I’ve used it - but didn’t like the fact the params for the neural net are unknown.  Over optimization with unknown genetic algorithms is pretty dangerous in the hands of a non-mathematician - which most financial advisors are not.

Personally, I’m not a genetic algorithm expert - which is essentially what traditional ai and nn use for optimization.  The backtest results using their software for modeling is impressive if you are good at selecting and/or building indicators - which is pretty plug and play with TS.  If you use it, I would suggest a lengthy walk forward test (1 year or more) before putting $$ into the strategies.  And I would be very careful to optimize known factors (ie; ma, macd, bb, stochastics, etc), you’re probably better off developing some one of a kind trend measurements and mean reversion measurements and letting the software train to optimize those factors.

Good luck and let us know if you move forward and have some success.

Thanks for your feedback AC, I’m not entirely convinced that AI/Neural Nets or any of the overly complex technical analysis stuff is anything more than hocus pocus. I’ve always considered myself more of a fundamentals guy but I don’t want to outright dismiss something due to unfamiliarity. So I’ve been looking at the various software packages that have trial versions - I hear NeuroShell is better than TradingSolutions - to get a better grasp on possibilities. I am still unable to get my mind around the fact that you can build a model of an unpredictable future. Do you have any software that you recommend I check out? 

Hocus pocus; not so much.  But a lot of advisors will not get the results I think they seek with the use of off the shelf programs like those you mention.  The big problem is most people will just plug in technical analysis factors that everyone already knows.  The software fits the data to the factor and magically the out of sample tests look amazing.

Lastly, the future isn’t nearly as unpredictable as you think .  To see what I mean, just do a little test in excel that assumes you are long the S&P only the days after the market goes down and short only after the days the market goes up.  The last 10 years you’ll average nearly 20% per year with no negative years.  Very simple mean reversion.

Cheers,

JW

I was a CTA for a number of years, and can say neural nets and all that other stuff is garbage. Hard-core optimization like what AI models create are dangerous.   



I would go with Tradestation, and just keep testing with simple models. The results really depend on your portfolio mix and time frame anyway.

[quote=Otane]I was a CTA for a number of years, and can say neural nets and all that other stuff is garbage. Hard-core optimization like what AI models create are dangerous.   



I would go with Tradestation, and just keep testing with simple models. The results really depend on your portfolio mix and time frame anyway.[/quote]

Would you say that the Renaissance Fund that made 80% last year and about 84% per year for over 20 years using genetic algorithms is garbage?  Just busting your blz:-)

Most advisors probably should stick to something very simple like tradestation; but neural nets are only dangerous when used by people who don’t understand their shortcomings and/or trust stuff they couldn’t build themselves.

[quote=AdvisorControl.com]



[quote=Otane]I was a CTA for a number of years, and can say neural nets and all that other stuff is garbage. Hard-core optimization like what AI models create are dangerous.   



I would go with Tradestation, and just keep testing with simple models. The results really depend on your portfolio mix and time frame anyway.[/quote]Would you say that the Renaissance Fund that made 80% last year and about 84% per year for over 20 years using genetic algorithms is garbage? Just busting your blz:-)Most advisors probably should stick to something very simple like tradestation; but neural nets are only dangerous when used by people who don’t understand their shortcomings and/or trust stuff they couldn’t build themselves.[/quote]





Renaissance Fund is hardly neural nets. Their database is clean, tick data that has been compiled over decades. Very few firms had that capability at the time, and don’t have access to that type of historical data today.



I worked for a systematic company that had a group which had the first type of Renaissance Fund in the early 90’s. This group came from Bankers Trust and compiled forex tick data, and traded it over several years. One of the members of the group left to work for Renaissance later, so I have an idea of what is going on.



And, it is not 80%, but 40’ish percent NET per year - still an outstanding return. You are not busting my biz, but you need to know your facts when busting it.



This trend in hiring programmers is for arbitrage, and was started by companies like Susquehanna in the early 90’s.

Renaissance Technologies is a hedge fund management company. Started in 1982 by James Simons, Renaissance currently has approximately $20 billion in assets under management. <SUP =reference id=cite_ref-0>[1] Since 1989, the company’s $5 billion Medallion Fund has averaged 35% annual returns<SUP =reference id=cite_ref-WSJ_1-0>[2]<SUP =reference id=cite_ref-2>[3], after fees and its flagship Medallion Fund is viewed as one of the most successful hedge funds

I think the returns are higher than the wiki facts when I last talked to my buddy.

Could be. I just Googled it for kicks. I have no idea if it’s accurate or not.

[quote=Otane] [quote=AdvisorControl.com]



[quote=Otane]I was a CTA for a number of years, and can say neural nets and all that other stuff is garbage. Hard-core optimization like what AI models create are dangerous.   



I would go with Tradestation, and just keep testing with simple models. The results really depend on your portfolio mix and time frame anyway.[/quote]Would you say that the Renaissance Fund that made 80% last year and about 84% per year for over 20 years using genetic algorithms is garbage?  Just busting your blz:-)Most advisors probably should stick to something very simple like tradestation; but neural nets are only dangerous when used by people who don’t understand their shortcomings and/or trust stuff they couldn’t build themselves.[/quote]





Renaissance Fund is hardly neural nets. Their database is clean, tick data that has been compiled over decades. Very few firms had that capability at the time, and don’t have access to that type of historical data today.



I worked for a systematic company that had a group which had the first type of Renaissance Fund in the early 90’s. This group came from Bankers Trust and compiled forex tick data, and traded it over several years. One of the members of the group left to work for Renaissance later, so I have an idea of what is going on.



And, it is not 80%, but 40’ish percent NET per year - still an outstanding return. You are not busting my biz, but you need to know your facts when busting it.



This trend in hiring programmers is for arbitrage, and was started by companies like Susquehanna in the early 90’s.

[/quote]

They are using genetic algorithms (self adapting, proprietary in nature) to trade their tick data.  It’s different then an NN - which is why I prefaced the comment by mentioning genetic algorithms.  What they are doing exactly, is clearly beyond anyone else in the industry as no one has anywhere near their track record.

Re: their track record.

They charge a 45% performance fee and I believe a 4% flat fee.  So if you take their net of mid 30’s you’ll get a gross in the mid 80’s.  Per Absolute Return/Alpha magazine (this months issue) they did 80% net last year with that particular fund.  Not too bad.

So none of us are entirely right or entirely wrong - my only point was that there are ways to use math and data to get better results.  A lot is garbage - but not all of it; it seems to get that way when used by people who don’t know what they’re doing.

Thanks for the killer insight Otane - it’s good to read you’re posts.

JW

There is a reason why they are the only fund doing this - it is the cash data. You can get futures IM tick by tick from the CME but it only covers US hours, and trying to smooth out the data because of the forwards is very difficult.



Another reason why they are successful is they push the markets with Goldman helping them. Goldman is probably the only house to risk trading capital, and they are known to be bullies. Look at how they said Oil would go to $150, and they pushed everyone and their grandmother to that level.



As far as their fees, I don’t think it is that high.

[quote=Otane]There is a reason why they are the only fund doing this - it is the cash data. You can get futures IM tick by tick from the CME but it only covers US hours, and trying to smooth out the data because of the forwards is very difficult.



Another reason why they are successful is they push the markets with Goldman helping them. Goldman is probably the only house to risk trading capital, and they are known to be bullies. Look at how they said Oil would go to $150, and they pushed everyone and their grandmother to that level.



As far as their fees, I don’t think it is that high. [/quote]

Your right on the fees; they charge 5/36.  This is from Wikepedia:

Renaissance Technologies is a hedge fund management company. Started in 1982 by James Simons, Renaissance currently has approximately $20 billion in assets under management. <sup id=“cite_ref-0” =“reference”>[1] Since 1989, the company’s $5 billion Medallion Fund has averaged 35% annual returns<sup id=“cite_ref-WSJ_1-0” =“reference”>[2]<sup id=“cite_ref-2” =“reference”>[3], after fees and its flagship Medallion Fund is viewed as one of the most successful hedge funds.<sup id=“cite_ref-WSJ_1-1” =“reference”>[2]
In line with other hedge funds, it charges a management fee and an
incentive fee. Its management fee is 5% and its profit participation is
36%, both of which are higher than the industry standard of 2% and 20%.

This equals about 80% gross return for a net 34% fee.  Crazy good.

[quote=Otane]Hard-core optimization like what AI models create are dangerous.[/quote]

There’s always a question in system development. Will it hold up or fall apart when presented with new data? My company’s program, Genetic System Search for Technical Analysis (http://www.gssta.com/), includes terms in the fitness formula to combat bad curve-fitting. If I turn off those terms in the genetic programming engine, the over-optimization is dramatic.

Other terms in the fitness function address the practicality of the system. For example, there’s a term which derates the investment systems according to the number of sequential losing trades. You’re likely to abandon a system which signals five losing trades in a row. Similarly, can you accept a heart-stopping drawdown for the vague promise of a rich payday on the other side? Probably not.