Overly Complex Coding and Trading Systems.

Hello Jure

Thanks for your contribution. I hope you continue sounds like you've experimented. That's exactly what we need. Which neural network software have you use. I am currently using neurosolution's I find it takes nominalized data from a standard division formula best.

I also run tests for uncorrelated data through DDR from Jurik tools.

Have you experimented with generic algorithms as well?

I have a train network that I have been using for quite some time now are you currently using any information which your net is providing?

Jure,
It sounds like you've had some experience using NN's. I however have not so if I say something that makes no sense forgive me, but had a few questions off of what you said. I'm only asking in regard to a NN that would have backward capabilities.


"You can't just make a bunch of interconnected "neurons", feed it with raw price (probably a window of past prices) and expect anything useful to come out. Feeding it with multiple inputs (such as various indicators, probably derived from price) will probably make things worse. "

--Now it sounds like the curse of dimensionality is something I should be afraid of, but indicators do seem redundant because there's no executable actions. Would actual trading systems with functions that create executable actions out of data help? Kind of..clear up and refine that data?


"There's also always the danger for the net to overfit the data. The net would literally memorize the training set, while totally failing at predictions on new data."

--Would you be able to memorize the data but dump "least effective" based on a defined set of parameters, while constantly accepting new data?


"I could go on, but ... the point is, there's a lot of voodoo required to make anything useful with NNs. A lot of people think a NN is some black box that magically learns to map history to future"

--Understandably noted But seeing as we know the future is and always will be just that.. based off of what claude seemed to be going for was more related to probablility. I think he was saying before.. referring to the probability of the next three bars meeting buy or sell conditions. That wouldn't be out of the realm of an NN would it?

SO... sorry to clog the thread, you won't see me do it very often that's for sure. But claude says, "A truly wise man, always has more questions than answers." I'm really excited to see where this goes and can tell I'm gonna nead a life-jacket. Claude, Jure, Be-mac, it'll be fun watching experienced minds woking together on this stuff.

wishes of appreciation and Best of luck,
Jax

Very interesting work with your NNs Claude.

I have been playing with some ideas that could use genetic algorithms myself. Basically I got the idea from reading articles about Financial Labs which uses GA to attain some great results (24% annum/4% maxDD). My research currently pertains to essentially creating various systems and then letting them loose where the GA selects for the best performing putting most weight on that, mating successful systems and adapting to new market structure (mutations). You may wish to check out literature on that.

No breakthough ideas here (yet!) but its pretty fascinating.

I realized that my previous post might have turned out a bit too negative. To clarify, I just wanted to address the "holy grail" thing. NNs are the holy grail of automated trading just like forex is the holy grail of money making. In other words, the potential is great, but so is the effort required to actually make use of some of the potential.

Claude:
I wrote my own software. Nothing fancy, just a library of what I accumulated through experimenting (I also experimented with GA, yes). I did check out some software packages out there, but I always found some limitiations that I didn't like. I must admit though, NeuroSolutions is pretty flexible.

I'm not using any NN trading model, except my brain . You should know, that I'm quite new to trading and most of my earlier work wasn't directly connected to that. Also, I have an idea that I'm working on, but it's not really NN based (it's some sort of a probabilistic state machine). I might use a NN (kohonen self-organizing maps) for some part of it though, like vector quantization.

JaxPacific:
I will forgive you, of course, but only up to 70*7 times ... and above. I hope for the same myself .

"Now it sounds like the curse of dimensionality is something I should be afraid of, but indicators do seem redundant because there's no executable actions. Would actual trading systems with functions that create executable actions out of data help? Kind of..clear up and refine that data?"

By "executable actions" you mean buy/sell signals? It would be definitely helpful in reducing the noise and simplify the input for the NN. But that doesn't mean anything by it self.

"Would you be able to memorize the data but dump "least effective" based on a defined set of parameters, while constantly accepting new data?"

That's exactly what a properly designed NN does. It "memorizes" only the most general features of the training set and rejects the rest as noise.

"Understandably noted But seeing as we know the future is and always will be just that.. based off of what claude seemed to be going for was more related to probablility. I think he was saying before.. referring to the probability of the next three bars meeting buy or sell conditions. That wouldn't be out of the realm of an NN would it?"

It's not out of the realm if you want to do it the hard way with no real reason. You don't need a NN for that.

----
All you really want here is to compute the probability of a successful buy or sell for a particular setup (correct me if I'm wrong). Let's say you use 2 indicators for defining setups. Those 2 indicators make up a 2 dimensional surface. For each point on that surface, you simulate a trade in both directions with some predefined exit rules to find if it turns out successful or not. When finished, you use this as success map for future trades. You enter trades only when the map says it has a % of success greater than some X you choose. You get the % by finding N nearest neighbours, like Claude mentioned, or by definding clusters (LVQ, Kohonen nets, etc), ... or, you could be more eccentric, have some fun and do the following:

- export the map for long trades as an image with white pixels representing success, and black failure (or maybe a shade representing a frequency of profit, because there might be more than one trade per pixel, depending on how you quantizated the map)
- make two images: one for long trades and one for short.
- open them up in Photoshop.
- apply gaussian blur with a desired radius (that's your N-neighbours equivalent sort-of).
- you could apply some other artistic filters, but be prepared for larger drawdowns ...
- use the shade of the pixels on the resulting blurred images as probability maps for initiating future trades.
- base money management on whitness of a pixel.


Sounds profitable?

----
I wrote so many answers, I feel totally stupid now .

Well I know you can save the FFT data. This is in the form of a long list of frequencies and amplitudes. I believe I've exported XY data too but if not, a good plotting program will have an inverse FFT function.

Hello Mahras

Thanks for the information I will definitely do some reading. Your current work seems to echo some of Dr. John F Clayburg latest work.

There are some big differences, though, and it is pretty controversial. He has developed a function called parallel. Essentially, he runs the same system with different parameters settings all at once in the background. The function identifies the one that is most profitable for current market conditions and switches over to those parameters.

It goes against everything you've ever read about setting parameters for a trading system I think the only reason he is getting anyone to listen is because of his other systems which are so profitable. Two of them being ranked top 10 at future magazine cyclones system and belly trader.

Anyhow, if you wanna to look it over he is kind enough to provide some coding and an explanation. You can get it here.

http://www.clayburg.com/omw_abstract.html

Good luck

Thank you Mike for the response. It looks pretty interesting. I hope you let us know if you come up with some more information.

This essentially creates a very curve fitted model which can never replicate its results. While they may work for the short run, if any fundamental change in prices occur it shall be unable to deal with it.

Like I said very controversial but the man is no fool it will be interesting to see what he's newest systems do.They are being tracked by third-party services.This is where we'll find out if it's curve fitted or has the man actually come up with something that is worth looking at.

As a side note, I have my presentations as well.

Claude,

Would it be possible to code Neural Net,Fuzzy Logic, AI on a platform like VT or Metatrader?

Hello amzi

The coding that I presented is inspired by neural networks, and shows the type of logic that they use but is far from being able to duplicate what a dedicated software would be able to do.

But if you're asking about the coding that is put up right here on FF any software with the capabilities of internal loops and optimizing capabilities could reproduce this coding. Unfortunately, these are two things that VT is not capable of yet. But Metatrader could.

Claude,

Thanks for clearing it all up for all of us.

I hope you could give us some sample so we could evaluate Neural Net potential on Metatrader and perhaps look at it in action apart from the routine trading activities.

I hope you could write some Neural Net for Metatrader on Trend Following System and Breakout System.

[Quote....]
Claude,
I hope you could write some Neural Net for Metatrader on Trend Following System and Breakout System.
[quote]


Claude... I don't wish to spend the time running these sorts of tests. Could you just send me what is going to be the High for Today and also the Low.

Better yet... could you just call my br_ker and tell him to Buy @ Today's Low and Sell @ Today's High for me? That'd grrreat. Thanks Claude.

only joking guys.