Australian (ASX) Stock Market Forum

Neural network software

Jump to new
Gringotts Bank

Gringotts Bank

Joined
30 June 2007
Posts
7,200
Reactions
1,225
Best low priced software I have used is Stockneuromaster. It does however have significant problems meaning that i wouldn't ever recommend buying it. All the other stuff I have tried (about 6 different softwares) are totally useless for various reasons.

Neuroshell's stuff (Ward Systems) has won the 'Stocks and Commodities' awards for best AI technology for 10 years or so, but I can't get my hands on a trial version. Cost is approx AU$2000 for the basic version.

Has anyone used neural nets, and/or Neuroshell software?
 
Gringotts Bank said:
Best low priced software I have used is Stockneuromaster. It does however have significant problems meaning that i wouldn't ever recommend buying it. All the other stuff I have tried (about 6 different softwares) are totally useless for various reasons.

Neuroshell's stuff (Ward Systems) has won the 'Stocks and Commodities' awards for best AI technology for 10 years or so, but I can't get my hands on a trial version. Cost is approx AU$2000 for the basic version.

Has anyone used neural nets, and/or Neuroshell software?
Click to expand...

Im working with my son (Doctor of Physics) on this(Very early days) and other challenges at the moment.

I notice DR Bruce Vanstone uses Wealth lab in his thesis.
http://works.bepress.com/bruce_vanstone/
I have his thesis if youd like a PDF just private mail me your email address.

We are looking at Metlab.(Have)
http://www.mathworks.com/products/matlab/
 
thanks tech, I've actually been in contact with Bruce over a number of issues and he's been quite helpful.

Just interested why you'd choose matlab over neuroshell?
 
Main reason is Kris has/it uses it.
Whether it suits what we wish to do with everything is yet to be seen.
He's the scientist and I'll be guided by him.
We are in very early stages so you may be right in the long term---cant really answer with a definative.
 
Gringotts Bank said:
Best low priced software I have used is Stockneuromaster. It does however have significant problems meaning that i wouldn't ever recommend buying it. All the other stuff I have tried (about 6 different softwares) are totally useless for various reasons.

Neuroshell's stuff (Ward Systems) has won the 'Stocks and Commodities' awards for best AI technology for 10 years or so, but I can't get my hands on a trial version. Cost is approx AU$2000 for the basic version.

Has anyone used neural nets, and/or Neuroshell software?
Click to expand...

I have not used it but had a cyberfriend who used it to develop a model for QQQQ. From watching and speaking to him I was struck by the amount of time it took to develop a good system and then maintain it. I was also a bit disappointed to see that what works for QQQQ did not work very well for other symbols. If that is generally true, then one can expect to invest a significant time for development and maintenace for each symbol.
 
colion said:
I have not used it but had a cyberfriend who used it to develop a model for QQQQ. From watching and speaking to him I was struck by the amount of time it took to develop a good system and then maintain it. I was also a bit disappointed to see that what works for QQQQ did not work very well for other symbols. If that is generally true, then one can expect to invest a significant time for development and maintenace for each symbol.
Click to expand...

That makes complete sense
as your attempting to train the software to anticipate positive moves.
Cutting out the noise so to speak.
It also shows that each symbol (from your friends experience) has its own nuances.
 
What's so great about Neural Networks specifically? I studied them, and other probability inference models at University and built a few pieces of software on the concepts. We also use inference models in a lot of the research done at the lab I currently work at (www.neuroimaging.org.au).

Personally I prefer Bayesian Inference or Support Vector Machines, all you need to be able to do is write a small piece of code that can quantify a set of conditions into vector space and then run that code over lots of your time series (in this case, a price chart) to train it. In any case, all three of these concepts (Bayes, SVM, NN) all accomplish the same task from a Turing perspective.

I am not saying anything is better than anything else or that you shouldn't use NN, I am just curious as to why NN was picked over other inference models considering the efficacy and relative easy learning curve of say, SVM.

Personally, the concept interested me too and I built my own bot from scratch in Python (which has freely available libraries for Neural Net stuff, and the SVM stuff is just plugged in using textfiles). I have written multiple SVMs and represented each SVM as a NN neuron. One SVM measures global macro stuff, one measures volatility stuff, some of the SVMs are the same but measuring on different timeframes, etc etc, so the SVMs are interacting and learning on their own while they teach the NN as a whole. This is actually my personal solution to the "over-training" problem listed in wabbits above article. Some of the SVMs have a 1 hour "memory", some have a 1 day memory and some are just sitting there checking what happens when price X is above yearly open Y across millions and millions of instruments. This allows the whole network to understand intraday and intraweek developments as they occur as well as keeping mind of the bigger picture. Add a genetic algorithm to "naturally select" which neurons work when and you've got your bot. Mine measure/predict the probability of an up or down close in the coming open of a specific set of instruments so the concept is vastly different from the average HFT NN which is all the rage these days.

But as in techs case, it is early days yet. I think to have a good probability inference model, you need to quantise a really huge vector space sort of like the stuff that used to be done at Princeton Economics before the words "HFT" even existed. Most of the academic papers on HFT NN stuff show that even though the NN can predict price action with uncanny accuracy, it generally still isn't enough to be profitable due to trading commisions and spread.

Which is why, surprise surprise, most of the successful HFT shops are liquidity providers in the US, who get cash-money rebates for providing liquidity using these bots. Their business model is only profitable because they have extremely low latency exchange access and the exchange gives them rebates for providing liquidity around the NBBO.
 
sinner said:
Most of the academic papers on HFT NN stuff show that even though the NN can predict price action with uncanny accuracy, it generally still isn't enough to be profitable due to trading commisions and spread.
Click to expand...
I don't understand. How does the uncanny accuracy fail to covert to profits?
 
Thanks sinner, you obviously have a very solid background in these sorts of things.

I wonder if you'd mind providing some links to either your own work, or maybe some SVM software that I could trial?

I'd also be interested in what you use nowadays for your trading - bot/manual/discretionary, short/medium term, stocks/indices/derivatives, local/international. And the big one - does it make money over the long term?

If it's all too complicated, (I don't have a maths or finance background) I may just end up buying some tips from a guy in the states who seems to have a good record with predicting the DOW five days out using NN's.

Thanks for the link wabbit. I like it.
 
Gringotts Bank said:
Thanks sinner, you obviously have a very solid background in these sorts of things.

I wonder if you'd mind providing some links to either your own work, or maybe some SVM software that I could trial?

I'd also be interested in what you use nowadays for your trading - bot/manual/discretionary, short/medium term, stocks/indices/derivatives, local/international. And the big one - does it make money over the long term?

If it's all too complicated, (I don't have a maths or finance background) I may just end up buying some tips from a guy in the states who seems to have a good record with predicting the DOW five days out using NN's.

Thanks for the link wabbit. I like it.
Click to expand...

Ive passed Sinners post on to my personal Rocket Scientist to see if he has anything to add.

GB
You could pay someone to design one and then have sole access rights.
The Question is what inputs do you choose.
The anwer of course is personal and may take a great many options to be tested---some obvious and others not so.

My chosen inputs are working well enough without NN input on Index Futures (SPI FTSE DAX).
 
For what its worth
Kris's reply to Sinners comments.

"That guy clearly knows his stuff. He's just asking "why use neural networks in that way?" Most people haven't even heard of them, let alone studied them and other machine-learning techniques like he would have (and now does in his day-to-day job?). My only comment would be to agree. Why use a specific technique over another when they haven't been tested against one another? It's an important question. Then again, you can always argue that if it works, who cares? There may very well be a better way to do it, but not everyone's enough of an expert to know the alternatives."
Click to expand...
 
I hadn't heard of SVM because it's not in the popular trading magazines (least not that I've seen) whereas NN's are everywhere in magazines, and also on the web.

I assume if I was to look at finance journals I could find lots on SVM, but I find them hard to read, what with all the statistical jargon. Also, academic literature is very specific and hard to generalize from. Academics are also notoriously bad a at making money (a subject all of it own!).

So I'm looking at more commercial stuff, and if SVM type software is available in a reasonably user-friendly format, then I'd definitely look at that.

atech, do you have any stats you could share on the performance of your system?
 
GB
Ive only reciently swapped to Index futures and trade a discretionary method based around VSA/my own indicators and interpretation of Price and volume.

I havent worked out actual perfomance (Should) but profit is fine so I havent really bothered.Time is my enemy!
 
Hi guys,

Forgive my delayed reply I forgot about this thread over the weekend. Some questions to answer:

I don't understand. How does the uncanny accuracy fail to covert to profits?
Click to expand...

Most of the HF bots are designed to work on the tick scale using low latency connection to exchange (or maybe even direct hook to the exchange) so the bot can't simply "go long at price X and wait for price Y" or some such business, rather it has to constantly scalp the bid/ask (just like the pros do it, in fact this is why HFT has been replacing the "specialists" who used to run the trading floors). The issue then becomes mainly one of transaction costs I believe. Since I have never actually been inside one of these HF trading firms I can't tell you for sure.

From my understanding, on strongly trending days, floor traders will get killed as their methods rely on being able to sell what you just bought for a tick or vice versa. So correctly inferring direction doesn't necessarily imply profits. I can tell you for sure that is the case.

I wonder if you'd mind providing some links to either your own work, or maybe some SVM software that I could trial?
Click to expand...

Considering the amount of time spent so far, you would be very hard pressed to get me to give up my finance code. Maybe for a few hundred thousand dollars I believe is the going rate for a functioning bot, and I can guarantee my bot is like no other written.

The software I use most is:
http://svmlight.joachims.org/
http://arctrix.com/nas/python/bpnn.py

However I will happily give you a totally free example concept for a simple bot that a daytrader might use to gauge the direction of the day before the market has opened.

Define your vector space. A vector is just a matrix with only 1 row. So lets say our vector space looks like this:

BullBot:
BullClose[0|1]:: Above20MA[0|1], AboveMonthlyClose[0|1], Volume>YesterdaysVolume[0|1]
BearBot:
BearClose[0|1]:: Above20MA[0|1], AboveMonthlyClose[0|1], Volume>YesterdaysVolume[0|1]

So, for example, if we ran a program over our data to convert the data into vector space a close higher than the open with close above 20MA, above last months close with Volume<Yesterdays Volume would look like this to BullBot:
1:: 1, 1, 0
and to BearBot it would look like this:
0:: 1, 1, 0

Now let's say we've run the program over 10,000 days worth of data we would have a bunch of vectors that look like this (just a random example):
1:: 1, 0, 1
0:: 1, 1, 1
1:: 1, 0, 0
1:: 0, 0, 0

etc.

This is our training data. We feed these vectors into our SVM or Bayesian Machine (one each per classification technique) and it should do all the hard work of probability analysis for us. There are some issues with auto-correlation and stuff, but for the sake of this example we will ignore.

Now to use the bot, we would grab yesterdays vector space before the next market open and feed it into BullBot and BearBot. The food would look like
X:: 1, 0, 1

Where we want the machine to evaluate X probabilistically for us. Each machine should think for a second and then spit out a 1 or 0 for X. If the direction for both the machines agree (i.e. BullBot says 1 and BearBot says 0 or BearBot says 1 and BullBot says 0 - but definitely not if they both say 1 or 0) then we can use that as our directional bias for the day.

This is just an example concept, to familiarise the ideas of converting parameters to vector space and some bot trading concepts. It is not a real live profitable bot (although a short-term bot like this is one of the neurons in my bot), so don't run off and try to code it.

I'd also be interested in what you use nowadays for your trading - bot/manual/discretionary, short/medium term, stocks/indices/derivatives, local/international. And the big one - does it make money over the long term?
Click to expand...

90% mechanical with leftovers discretionary, almost all forex (London hours) with a little bit of FTSE, DAX and SPX thrown in as required. I also trade leveraged gold, but the technique used means usually only 1 or 2 trades per quarter. Check out the thread "Trading Chaos" for a system I like a lot. I hate the question "does it make money" or worse "what is your performance using this system". Find out for yourself or go get a job since trading isn't for you! Other mechanical stuff in the works: I am working on a long term "buy and hedge" portfolio system which just monthly buys/rotates the top 10 ASX100 stocks based on my personal ranking system and hedges the dollar amount using futures or options if necessary. Also working very very hard on a Delta-Neutral options trading model (in fact I spent my Saturday up until 4am doing research on this topic how cool am I) so I can start buying and selling vol. Business plan on that almost complete now. The goal across all systems is to be robust, but with as little time in front of the screen as possible.
 
sinner said:
Also working very very hard on a Delta-Neutral options trading model (in fact I spent my Saturday up until 4am doing research on this topic how cool am I) so I can start buying and selling vol.
Click to expand...

oooo...nice, will be interested to see your attempts.
Though curse the path dependency of long/short replication as well as tail event risk, even if you can predict vol accurately. GL
 
sinner said:
Hi guys,

Forgive my delayed reply I forgot about this thread over the weekend. Some questions to answer:
Click to expand...

No need to seek forgiveness when the information is of such density :D


Most of the HF bots are designed to work on the tick scale using low latency connection to exchange (or maybe even direct hook to the exchange) so the bot can't simply "go long at price X and wait for price Y" or some such business, rather it has to constantly scalp the bid/ask (just like the pros do it, in fact this is why HFT has been replacing the "specialists" who used to run the trading floors). The issue then becomes mainly one of transaction costs I believe. Since I have never actually been inside one of these HF trading firms I can't tell you for sure.
Click to expand...

HF, Flash, LL, etc etc
Is now actually getting more influential in the market (this is my area of expertise in a round about way) but has been around for about 6 years.
Exchanges are now offering "Co-Location" with member/customer gateways being located in the same data-center as the matching engines, some even allow a customers own server next to it..

ASX plan on this is the short near future.

latency is sub ms for these systems, but when they go wrong they go very wrong.
 
latency is sub ms for these systems, but when they go wrong they go very wrong
Click to expand...

Speed is always the issue.But then again why not broaden the timeframe to the point where it isnt?

Would you just be able to use an equity curve switch for "when they go very wrong?"
 
tech/a said:
Speed is always the issue.But then again why not broaden the timeframe to the point where it isnt?

Would you just be able to use an equity curve switch for "when they go very wrong?"
Click to expand...

Some of these firms the biggest, trade up to 200,000,000 shares in a day.

They just want more and more volatility

Personnel bots should have about 4 checks for every decision IMHO
 
Top