lower.
Actually "net profit" could be higher if the 2 systems are required to be in agreement, and thus eliminating more losing trades than winning trades. Just like driving a car, the more rules you follow, the less accidents you have. Stopping at red lights but running stop signs is not better than than stopping at both. Ed
I disagree with you. Signal from two systems will be rare, that's why the net profit will be _definitely_ lower. The question is about risks, how would it influence on NetProfit/DD? I do not have an answer, it needs to be simulated, may be using monte-carlo, don't have time. I run several systems simultaneously and it reduces my risks. Also when you add more and more rules to your strategy you are at the risk of over-fitting.
I agree that relying on multiple signals to confirm the entry/exit of a trade when dealing with NNs is a losing battle. You will barely have any trades and the ones you do have I have not found to be better. Having multiple independent strategies does reduce your risk. In essence that is what I am doing here.. each NN is its own independent strategy all racing to see who will generate a buy or sell signal for the "overall" strategy.
I don't understand, how do you finally trade? If system 1 has long signal, will you go long 1 contract? If system 2 has long signal will you add and have 2 longs? If you run 50 different NNs running together, does it mean that one day you may be long/short 50 contracts? (I run my systems in that way)
No, I run it on a first come first serve basis. All the long neural networks are wrapped into position != long { } then all short neural networks are wrapped into position!= short {} That way if the system is already long, then all the long neural networks sit on the side lines. If a short signal happens however, then the short neural network takes over and the position goes from long to short. Even with nearly 30 neural networks so far in the system, I rarely have 2 long neural networks overlapping each other. This is by design. I do not need 2 neural networks telling me the same thing. I want them both telling me a different story
Then when do you become flat? If you consider data since 06/2009 as Out Of Sample, there was strong up-trend. You have only long signals, is it fitted? Waiting for information about short systems...
The last strategy statistics given on 06-20-10 contained both long and short neural networks. The short neural networks contributed just shy of 1/3 of the PL for the out of sample from 6-09 till now back test. The system is using set exit points. Earlier in this thread I mentioned using x bars for an exit. That is how the nets are trained for easy comparison. The exits used in the strategy here is a set stop and set profit target entered when the trade is entered. Exit happens if either the pt or stop is hit or the end of the day arrives.
This concept/trading idea is a risk management clusterf**k waiting to happen. 1. You have no idea as to what this strategy's risk truly is, both in terms market dynamic and worst case scenario. 2. The combining of "systems" further obscures your true risk since you do not know the statistical historical correlation of N-systems. 3. No effort can be made to quantify or reduce any type of abnormal market risk because the "combining" of these systems is essentially one big selection process that uses an unknown market dynamic to guide the selection process. You're fitting a curve fit... 4. Out of sample tests are not indicative of anything in this case because your NN will only have forecasting capability based on your training set. Since your rules are essentially "blind" rules with no "proof of concept" they may as well be random rules. What you've essentially done is put 100 monkeys with an affinity for keyboards in front of trading desks and asked them to hit buy/sell buttons a few hundred times. You've then taken the "profitable" monkeys and placed your bets on them... Christ... just go study a real edge instead of wasting your time with monkeys.