>I'm still trying to understand how one would utilize neural
>nets in trading real money. Is this basically a black box out of
>which comes signals that may give a 55% edge?
You would use a neural net as you would any other trading system. It is up to the individual to integrate the nets into their style of trading. For example, if you are a fan of the stochastics inidicator and you have an existing system, you could create a net that would forecast the stochastics indicator a few days into the future, so as to allow you earlier entry into the market. Or if you are building a system from scratch you could predict the % change in close a few days into the future. Then if the net predicts a 2 or 3 percent change, you buy and then sell after the gain. If it doesn't happen it is still up to the individual to use proper money management rules and exit the trade. The trick to using a net, like any other trading system, is to follow the system through thick and thin. If you deviate from your system you are giving up that 55%+ edge. Remember 55% means 11 out of every 20 trades is successful. That may not seem like much, but it is profitable if you apply money management rules. And it is not uncommon to see systems that trade correct 65-70 percent of the time, so 55% is a minimum...
>If so how do you integrate other forms of TA that might
>give conflicting recommendations though additive to the
>ultimate outcome? Of course psychologically how do you
>jump off a cliff if the system says to? If you selectively
>take the signals that are liked, results must suffer.
That is up to the traders discretion. No one system is right all of the time. I'd say go with the majority. And like I said, you can integrate nets into a working, existing system, and will probably achieve better results than by using each system by itself.
>How do you know when the NN no longer works without
>suffering an extreme drawdown, or if the drawdown is
>to be expected?
This is a problem inherent in all trading systems. The market is in a constant state of adjustment. A system that works one day will cease working the next because too many people have discovered it and are now using it or external factors such as political events come into play. No system will work forever. Period.
However, I think the neural nets are the best form of trading systems in that they can adjust to factors that influence the markets. For example, if you feed a neural net a diverse range of indicators and decent trading history (3 to 5 year) it should be able to pick up patterns for both a bull and bear market. A robust net would then be able to compare what is happening in todays market with what has happened historically and make a forecast. Now eventually the patterns will change to the point at which the net can't make a good prediction.
You can however run a correlation between predicted and actual output. What this means is that you are comparing the predicted output and the actual output to see how well they mimick each other. If the net scores low on the correleation, then you may want to stop trading the system as the net can't make a good prediction. If the two correlate well then the net is making good predictions. You can also set up another net the forecasts correleation into the future. This could help you 'see in advance' whether the system will do well. It is also important to note that most NN systems won't necessarily give you the correct range of the move, but they will give you the right direction.
Again though, proper money management rules should be applied to limit the effect of any unforeseen drawdowns.
Optim |
