QuoteOriginally posted by: crowlogicNN, feed forward, recurrent, time-delay, dynamic, first order, second urder training, unscented, filtering, generatic topograhic mappings, auto-associative, long-term-short-term, hierarchal, etc?Huh? A neural net is just an evolved stimulus-response catalog which abstracts the common elements of patterns as the advantageous way to summarize data.My web sites says "something along the lines of a neural network" because you can achieve the same result as a neural network, without going to all the trouble to use a textbook neural network to do it. Any erosion or growth process in nature is an example of a structure that has "learned" what will survive in its environment. Any outcome which stabilizes and recurs after a recurring process has learned about its environment, by revealing or reflecting some record of it.And no, it is not my knowledge of neural networks that makes me optimistic about the strategy. What makes me optimistic, is what I use the procedure to accomplish, the relevant summary of data I've chosen to isolate. I've forgotten most of what I ever knew about standard neural networks, since most of what comes out of them is an artifact of the network itself, rather than of the data being studied. So your selection of which one to use, rather than the network itself, would dictate the result. Or something like that...I've used simple classification methods for years, and it occurred to me how one could be used to get the economic result I had been working on. The most inspirational thing I ever read on the topic is ICMAUS, but you don't need anything like that to do the looking, once you know what you are looking for. What I mainly mean to communicate with my description of the strategy, is that I don't use any of the common/elegant time-series from physics tools.The program doesn't discover the strategy, just execute the strategy which I conceived over a period of years, by turning inputs into outputs. In truth, I can't always remember off the top of my head why I want it to do what it does. It takes some hard thinking, for me at least, to understand/remember the concept.

The main problem I remember with the more biologically-inspired experience recorders - the ones that try to look like a brain at the same time as selecting inputs - was timing. Suppose signals propagate through it to achieve some kind of complex mixing, where the propagation is animated in a manner similar to a microprocessor shifting bits on a sawtooth voltage signal. The thing has to be either hopelessly complex to learn some simple rule, or be contrived with that rule in mind, to have the signals it needs at the same time be available in a way which they can become associated. In retrospect, maybe I could have left things on longer than a single clock cycle or something, though I imagine I would have thought of that.Anyway that was my experience, though I don't remember the particulars. I do remember that using nerves that only went one direction was miserable, and using inhibitors didn't help much. I tried to design a number of simple creatures, and all I remember is that the "timing" never worked. So now I just design something to do exactly what I want it to do, without worrying about trying to make it look like nerves. It's a much more direct way to specify what you want to try, what you want to give it a chance to learn, or what "space" you want to search and with what sort of economization. Just use nested loops. You can do 10,000 times what a neural network can do, in 1/10th the time - and without having to be particularly bright or a good programmer - using nested loops.In summary, it is silly to take a good computer, and make it work like a bad brain. Neural networking is a toy. My only faint interest in the neurons any more, is that I am no longer satisfied with the explanation of human consciousness in "The Sensory Order.". Though I wouldn't waste my time thinking about it.

QuoteOriginally posted by: farmerThe main problem I remember with the more biologically-inspired experience recorders - the ones that try to look like a brain at the same time as selecting inputs - was timing. Suppose signals propagate through it to achieve some kind of complex mixing, where the propagation is animated in a manner similar to a microprocessor shifting bits on a sawtooth voltage signal. The thing has to be either hopelessly complex to learn some simple rule, or be contrived with that rule in mind, to have the signals it needs at the same time be available in a way which they can become associated. In retrospect, maybe I could have left things on longer than a single clock cycle or something, though I imagine I would have thought of that.Yes, timing is definately the most important part of neural information processing. Your description of signal propagation in neural nets is pretty accurate... the complex interweiving of signals hitting at just the right phases to induce a coherent response is how real neural nets assimilate, detect, predict, fill in missing gaps, etcVery very fiew 'neural' models accurately describe real signal propagation in actual neurons, only a certain subclass do: hopfield, hebb, etcHebbian learning is actually very 'natural' and elegent and does not take a long time or use a complex algorithm to learn complex temporal signals... think: physics simulation.The whole goal of biological neural modelling is to create equations simple enough to be biologically plausable, yet accurate enough to replicate the observed experimental data recording from individual and global neuron firings. Simplicity rules out complex training or adaptation algorithms and the system must also deal with inherently noise(volatile) data.QuoteAnyway that was my experience, though I don't remember the particulars. I do remember that using nerves that only went one direction was miserable, and using inhibitors didn't help much. I tried to design a number of simple creatures, and all I remember is that the "timing" never worked. So now I just design something to do exactly what I want it to do, without worrying about trying to make it look like nerves. It's a much more direct way to specify what you want to try, what you want to give it a chance to learn, or what "space" you want to search and with what sort of economization. Just use nested loops. You can do 10,000 times what a neural network can do, in 1/10th the time - and without having to be particularly bright or a good programmer - using nested loops.In summary, it is silly to take a good computer, and make it work like a bad brain. Neural networking is a toy. My only faint interest in the neurons any more, is that I am no longer satisfied with the explanation of human consciousness in "The Sensory Order.". Though I wouldn't waste my time thinking about it.99.99% of neural networks are indeed toys.See the work of John Hopfield, this is definately not a toy:Contest prompts thinking about thinkingAlso see: Baby robot learns like a humanIf the neural models you played with then had only one direction connections and operated in discrete time then you must have been using some really primitive stuff.A coworker of mine designed neural net solutions for boeing and military research for commerical airliner autopilot systems. The neural solutions they had to optimize ran for 4 to 5 days on a Cray (this was about 8 years ago). The systems performance far exceed all of the hand-tuned or more classifical control systems. The pilots rejected the system because they didn't understand why the machine was making the decisions it was making and they were spooked by it.In short: NN's have a bad wrap due to the obsession of this stupid crap floating everywhere called 'neural nets' now, which are mainly a few matrix multiplications and a nonlinear function appled to the output.See this excellent book for a non-technical explanation of why nearly all traditional AI research is crap: On IntelligenceGood luck,Stephen

QuoteOriginally posted by: crowlogicSee the work of John Hopfield, this is definately not a toyI don't know, I thought I got some laughs with it in the D.E. Shaw thread. And I've never seen it used for much else. Didn't he come up with the minimal layout it in like 1977? I would have thought he'd have come up with something more interesting than recognizing pictures by 2001. The only real NN application I've heard of is recognizing handwritten signatures, and I doubt that is a Hopfield network.QuoteOriginally posted by: crowlogicIf the neural models you played with then had only one direction connections and operated in discrete time then you must have been using some really primitive stuff.Primitive? The brain that wrote that sentence has one-directional connections. And if there is some way to use a computer to make something that operates in continuous time, that would be quite a revelation to me.After reading "The Organization of Behavior", I wanted to make my toy as close to the human brain as possible. Because I knew the human brain worked, and I had one handy to study. Economic signals are two-directional. But I wanted to expand my structures gradually once I got them working, and I thought there was good evidence human neurons were suitable for that.Hayek actually came up with his more general model decades before Hebb. And Hayek's model is physiologically feasible.

QuoteI don't know, I thought I got some laughs with it in the D.E. Shaw thread. And I've never seen it used for much else. Didn't he come up with the minimal layout it in like 1977? I would have thought he'd have come up with something more interesting than recognizing pictures by 2001. The only real NN application I've heard of is recognizing handwritten signatures, and I doubt that is a Hopfield network.He discovered that stuff in the early 80s. That's not what I'm talking about, as neat as i t is.Do a search for his recent papers (last 2 to 3 years) on neuro-biology and auto-associative invariant memory representations.QuoteQuoteOriginally posted by: crowlogicIf the neural models you played with then had only one direction connections and operated in discrete time then you must have been using some really primitive stuff.Primitive? The brain that wrote that sentence has one-directional connections. And if there is some way to use a computer to make something that operates in continuous time, that would be quite a revelation to me.How are you defining one way? Yes, information builds up until a trigger releases a pulse and the integrator is reset. So information 'flows out' along the dendrites and stimulates other neurons, whic feeds back around and eventually into the source again. But yes, the chemical structures makes the pulse travelling out, one-way-only and then branches out, but it can come back.Numerical integration with some of the more advanced techniques comes pretty damn close to being actually continuous as long as you detect 0-crossings properly and set your tolerance as well and your time derivatives dont do anything physically implausible.QuoteAfter reading "The Organization of Behavior", I wanted to make my toy as close to the human brain as possible. Because I knew the human brain worked, and I had one handy to study. Economic signals are two-directional. But I wanted to expand my structures gradually once I got them working, and I thought there was good evidence human neurons were suitable for that.Hayek actually came up with his more general model decades before Hebb. And Hayek's model is physiologically feasible.Yes I know Hayek's models too, i've played with them to using numerical integration.All of these neural models are powerful but were simplified due to the lack of cpu power to properly evaluate the integrals, but now that CPU is cheap this kind of stuff can take place.

QuoteOriginally posted by: farmerThe only real NN application I've heard of is recognizing handwritten signatures, and I doubt that is a Hopfield network.Surely the post code/zip code reading software is the most famous/succesful? Also there are some adaptive optic telescopes now where the mirrors are controlled by neural nets. Although neither example is a hopfield network.As far as simply modelling (rather than control) Ising models are still v popular in physics which are special case of hopfield network.

QuoteOriginally posted by: MikeCroweSurely the post code/zip code reading software is the most famous/succesful?I always thought that if you don't use "REDRUM" font, it got bumped to a hand sorter and took two extra days. But maybe that's why girls dump me after the first valentines-day card...

QuoteOriginally posted by: crowlogicIf you are serious about trading you should really check out Lime BrokerageI've used IB's java api before.. it sucks, they limit your message rate and you can get burned in fast moving markets and even when markets arent moving fast you are dealing with slow, incomplete data. Also, you still need that huge beast of a program running in the background just to use the API.Lime's api is completely java based.. very very fast, simple to use, pretty good design. I can send you the javadoc's if you like. Also, just last month they were the top liquidity provider for the NYSE.Top Liquidity Providers in NYSE-Listed Securities - March 2006 I don't work for em them have any other interest besides being a very happy customer.Hi all,crowlogic: I'm interesting in trading from few weeks ago...I'm a software analyst and developer and I want to develop a platform like your.I develop in many languages but my favorite is Java and surfing internet I find Interactive Brokers. They haven't a well guide to develop with their APIs (also no javadoc ).Now I read your post about Lime Brokerage..Seeing LB site it seems not professional. I can't download nothing. No serious informations about APIs and not prices...It's strange because they are famous and I believe good!Can you give me (us) more informations and, if you can, docs of the APIs and so more?Thank you very much.Z.P.S.: I have much more questions on automatic trading but I make them in future...

Not to be an ass, but are you serious? Interactve Brokers is such an amateur organization.. its pitiful. Their API is absolute garbage and they require this huge monolithic GUI application to be running to use the API last I checked. They also have an arbitrary 50 message per second message rate and NO gaurantee that you are getting raw price info.Lime Brokerage is only for hedge funds and 'sophisticated traders' which is why you dont see pricing info on their site, they dont advertise to the public, InteractiveBrokers does thus they have all sorts of resources to waste supporting non-tech-savy end users whereas Lime doesnt.QuoteOriginally posted by: zorrothekidQuoteOriginally posted by: crowlogicIf you are serious about trading you should really check out Lime BrokerageI've used IB's java api before.. it sucks, they limit your message rate and you can get burned in fast moving markets and even when markets arent moving fast you are dealing with slow, incomplete data. Also, you still need that huge beast of a program running in the background just to use the API.Lime's api is completely java based.. very very fast, simple to use, pretty good design. I can send you the javadoc's if you like. Also, just last month they were the top liquidity provider for the NYSE.Top Liquidity Providers in NYSE-Listed Securities - March 2006 I don't work for em them have any other interest besides being a very happy customer.Hi all,crowlogic: I'm interesting in trading from few weeks ago...I'm a software analyst and developer and I want to develop a platform like your.I develop in many languages but my favorite is Java and surfing internet I find Interactive Brokers. They haven't a well guide to develop with their APIs (also no javadoc ).Now I read your post about Lime Brokerage..Seeing LB site it seems not professional. I can't download nothing. No serious informations about APIs and not prices...It's strange because they are famous and I believe good!Can you give me (us) more informations and, if you can, docs of the APIs and so more?Thank you very much.Z.P.S.: I have much more questions on automatic trading but I make them in future...

Oops, forgot.. here is the requested javadocs.Also, they only deal with fairly high volume stuff.. you need to be able to move at least 50,000 shares a day I believe, unless it has changed recently. Prices are direct pass thru.. you negoiate with them based on what sort of volume you can do.

Thank you very much for APIs and informations. I'm a beginner trader. I'm starting now with my ideas of automatic trading and I'm seeing what the market offers.I believe you about Lime Brokerage and I think the same about IB but I can't find now a better offer for me.I'm very far from the 50,000 shares a day. I hope to use Lime B. APIs soon but...can you advise me about a good platform for testing my code, trade few shares and with contained prices?Thank you,z.P.S.: I'm seeing LB javadoc. Very simple and good developed!

I'm not sure of any other quote providers.. I checked into several but they all had their problems (windows based proprietary interface), expensive, incomplete, etc.IB is not all *that* bad.. probably one of the best places you that you can start off on your own.. you might shop around, I know there are some places that provide nice quote feeds but i havent used them. If you are clever you can hack some java and reverse engineer some of these web-based java applets that do realtime streaming quotes (many of the ECNS have them for free!)... I did that back in the day.Lime's java api is very clean, robust and fast..definately the best I've seen so far.QuoteOriginally posted by: zorrothekidThank you very much for APIs and informations. I'm a beginner trader. I'm starting now with my ideas of automatic trading and I'm seeing what the market offers.I believe you about Lime Brokerage and I think the same about IB but I can't find now a better offer for me.I'm very far from the 50,000 shares a day. I hope to use Lime B. APIs soon but...can you advise me about a good platform for testing my code, trade few shares and with contained prices?Thank you,z.P.S.: I'm seeing LB javadoc. Very simple and good developed!

QuoteOriginally posted by: crowlogicI'm not sure of any other quote providers.. I checked into several but they all had their problems (windows based proprietary interface), expensive, incomplete, etc.I'm seeing that...QuoteIB is not all *that* bad.. probably one of the best places you that you can start off on your own.. you might shop around, I know there are some places that provide nice quote feeds but i havent used them.Yes. I think so...QuoteIf you are clever you can hack some java and reverse engineer some of these web-based java applets that do realtime streaming quotes (many of the ECNS have them for free!)... I did that back in the day.Yes I can hack Java code. I'm thinking also to parse HTML quotes in web based stocks information systems. For the Java the problem is the illegality of this action...QuoteLime's java api is very clean, robust and fast..definately the best I've seen so far.I hope to use Lime's API soon...Now I'm studying financial systems and I'm modeling my platform: database, all programming stuffs such objects, agents, interfaces, algorithms, etc.I have to study more...I'm not sure on my DB model. Can you tell me good books to model (and simply understand) financial systems for a "computer science mind"?Thank you,Z.

QuoteOriginally posted by: zorrothekidQuoteOriginally posted by: crowlogicQuoteQuoteIf you are clever you can hack some java and reverse engineer some of these web-based java applets that do realtime streaming quotes (many of the ECNS have them for free!)... I did that back in the day.Yes I can hack Java code. I'm thinking also to parse HTML quotes in web based stocks information systems. For the Java the problem is the illegality of this action...HTML quotes are not good because they are polling rather than event driven so you will have a lot of network overhead, missed quotes, duplicated quotes, etc, all this stuff is extremely important in the high-frequency setting. And about the illegality, who cares? They dont know. Just change your user-agent string to that of a normal browser.QuoteQuoteLime's java api is very clean, robust and fast..definately the best I've seen so far.I hope to use Lime's API soon...Now I'm studying financial systems and I'm modeling my platform: database, all programming stuffs such objects, agents, interfaces, algorithms, etc.I have to study more...I'm not sure on my DB model. Can you tell me good books to model (and simply understand) financial systems for a "computer science mind"?Thank you,Z.Oh how I wish there was..I'm not aware of any, I've basically had to switch out of computer-science mind and into mathematical/algorithmic mind. This process has taken me well over a year of intensive study and the more math I learn the more I realize there is that I need to learn. It basically boils down to, you need to study stochastic process evolution and then once you understand the math it'll become apparant how this turns into algorithm.Also, re: databases. My initial idea was to work with SQL.. that turned out to be entirely too slow and cumbersome. ended up going with Berkeley DB (in-memory/on-disk embedded database, lowlevel, extremely fast).Good luck,Stephen

QuoteWhat is a platform? I've written data-management code to receive, store, and retrieve data, decision code to turn data into buy-sell decisions, order-routing code, and visual monitoring code. But I've never coded a platform, so far as I know.To me it doesn't seem feasible to implement every connection by hand. There are platforms like the ORC Liquidator. I would be interested to hear from experiences from this or similar products.Here is a descprition on the multi-access featureQuote Direct access to 120 electronic marketsOur large coverage of markets ensures that you, through the Orc System, will have access to the markets that you want to trade on - both locally and globally.Our unique offering includes simultaneous access to exchanges, whether they be through your membership or through non-membership connections via brokers.Alongside your own membership connections, Orc ExNet gives you direct market access to additional markets without the need for exchange membership. You can use the Orc ExNet service as a compliment to your regular market connections to take advantage of arbitrage and trading opportunities on several markets at the same time.In addition, the Orc System can also be connected to leading market data vendors.