Has Any of You Made Money with Either Neural Network Software or Ninja Trader Tools?

[Vince50 10]

I am not a programmer and not that Internet savvy. Im a trader. A chart reader! And I have always wondered if I was at a disadvantage by not being able to use the Expert Advisers and Backtesting stuff MT4 Platforms give you. As cwell, I didnt find Ninja Traders tools or back testing stuff to be easily usable. So I am wondering, are Neural Network software just a constant rehash of the "magic black box" they keep trying to sell traders on, and who has found a platform that letting them do your backtesting on your system turned out to be profitable?

 

I ask because trade station programmers wouldnt be getting $100 an hour to program whatever you want if backtesting thru coding really was the answer. They would just code themselves up a neat simple winning system, no?

 

Also, what is all this backtesting tools with Ninja Trader and MT4. I post this here because it seems they focus this kind of stuff on the futures market first, then Forex and then stock traders. So..am I missing something guys, or what?

 

BTW.....I went thru a lengthy browser based software program that let me enter any stock I wanted and tweak the results with 2 or 3 dials ( and I could get most stocks to give me a back tested return over 2 years of over 100% annually and almost always 50% or better W/L/ ratio. In forward testing on paper and also in real trading I had a dismal 80% loss ratio!!! So I decided to make lemonade out of lemons and take the opposite trade, (go long if they said short) and I still go decimated. So after 18 months of tweaking this thing, I couldnt understand why it wouldn't work either thru massive winners or massive losers. Has any human being done well with a Neural network program/software or is it all junk?

 

I now am finding success with Harmonic patterns with Fibonacci work so all that is past me now. I just want to know if I gave up too soon on those platforms.

Linda Raschke pointed out, not to my liking that Charles Dow, and Elliot and all the greats of the stock market in the 30's and 40's did all their charts by hand. And she believes that gives you an edge of some sort. I personally will never trade again if I ever had to hand draw my charts. But maybe nueral networks and Expert advisers are going one step too far in the opposite direction. Comments please..............

[BlueHorseshoe 10]

I am not a programmer and not that Internet savvy. Im a trader. A chart reader! And I have always wondered if I was at a disadvantage by not being able to use the Expert Advisers and Backtesting stuff MT4 Platforms give you. As cwell, I didnt find Ninja Traders tools or back testing stuff to be easily usable. So I am wondering, are Neural Network software just a constant rehash of the "magic black box" they keep trying to sell traders on, and who has found a platform that letting them do your backtesting on your system turned out to be profitable?

 

I ask because trade station programmers wouldnt be getting $100 an hour to program whatever you want if backtesting thru coding really was the answer. They would just code themselves up a neat simple winning system, no?

 

Also, what is all this backtesting tools with Ninja Trader and MT4. I post this here because it seems they focus this kind of stuff on the futures market first, then Forex and then stock traders. So..am I missing something guys, or what?

 

BTW.....I went thru a lengthy browser based software program that let me enter any stock I wanted and tweak the results with 2 or 3 dials ( and I could get most stocks to give me a back tested return over 2 years of over 100% annually and almost always 50% or better W/L/ ratio. In forward testing on paper and also in real trading I had a dismal 80% loss ratio!!! So I decided to make lemonade out of lemons and take the opposite trade, (go long if they said short) and I still go decimated. So after 18 months of tweaking this thing, I couldnt understand why it wouldn't work either thru massive winners or massive losers. Has any human being done well with a Neural network program/software or is it all junk?

 

I now am finding success with Harmonic patterns with Fibonacci work so all that is past me now. I just want to know if I gave up too soon on those platforms.

Linda Raschke pointed out, not to my liking that Charles Dow, and Elliot and all the greats of the stock market in the 30's and 40's did all their charts by hand. And she believes that gives you an edge of some sort. I personally will never trade again if I ever had to hand draw my charts. But maybe nueral networks and Expert advisers are going one step too far in the opposite direction. Comments please..............

 

Hi Vince,

 

Like so many questions that are asked on TL, I'm afraid that the answer is probably "yes . . . and no". There are undoubtedly a handful of firms (mostly, though not exclusively) HFT, making a hell of a lot of money using A.I. approaches, but I would guess the number of retail traders who successfully apply anything of this sort are very small in number. I would guess that no off-the-shelf software of this type will give you any consistent and significant profits. That's just my opinion though.

 

So, half a dozen maths and computer science PhDs at a multi-billion dollar hedge fund = bleeding edge A.I. algorithmic trading success.

 

Something called an 'Expert Advisor' that you buy from some clueless internet marketeer's website for $999 = probable drained account.

 

It's interesting that you mention LBR. She has made a point for many years of stating that the human mind is far more adaptive to change, and far better at recognising patterns than algorithms are, and that this gives discretionary traders an edge over automated strategies. While this may be true for anything that you or I may automate, it's probably not true when a trader finds themselves on the other side of some thing put together by, say, RenTech. All the indications are that professional A.I. trading has become very, very sophisticated.

 

As a side note, Neural Nets specifically apparently aren't that popular anymore - A.I. trading moved on to things like genetic algorithms, modelling hidden markov chains, kalman filters, and a whole host of other things that involve the back-propogation of error. Unless you really know what you're doing, Neural Nets are just a complex method of curve-fitting.

 

For an easy and casual read on all this, I recommend Scott Paterson's "Dark Pools".

 

BlueHorseshoe

 

ps The direct answer to your question is that I have never made money trading with Neural Nets

[michael45 10]

I am not a programmer and not that Internet savvy. Im a trader. A chart reader! And I have always wondered if I was at a disadvantage by not being able to use the Expert Advisers and Backtesting stuff MT4 Platforms give you. As cwell, I didnt find Ninja Traders tools or back testing stuff to be easily usable. So I am wondering, are Neural Network software just a constant rehash of the "magic black box" they keep trying to sell traders on, and who has found a platform that letting them do your backtesting on your system turned out to be profitable?

 

I ask because trade station programmers wouldnt be getting $100 an hour to program whatever you want if backtesting thru coding really was the answer. They would just code themselves up a neat simple winning system, no?

 

Also, what is all this backtesting tools with Ninja Trader and MT4. I post this here because it seems they focus this kind of stuff on the futures market first, then Forex and then stock traders. So..am I missing something guys, or what?

 

 

BTW.....I went thru a lengthy browser based software program that let me enter any stock I wanted and tweak the results with 2 or 3 dials ( and I could get most stocks to give me a back tested return over 2 years of over 100% annually and almost always 50% or better W/L/ ratio. In forward testing on paper and also in real trading I had a dismal 80% loss ratio!!! So I decided to make lemonade out of lemons and take the opposite trade, (go long if they said short) and I still go decimated. So after 18 months of tweaking this thing, I couldnt understand why it wouldn't work either thru massive winners or massive losers. Has any human being done well with a Neural network program/software or is it all junk?

 

I now am finding success with Harmonic patterns with Fibonacci work so all that is past me now. I just want to know if I gave up too soon on those platforms.

Linda Raschke pointed out, not to my liking that Charles Dow, and Elliot and all the greats of the stock market in the 30's and 40's did all their charts by hand. And she believes that gives you an edge of some sort. I personally will never trade again if I ever had to hand draw my charts. But maybe nueral networks and Expert advisers are going one step too far in the opposite direction. Comments please..............

 

Don’t get me wrong, when I was surfing around the internet, there were a few really good systems out there for doing backtesting. Have many kind of software in backtesting. But I do not have experience about them, I can not recommend these software.

[Perrin 10]

It makes you wonder how anyone can make money when you put it like that

 

It's like we're up against The Terminator.

 

By the way, thanks for mentioning Scott Paterson's "Dark Pools" - that's another book for my library

[BlueHorseshoe 10]

 

It's like we're up against The Terminator.

 

 

That's really funny because the UK Paperback version of the book has a picture of what is clearly meant to be a Terminator's hand scrolling across a series of prices

 

A lot of this sort of thing happens in sub-second timeframes (all the HFT scalpers), is far more prevalent in stocks than futures (the futures exchanges mostly operate with FIFO orders books, which level the playing field somewhat against the HFTs), and should not be an issue unless you're trying to scalp (in the traditional sense).

 

The book's a very interesting read though.

 

If you wanted to learn more about all this in detail then you could read my posts in the "Daytraders, Do You Know Your Enemy?" thread, which was an abortive online attempt to develop an algorithm for estimating liquidity and position in queue in the CME order books.

 

BlueHorseshoe

[Vince50 10]

Hi Vince,

 

Like so many questions that are asked on TL, I'm afraid that the answer is probably "yes . . . and no". There are undoubtedly a handful of firms (mostly, though not exclusively) HFT, making a hell of a lot of money using A.I. approaches, but I would guess the number of retail traders who successfully apply anything of this sort are very small in number. I would guess that no off-the-shelf software of this type will give you any consistent and significant profits. That's just my opinion though.

 

So, half a dozen maths and computer science PhDs at a multi-billion dollar hedge fund = bleeding edge A.I. algorithmic trading success.

 

Something called an 'Expert Advisor' that you buy from some clueless internet marketeer's website for $999 = probable drained account.

 

It's interesting that you mention LBR. She has made a point for many years of stating that the human mind is far more adaptive to change, and far better at recognising patterns than algorithms are, and that this gives discretionary traders an edge over automated strategies. While this may be true for anything that you or I may automate, it's probably not true when a trader finds themselves on the other side of some thing put together by, say, RenTech. All the indications are that professional A.I. trading has become very, very sophisticated.

 

As a side note, Neural Nets specifically apparently aren't that popular anymore - A.I. trading moved on to things like genetic algorithms, modelling hidden markov chains, kalman filters, and a whole host of other things that involve the back-propogation of error. Unless you really know what you're doing, Neural Nets are just a complex method of curve-fitting.

 

For an easy and casual read on all this, I recommend Scott Paterson's "Dark Pools".

 

BlueHorseshoe

 

ps The direct answer to your question is that I have never made money trading with Neural Nets

 

What I find dismaying shoe, is that curve fitting itself would not produce countless winning systems ,simply because patterns repeat. So, let us for the moment use only the past and the present. Lets say we had a neural network that calculated 50 trades in the past with a return of 14% on those trades. Now let us say we use the same criteria and come up with another 25 trades during this week, the present. Why would we assume that the future, would hold a different result. Well everything is in flux,true. But why would we assume the result would be different enough for us not to want to trade the system!!!? Yet as I admitted in my purchase of a neural software, the whole curve didnt just change, it fell apart completely. (That will leave us to discuss another subject of "Are all systems destined to fail" ) ) an important question to be sure...for another time. -)

 

But no she, getting back to the HFT's, I never mentioned them, you did. Because I dont consider them in the same light as Linda Rascke speaks,nor where there HFT'S around very much probably, when she even made the statement and might not have even known what they are. No, to me, HFT's are just a system within itself...........that may also be destined to fail ! Because in the cutthroat world of trading, do you not think a group of traders and programmers are working on something to counter that or suck the advantage out of HFT's? How about the flash crash we had last year? Could that be one example where all those HFTs profits of many weeks or months disappeared into one syndicates pocket in one day. And it is not a repeatable event for reasons both legal and for security. Do I think the HFT'S are going to one day implode as did the great Internet boom of the 90's did..."after" making many people rich............absolutely. So to me shoe, Expert Advisors and Neural networks arent related to Hft's which is more like just ultra fast scalping to me. With a topping of specific programming. Where is the "art" the visuals on a chart as any real system has? I think HFT'S are making a lot of software sellers rich with claims of how you can "do that too." And book writers who write about them as if it was a conspiracy on the level of the Illuminati and that whole secret society stuff which may or may not be true...to a point, but it rings the cash register. I think there are far fewer of them than we think or price in stocks would never fluctuate more than 1-2% on any given day. And we wouldnt have the s&p returning around 16% for the year. The HFT's would have stolen half of that already. Right?

 

Oh, dont leave out the conspiracy that China is allowing our dollar to stay stable while they stockpile all the gold in the world which is going to $10,000 an ounce. And next to that book I have one that assures me that we have more oil than we could go thru in 100 yrs, and its next to my other book that tells us(by a professor of geology!!!) we are being lied to and oiul will run out in our lifetime and govt doesnt want to create a panic. LOL Never has the media/Internet/Forex/stock/futures markets been such a well oiled machine of bullshit, hah? -)

[BlueHorseshoe 10]

But no she, getting back to the HFT's, I never mentioned them, you did. Because I dont consider them in the same light as Linda Rascke speaks,nor where there HFT'S around very much probably, when she even made the statement and might not have even known what they are. -)

 

I last heard Raschke make this statement in a TradeStation webinar about four weeks ago.

 

I mention HFTs because that is probably where the bulk of self-learning algorithmic trading is focussed. Nadex is a good website to visit to learn more about what these things do and why the need to use things like Neural Nets to continually adapt and evolve.

 

BlueHorseshoe

[michael45 10]

It makes you wonder how anyone can make money when you put it like that

 

It's like we're up against The Terminator.

 

By the way, thanks for mentioning Scott Paterson's "Dark Pools" - that's another book for my library

 

Where I can access this book easily ?

[Vince50 10]

Where I can access this book easily ?

 

Shoe, Please allow me the pleasure of answering this young mans question (smile). LIKE EVERY OTHER BOOK IN THE WORLD THAT YOU WANT....YOU GO TO AMAZON.COM FIRST. ALWAYS CHECK THERE FIRST. The book is $18 new and in the store its $27.00

So where else would you want to buy it. Im not yelling at yuh, Im just being a little silly. hA! HA! :haha: Buy yeah try amazon, but also becuase amazon.com lets you read some of the pages free before bying and with this book, they give you a loooooooooooot of free pages. Almost 1/4 of the book it seems. I started reading it myself. Now...........

 

Shoe, ive worn your ear out enough, so you dont have to read further. im going to teach steven some of the games i watched being played when I was a broker.,You might know the rest of what im going to say.

 

I dont know if the book gets into it, and I know about many of the games they mention in there already, having worked on wall street myself. But back in the 90's they had what they called SOS Bandits. You remember that term shoe? =) Liker 100 yrs ago. Anyway, that stands for Small Order Execution specialists and they would basically look to do manually and by phone, what these HFT's do automatically. I can say more but basically if you could always buy on the bid or in between the bid/ask, and just 50% of the time get rid of your stock on the ask, would you not make a boatload of money. Well, thats what those guys did.

 

BUT, THAT IS NOT TRADING IMHO. Its just being a manipulator. Thats a job. Thats real work. That isnt outwitting anyone the way we all are trying to do it here. What pleasure can they have that we get by watching a pattern unfold and being on the right side of it! It may also be on the thin line of whats legal and whats not as well. I would never get involved. I wouldnt be shocked if they had organized syndicates,bad people involved in that stuff. In fact Lucky Luciano when he was given a tour of the NY stock exchange, commented "I think I joined the wrong mob," God knows what shady dealings going on every moment that they let him in on !

 

Steven, I had a friend who worked for a firm that put a stock out at .49 bid-..50 offer . They paid their brokers 20% of total sales. How is that possible? Because as my friend tells me, whenever someone wanted to sell, they held up orders until all the selling came in in the same 10 minutes and the stock went down to .02 on the offer!!!! A week later it was back up to .49 bid again. I dont know how my friend slept at night but I say these hft's do the same thing with more liquid bigger stocks but do it a penny or 2 at a time and so fast you cant see it. Before it was dome by phone calls, now its done with phone calls and computers. and then maybe another group just computers.

 

Thought of the day: I always wondered, if you could find a way to make 1/4 penny on every share trading at 5 million shares or more per day, thats $ 50,000 a day. So if a handful of syndicates figured "my fantasy" out lol it still isnt that much money if done on 5-10 stocks per day. In forex it wouldnt even be noticed!

 

As I told shoe, steven , with the market up 16% the last 12 months, how much can they really be stealing ,relative to the amount traded? I will say this, I think the idiots who spread this info scared a lot of big money into bonds and safer investments and thats where the real damage was done!!!

[Perrin 10]

That's quite a big post for a simple answer but I enjoyed it all the same.

[SBGTrading 10]

I didnt find Ninja Traders tools or back testing stuff to be easily usable.

 

There's a very usable tool for backtesting any strategy in NinjaTrader...it's called ProphetViewer, and available at Prophet Viewer |

[Predictor 11]

Vince,

 

The way you pose your question is almost insulting. You basically want to know if someone with no effort, no skills, and no ability can make big bucks by pressing a button and generating a system? And, you probably know the answer to that: probably not.

 

>>I am not a programmer and not that Internet savvy.

 

I always say stick to your strengths. Given you can't program and aren't computer savvy then it doesn't sound like this approach is for you.

 

>>I ask because trade station programmers wouldnt be getting $100 an hour to program >>whatever you want if backtesting thru coding really was the answer. They would just >>code themselves up a neat simple winning system, no?

 

Yeah, and I guess the myriad of trading firms wouldn't be able to hire programmers either if that were true nor would they offer them big fat 200k/year salaries. No? They'd just code themselves up a winning system.

 

>>Also, what is all this backtesting tools with Ninja Trader and MT4. ....

 

Not sure what you tried to say but I couldn't understand that.

 

>>couldnt understand why it wouldn't work either thru massive winners or massive losers. >>Has any human being done well with a Neural network program/software or is it all >>junk?

 

Sure, there are many firms who use such methods and make huge fortunes. Is that what you wanted to hear? I'm sure its true but don't know any personally. Most of the CTA's and funds use systematic trading strategies of one sort or another and many uses machine learning techniques including neural networks.

 

It sounds like you're really asking.. is it worth my time to learn this stuff? Well only you can answer that. If you're really having success already and aren't driven to learn new things and don't have partner then you're going to be at a disadvantage when it comes to conducting research. Your best bet might be to subscribe to a research/quantitative newsletter or subscribe to another's trading system. Some of the quants offering information/research are Quantifiable Edges, MarketTells, Birinyi,

 

--

Curtis

OrderFlowDashPro - Enabling Trader Success

Edited December 8, 2012 by Predictor

[BlueHorseshoe 10]

 

I dont know if the book gets into it, and I know about many of the games they mention in there already, having worked on wall street myself. But back in the 90's they had what they called SOS Bandits. You remember that term shoe? =) Liker 100 yrs ago.

 

Hi Vince,

 

Sorry - I only just stumbled across your reply!

 

The book (Dark Pools) goes into quite a bit of detail about the SOS Bandits, yes. Basically, the original bandits operated out of the same office as the programmer who put the ISLAND exchange together, so they arguably had an unfair advantage in exploiting it, and probably about the best collocation in history (on the shelf above the exchange server!).

 

They were a shady bunch by all accounts, and I think lots of them found the way to jail about the same time that ISLAND went legit.

 

In terms of your later point, it's exactly the sort of 'loophole' that I continue looking for - I'm not fussed about the ego satisfaction of watching a chart pattern unfold having 'predicted' the market ket correctly - give me bucket loads of cash and I'll be happy

 

What did Sekoyta say - "everyone gets what they want from the markets".

 

BlueHorseshoe

[nelsonhimes 10]

Its hard to justify which are better but for that you have to analyses the things and the factors behind it. There are many books available in the market from where you can get better idea and knowledge about it.

[Vince50 10]

Hi Vince,

 

Like so many questions that are asked on TL, I'm afraid that the answer is probably "yes . . . and no". There are undoubtedly a handful of firms (mostly, though not exclusively) HFT, making a hell of a lot of money using A.I. approaches, but I would guess the number of retail traders who successfully apply anything of this sort are very small in number. I would guess that no off-the-shelf software of this type will give you any consistent and significant profits. That's just my opinion though.

 

 

As a side note, Neural Nets specifically apparently aren't that popular anymore - A.I. trading moved on to things like genetic algorithms, modelling hidden markov chains, kalman filters, and a whole host of other things that involve the back-propogation of error. Unless you really know what you're doing, Neural Nets are just a complex method of curve-fitting.

 

For an easy and casual read on all this, I recommend Scott Paterson's "Dark Pools".

 

BlueHorseshoe

 

ps The direct answer to your question is that I have never made money trading with Neural Nets

 

Shoe, you wrote the above. We have had many interesting back and forths. I admire your opn mindednes. But as for what you and most people say about curve fitting, I think there is a lot ot be said for its predictability value over any random indicator could ever have.

 

Lets look at it "on the street." If I am a policeman or just work in a high crime area and I notice that out of 10 stick-ups and car jackings I witnessed in the past year, only one of the perps wore any kind of clothing that you could even call "appropriate," and none of the perps wore a collared shirt nor a tie, or even shoes.

 

Is it not in my best interest to instantly curve fit this data so that when Im driving my car in the same area(or maybe ANY area!) That my reaction should be very, very different to a guy who yells at me from the corner while Im stopped at a light "if I know where Joes Pizza Parlor is" if his pants hang down to his butt and he has 4 gold teeth in the front and sneakers with the tongue defiantly sticking out compared to a man with suit, tie, briefcase, and polished shoes, and no gold teeth...IRRESPECTIVE OF THE DEMEANOR OR MOOD THIS PERSON SEEMS TO BE IN?

 

I LIKE CURVE FITTING...IT KEEPS ME ALIVE. The eager to please politicians and naive activists have a name for this curve fitting.They call it "profiling" and unfair.Makes people feel singled out,they say. I call ignoring it-- ridiculous and harmful to your health. What amazes me is..the same people who bring this subject up are never people who walk in a "bad area" with a suit and tie. They have the sense to blend in! Why dont people who walk in decent areas also show the same respect. Ohhhhhhhhhhhhhhhhhhhhhhhh...the constitution argument...again. &&^%$#@!

 

Anyway, I am a firm believer that if when Bernanke talks, the market has gone down the last 5 times out of 6, ten minutes before he started speaking, you can bet money I will not be in any short term longs until his speech is long over. Am I racial profiling Jewish men with grey facial hair? Doesn't matter,does it? It keeps me safe from blowing my acct! End of story.

 

Also, I hear that many people who are in commodities have made a fortune curve fitting seasonal stats that fit each grain or anything growing from the ground. Books have been written on it. Not my cup of tea, but I respect it greatly none the less.

[Perrin 10]

I've finished reading Dark Pools, so I just wanted to say thanks to BlueHorseshoe for the recommendation. For me, it was useful in getting to know a bit more about the trading world that we inhabit.

 

I'm still wondering (but I am less concerned now) about the impact and implications of algorithims and High Frequency Trading making buys & sells with no human participation.

 

I realise that humans create these programs (at the moment). At the moment, humans doing swing trading don't need to worry about HFT so much. At the moment. Although perhaps they do make the markets more choppy? I'm not sure. But will it always be the case that swing traders don't need to worry? I don't know. Things change over time. I'm sure the pit traders considered themselves to be essential. But that changed. In other industries, before machines were commonplace, huge numbers of people were employed farming the land. Or making clothes. Now those people are not requried, certainly not in the quantity they once were.

 

So can this happen to trading? I imagine most traders (myself included) will come to the conclusion that day trading or swing trading is still a way to make money. Perhaps scalping is not a way to make money for retail traders, unless using slightly longer timeframes - but would that be called day trading? Perhaps scalping is only viable (as a long-term career) if you own a computer program with your computers situated next to the exchange (the HFTs). I don't know.

 

Some people might say that scalping was never a good way to make money, in the long run, even before HFT? I don't know enough about scalping to have a view strongly either way, but if this is the case then perhaps nothing has actually changed regarding the way us small retail traders can make money in this business.

 

The way I read it, is that the algos and High Frequency Traders have taken the place of the pit traders.

 

What I read reassured me that I will still be around for a while yet, buying & selling. However my grandparents and parents all had to adapt to the changing world in their jobs and lifestyle. So perhaps I will have to adapt in the future in order to make money in the stock market. Although I cannot see how I will have to adapt too much in the way that I personally trade, as long as the market continues to flow based on human nature & emotion.

[BlueHorseshoe 10]

 

I realise that humans create these programs (at the moment).

 

Hi Perrin,

 

Glad you found the book enjoyable - it was recommended to me by someone else on TL, but I forget who.

 

Some of the strategies, as you probably gathered, are a product of machine learning, although obviously these algorithms have to be created by a human and then given a selection of data to mine.

 

I think that HFT replacing Pit Traders pretty much sums it up, and I'm sure pit traders probably used to play games with one another in just the same way.

 

I also agree about swing trading; in fact, I think that the rise of HFT has probably made the markets more suitable for swing trader. Though algorithms also occupy this space, the advantage of speed becomes irrelevant - trading from daily bars all that really matters is the effectiveness of the strategy. Or that's my take on it.

 

BlueHorseshoe

[syswizard 10]

I believe that using AI and Neural Networks and Genetic Programming have all fallen to the wayside because of curve-fitting. I guess the issue is: how do you avoid it using these sophisticated techniques ?

IOW: How can you teach your neural net NOT to curve-fit ?

I don't think it can be done.

[BlueHorseshoe 10]

I believe that using AI and Neural Networks and Genetic Programming have all fallen to the wayside because of curve-fitting. I guess the issue is: how do you avoid it using these sophisticated techniques ?

IOW: How can you teach your neural net NOT to curve-fit ?

I don't think it can be done.

 

Hi Syswizard,

 

If I had a definitive answer to that then I would probably be a far richer man!

 

However, I focus on minimizing the number of optimisable parameters within a strategy, by having parameters that produce acceptable returns regardless of value, and by employing fully recursive statistical measures rather than parameter dependent indicators. Another interesting thing to study is the predictability of shifts in the optimal parameters over time (for instance, is a more significant jump in the optimum more or less robust than a steady shift in the optimum?).

 

The general concensus amongst quantitative traders and mathematicians seems to be that optimisation is a good thing, whereas over-optimisation isn't - an easier concept to advocate than to put into practice in my opinion. Eckhart refers to the 'optimisation paradox', whereby a historical optimum will likely perform slightly worse than the future optimum, but far better than a random value.

 

I do think that Neural Networks, at least in the forms I am familiar with, carry significant risk of curve-fitting.

 

BlueHorseshoe

[UrmaBlume 10]

I believe that using AI and Neural Networks and Genetic Programming have all fallen to the wayside because of curve-fitting. I guess the issue is: how do you avoid it using these sophisticated techniques ?

IOW: How can you teach your neural net NOT to curve-fit ?

I don't think it can be done.

 

Over thirty years ago when DARPA had first allowed developers to make Neural Networks and certain other tools available to the public, I bought N-Train, a NN development kit offered by Dr. Jeffery Katz. I developed a model to trade the Swiss Franc.

 

The model produced answers that were very close to perfect. I hooked it up, ran it and it never made a winninhg trade. Thus began my leaning of the effective application of data mining, artificially intelligent tools, feedback mathematics and genetic optimization.

 

Experienced users of such technologies refer to curve fitting as an instance of the technology memorizing the data rather than generalizing rules from the data.

 

There are several ways this memorization can be avoided. One is to have enough data - some say there should be at least a couple of hundred records for each input to the model. Also at least 20 percent of the data should be unseen by the technology during development/training and then test the developed/trained model on the unseen data.

 

The degree to which the model performs more poorly on the test data is one of the ways to detect curve fitting.

 

Also the model should be built using different versions, branches and sets of these data mining/AI/feedback mathematics tools and test results should be similiar if concept and data contain inputs that truly offer a solution.

 

In any of these projects pre-processing the data is 80% of the engineering and 90% of the work. The greater/better the relevance, scaling and other preprocessing of the inputs to the target - the greater the chance of success.

 

 

cheers

 

UB

[BlueHorseshoe 10]

Over thirty years ago when DARPA had first allowed developers to make Neural Networks and certain other tools available to the public, I bought N-Train, a NN development kit offered by Dr. Jeffery Katz. I developed a model to trade the Swiss Franc.

 

The model produced answers that were very close to perfect. I hooked it up, ran it and it never made a winninhg trade. Thus began my leaning of the effective application of data mining, artificially intelligent tools, feedback mathematics and genetic optimization.

 

Experienced users of such technologies refer to curve fitting as an instance of the technology memorizing the data rather than generalizing rules from the data.

 

There are several ways this memorization can be avoided. One is to have enough data - some say there should be at least a couple of hundred records for each input to the model. Also at least 20 percent of the data should be unseen by the technology during development/training and then test the developed/trained model on the unseen data.

 

The degree to which the model performs more poorly on the test data is one of the ways to detect curve fitting.

 

Also the model should be built using different versions, branches and sets of these data mining/AI/feedback mathematics tools and test results should be similiar if concept and data contain inputs that truly offer a solution.

 

In any of these projects pre-processing the data is 80% of the engineering and 90% of the work. The greater/better the relevance, scaling and other preprocessing of the inputs to the target - the greater the chance of success.

 

 

cheers

 

UB

 

Hi UB,

 

Thanks for an interesting post.

 

In your opinion, given the caveats you discuss, is the "answer" to adapt and develop neural nets with these pitfalls firmly in mind, or to look for different approaches altogether?

 

Would you be able to list some of the AI processes you have experimented with?

 

Cheers,

 

BlueHorseshoe

[UrmaBlume 10]

Hi UB,Thanks for an interesting post.In your opinion, given the caveats you discuss, is the "answer" to adapt and develop neural nets with these pitfalls firmly in mind, or to look for different approaches altogether?Would you be able to list some of the AI processes you have experimented with?Cheers,BlueHorseshoe

 

Thanks for the kind words.

 

I work with a wide selection of tools that includes Neural Networks, Rules Generators, Decision Trees, MARS (multivariate adaptive regression splines), Genetic Optimizers and Selectors, Bayesian/probabilistic networks plus various tools for intelligent data extraction, scaling and general pre-processing.

 

Given the proper understanding of the tool, its data preprocessing requirements and the data domain - any of them, individually or in tandem, can be used to develop successful trading bots.

 

The point I tried to make in the post above is that it takes more than just buying a NN development kit and throwing the output of some indicators that don't work to begin with to make money using these technologies.

 

There are masters of technology that can't build successful trading models and there are trading wizards who can't build intelligent trading models. It takes way more than just professional expertice in both domains to have a chance at this game - it takes an artist at both.

 

Just like in trading in general, it is only a few that succeed.

 

"Success in the markets is not about instinct, divine inspiration or spontaneous intellectual combustion. It is about intelligent data processing, sound method and the management of risk and resource that is both effective and adaptive to change."

 

cheers

 

UB

 

 

Below is a picture of my desk where I track and train these little bots of mine: 4 dell workstations and 23 19" monitors