Mike: Thanks for your input. I know I am fitting a curve. Read back to the first post of this thread. That was the point of this exercise. The exercise is to do something that most (yourself included) thinks is a bad idea. Then see the outcome. I like your analogy using monkeys however its way off base. Perhaps the vagueness of the selection process and training of the neural networks as presented in the thread leads you to that conclusion. However, its far from the truth.
Fair enough - thanks for not taking my post too strongly, I re-read my post and realized I sound kinda harsh, sorry about that. What you're doing is an idea I'm familiar with, however, I'm form the old school. A good trading idea needs to exploit something fundamental. If one can't explain that fundamental something in one sentence, I don't buy it That said, I know more than I should about neural nets, I spent too much time with them in grad, school... the critical component, *always*, is the input. Garbage in = garbage out. What might be more significant is all ES/YM data from 1997-present. It appears as if you're using 2007 forward and there is lots of vola during that time...
I am using Esignal and is all the historical data its given me. Over the years I have bought tons of data. Through computer crashes etc I have managed to lose all of what I once had.
Statistically, linearly combining independent random variables (which is what you're doing here) doesn't increase the expected value. (It *does* reduce the variance.) In other words, if your neural networks do not have positive expectancy, then combining a large number of them doesn't help.
Correct. Every neural network is profitable on its own. The problem is to have a good neural network it needs to be pretty specific. Therefore it wont trade much.. What i'm doing in this exercise is taking the best neural networks and adding them to a larger strategy. This way you can have a lot of highly specific neural networks trading and produce enough trades/opportunities to be worth while. No one wants to run a single strategy that only produces 3k profit over 3 years with huge gaps of no activity... but you combine 50 of those together and now you have something that trades often and produces a good result. That is the basis behind this whole exercise
No, the fact that you've curve-fit them does *not* mean that the neural networks are "profitable on its own". Not at all. No, combining 50 of them doesn't improve the (expected) result. It only increases the number of expected trades, and decreases the variance.
Your missing the point.. YES they are all profitable on their own. I do not add any neural network to the system that isn't. Read the first page of this thread.
No trades taken today. Although should have been a trade for a $90 loss. I still do not have NinjaTrader configured properly to actually trade. The print out today showed the strategy never saw any data bars. Hopefully I have this issue worked out for tomorrows trading.
Frosty, back from your sabbatical? I miss your journal !!! Follow my trading progress @ http://www.elitetrader.com/vb/showthread.php?s=&threadid=200655
I am up to 46 neural networks, here are the current key stats: P&L: +$137,000 (1 contract only) PF: 3.52 Sharp: 1.09 Win %: 63.8 Average Trade: +$175 Trades: 779 DD: -$1145 I removed several neural networks and of course added several new ones.