Frost, What criteria do you use to add/delete a network? When you forward test your networks do they all perform decent, or do you weed out a large percentage? jet
The idea presented at the start of this thread sounds analogous to the 'AdaBoost' algorithm, where by weak classifiers (a NN for instance) are combined together to form a stronger classifier. The data is re-weighted so that each classifier in the ensemble concentrates on a particular aspect of the data.
Here's my experience with a similar approach. As soon as I forward test a significant amount of time (e.g. 1 month for day trades or 6 month for swing trades), the performance of the neural network or curve fitted algo significantly suffers. It may have been good on a small walk forward test but As soon as I increased the timespan it changed. How do you construct your testing right now, how much input time, how much walk forward time ? What happens if you segment your walk forward time and train the algo again, does it stay profitable? e.g. Week 1-4 training input, W5 is walk forward testing. Algo is still highly profitable, then you go with W2-5 as training input/curve fitting and W6 for walk forward, does it stay profitable? I never got lucky with that approach Follow my trading progress @ http://www.elitetrader.com/vb/showthread.php?s=&threadid=200655
To add a network it has to: 1. Be profitable on its own 2. Have low risk characteristics 3. Add profit to the overall system without increasing risk much Yes I weed out most of them. Only a few make it
Looking forward to this week. Hopefully the strategy performs well. I have been looking very closely at the entries of the strategy and am a bit concerned. Many trades have fairly poor entries. A lot of heat is taken on most trades. I have tried adding various filters to cut down on the heat, but hurts profitability a bit. I need to come up with some ideas on how to reduce risk further
This is like traveling back in time 10-15+ years. I watched forums , magazines and other places as people tuned, tweaked and massaged systems of Neural Nets. After a lot of effort, none of them ever really seemed to make money in the real market for long periods...
Strategy made 2 trades today. 1 long for a loss and 1 short for a gain. Overall the strategy made +$35 for the day
Actually around my computer so I can see the trade today. Bot is short from 9832.... Is there any interest in these sorts of trade by trade updates? I have an unused testing account on collective2 I can convert to a real account and show for free if there is any interest?
Nice job !!! I'd look @ them on collective2 if you'll post them follow my progress @ http://www.elitetrader.com/vb/showthread.php?s=&threadid=200655