I think what will happen is your selected NNs will be trained to pick up on outliers, i.e., those trades that make tons of profit and happen to coincide with whatever inputs you have. So you will be fitting to those very rare, coincidental trades that made money, i.e., so rare that they will likely be singletons in history. It really depends on how tight your selection criteria is, I guess. Essentially what you are doing is clustering your trades based on inputs and profit/loss. The tigher your clustering criteria, the more likely you will get singletons, and hence they will not be generalizable. I think you would need to decide on your cluster distance criteria based on some kind of relative information content of your clusters for them to be useful.
Auto Trading Idea
- Thread starter frostengine
- Start date
This makes sense....and of course leaves you with a really small sample size of backtested trades. And this is not good in forward tests which may be live (real money).
I think you've got to optimize the number of trade opportunities with a cut-off at the low end. Too many trades = high commission cost; Too few trades = less reliable results. Always look at the average time-in-trade as well as the average time-between-trades statistic.
This might explain why HFT has proven to be so successful.
Quote from RusselG:
These new updates are great -the daily live account statement is just saving my strategy! I am traveling at the moment and it is a bit hard to be in front of the computer or on my smartphone all the time
It is all over the forums that Zulutrade changed the rules for the providers - now they finally will be penalized when they are loosing our money
Providers watch out the big Z's watching you Quote from MillieJ:
It is all over the forums that Zulutrade changed the rules for the providers - now they finally will be penalized when they are loosing our money
Ohh yeah? Really, where did u read this? ..I was away some time and did not managed to follow anything...
Frost, your goal is to design robots with highest amount of profit in relation to risk.
No serious funds/ctas/firms will want to trade such "strategy".
why? they all want to make as many trades as possible and churn their investors to death.
the idea of having 10 or more individual strategies per market is of course very good and logical.
if each strategy protects the downside risk by set of intelligent exits, the whole portfolio can generate 100% winning months.
but as i said, nobody would want to trade it.
if you have your own money to trade 10 systems/market on 5 markets - move on with the idea.
No serious funds/ctas/firms will want to trade such "strategy".
why? they all want to make as many trades as possible and churn their investors to death.
the idea of having 10 or more individual strategies per market is of course very good and logical.
if each strategy protects the downside risk by set of intelligent exits, the whole portfolio can generate 100% winning months.
but as i said, nobody would want to trade it.
if you have your own money to trade 10 systems/market on 5 markets - move on with the idea.
Ed, since you were looking at this back before decimalization I'm assuming you'd be willing to talk details. What were your inputs into your neural nets back then?
ha. ha.
Paul, is that you?
If it is then you know I (and CoolTrade) are light years ahead of where I was in 2001.
Most of the inputs were the eminis and the data on both sides of the Level II screens. But I don't even bother with the Level II information anymore and my strategies, now, are way better than they were back then.
Cheers!
Ed