Attached a little document dating from 1994 (if I am not mistaken 486 were the strongest PC's then)
Could post some more similar things, but I think there is no need

To clarify : I am not a NN skeptic, but seeing the magic the simple NN mentioning does to people makes me wonder. Right now it is more like a religion than serious investigation. This document is just a simple voice from someone that really did investigate and worked on NN. Might be worth reading (no math, no complicated explanations, no super-duper high end, clever only stuff).

Attached Files

neuralpitfalls.pdf (230.2 KB, 185 views)

Brings to mind the story of how in the 80-90's the US military decided NN's were the biz, so decided to train NN's to detect camouflaged tanks in a forest. After training, they worked spectacularily, flawlessly detecting the tanks in a large series of similar photograph. But just to be on the safe-side, they went out and took more photograph of tanks in forests and rechecked the NN - only to find they failed completely! Turned out that all the NN's were managing to detect was the light levels that existed on the days the photos were taken...

Moral of the story is that although some of the best and brightest were sure how the NN's worked, in reality they had no idea, but the initial results fitted nicely with what they were hoping to see! Sure sounds like 'curve-fitting' to me, and nothing I have read about NN's since makes me think anything has really changed since then...

Hi Zupcon,

Your way of thinking seems to be the logic itself. But it is not the correct way of analyzing this.
Someone in the webinar made also the remark and I descided not to react on it because it would bring us way to far from the topic and way to technical.

If you want to analize this then split the trades up in to 8 catogories.
Yep eight!
In the clear uptrend that you can spot with the eye ( between a low and an high)
the winning long trade.
the winning short trade
the losing long trade
the losing short trade

Then spot with your eye the down trend ( between a high and a low) again the:
the winning long trade.
the winning short trade
the losing long trade
the losing short trade

You will spot emmidiatly 2 remarkable facts: the winning long trades get smaller and the losing long trades gets bigger.
Depending on how good the system works or how much is it was curve fitted
you will see a shift of the size of the other winners and losers.
The hitrate will not shift that much but the size of the winners and losers will shift this way that the overall result turns into an overall loss.

Very important for a system develloper is to look that there is as less corelation possible between the prize en the equity curve.
That is even so important that there are formulas for this to calculate this coorelation.
Why ?...because professional traders can show to their customers or clients that they have a system that has no correlation with the market that they trade. Lets give an example: If a client is trading the stock markets and I have a system that is trading on the futures of these stock markets. If my equitycurve is correlated to that particular market that means if that stock index goes down that the client who is already trading that stock market will loos money and with my system he will again loos money...so 2x times losing.
If I can show him that if that particular market goes down and my equity curve goes up or that there is no correlation between my results and that stock index then he knows that he is well protected.

I don't know if you red already some postings of me where I explain that each system can be placed on a spectrum. For ex. the trend following system that have +/- 37% hitrate will have a winnig trade that is 2x times a big as a loosing trade.
The systems that are on the other side of the spectrum are the systems that have a 65% hitrate and will automaticly have losses that are 2x times as big as the avg winners ( no way around this). One can not build a system that has the best of both worlds...no way).
Well on that same spectrum we can place on one extreem side of that spectrum the systems that are over-optimized or over-curve-fitted (like the "better EA)
They will perform extreemly well in certain market conditions and do very bad in other market conditions in this way that one will see a very strong correlation between the prize and the equity curve itself. Or see an extreem inverted correlation.
On the other side of the spectrum we can place the systems that have a BAH strategy (buy and hold). That means that the strategy would be buy and don't do anything anymore. That is the other side of the spectrum because there the rules are as good as zero and no optimization at all or backtesting took place. The result of that strategy is that one wil see that equity curve and price will show again an extreem correlation.

I'm going to end this posting with some sort of a cliche....Learn to trade what you see today and don't depend on a EA that you 'teached' how to trade based on the history that maybe will not repeat itself.

Or an other some sort of a wisdom....till 10 years ago there was no such thing as EA's. But still there were over the last 100 years many famous traders that were very succesfull in trading. Bring those people back today and they will be as succesfull again (because they know how to trade).
Place the people from today with their EA's and their dual core laptops back in history and they will fail just the same.


regards...iGoR

PS. Don't get me wrong, there are people or institutes that are using systems.
But they known that they can have drawdowns that last for many months even years the same as certain indexes or markets do ( stock markets were down between 2000 and 2003) but they accept this because market conditions can change.
If people here on this forum would receive a good working EA that would have lost money over the last 3 months they would say that the EA is just plain crap.

__________________
Success comes with knowledge. Knowledge comes with experience. Experience comes with time and hard work... (iGoR)
MT Intelligence - Stats for fxigor1

Hi Igor,

i just saw the clip....great stuff... also if you have an EA that making 130K from 10K ....so its ok if he is loosin 5% the next few month,,, :-)

all the best...you are ok!

__________________
VISIT US AT:
http://WWW.MT4free.com