Forums  > Trading  > Price patterns  
     
Page 1 of 6Goto to page: 1, 2, 3 ... 5, 6 Next
Display using:  

starchild


Total Posts: 90
Joined: Nov 2007
 
Posted: 2008-03-13 21:50

This paper shows that some candlestick patterns on daily stock prices have predictive power:

Caginalp & Laurent, "The predictive power of price patterns," Applied Math Finance, 1998.

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=932984#PaperDownload

I want to do some reading on finding patterns that have predictive power. Any recommendations? (I have a EE signal processing background).

It would be really cool to come up with an algorithm that systematically searches for such patterns.

Thanks!


When I see a bubble, I buy it, because that's how I make money. -- G. Soros

urnash


Total Posts: 564
Joined: Sep 2006
 
Posted: 2008-03-14 08:51

First of all, welcome to NP.

By using the search button on this site you can find quite a lot of discussion about technical analysis. Most people laugh about it, others say that it is just another model. You could also have a look at this thread, and at the book Evidence Based technical analysis by Aronson (which I haven't read). HTH


Assuming normality means never having to say you don’t have enough data. -- J. Michael Steele

starchild


Total Posts: 90
Joined: Nov 2007
 
Posted: 2008-03-14 09:57

I understand that most technical analysis is BS. What I liked about this paper is that the authors give rigorous definitions for what it means to be a certain candlestick pattern, then test the predictive power. According to the resuts, the predictive powers look non-negligible. I wanna test it myself sometime.

I saw two broad classes of trading strategies in academic literature. One group of papers try to estimate market/stock behavior, go long or short each and every day, and beat the market (typical approaches are NN, HMM etc). Another group of papers try to find some patterns that lead to predictable behavior (strategy: get into the market only if a predictable pattern emerges, stay out otherwise).

Whatever I tried along the lines of the first approach failed (even a few papers that claim to have beaten the market look suspicious). The latter one makes much more sense, and I'm very curious if there are some patterns that indeed lead to predictable behavior (after all that's what [some] hedge funds do, right?).

Anyways, thanks very much for the suggestions! I'll look at them more carefully.

 


When I see a bubble, I buy it, because that's how I make money. -- G. Soros

svquant


Total Posts: 113
Joined: Apr 2007
 
Posted: 2008-03-14 16:54
Both methods work - pattern based trading and forecast based trading, i.e. NN, HMM, ARIMA, filters. Of course both methods will also fail if you are not careful, have incorrect inputs, and just datamine and overfit. Of course figuring out that last part separates those who make money in systematic trading from those who loose.

As for pattern based trading Trout got started that way and Crabel too. Both very successful. Some academic papers based on those forecasting techniques do not always implement a stop-and-reverse trading strategy. Perhaps look at modifying the techniques in this way and you might turn an unsuccessful academic system into something more?

Be very careful in terms of price patterns and datamining - I've seen many a trader get burned bad when they did not realize they overfit the data and did not have a general behavior based pattern.







MrMagoo


Total Posts: 192
Joined: Jan 2008
 
Posted: 2008-03-14 18:08
Actually, naive extrapolation of past prices (most of technical analysis) yields the same in-sample results of some behavioral finance market anomalies ie. short-term serial auto-correlation.
The problem imo is that you don't get a picture safe enough to trade. For instance, if you assume a car driving over 50 mph you would reason by induction it will be driving over 45 mph a moment later. That may work in a perfect long road, but this simple heuristic tells nothing about the chance of curves, crashes, traffic signals, holes, drunk drivers, storms, etc.

"One who says it can't be done should not interrupt the person doing it."

starchild


Total Posts: 90
Joined: Nov 2007
 
Posted: 2008-03-18 09:12

It's easy to get discouraged that systematic trading based on daily data is impossible. Your comments and the paper gave me more motivation to experiment. Thanks for the insightful advice!

PS: svquant, I sent you a PM (email), please let me know if you haven't received it.


When I see a bubble, I buy it, because that's how I make money. -- G. Soros

intradaybill


Total Posts: 110
Joined: Mar 2008
 
Posted: 2008-03-24 16:22

starchild, very interesting article. Check out also this more recent peer reviewed one:

http://www.westga.edu/~bquest/2008/candlestick08.pdf

I'm looking at a program called APS Automatic Pattern Search that discovers price patterns. Candlestciks are a special subset of price patterns. I think this is also an interesting article:

http://www.tradingpatterns.com/About_Us/articles/article3/article3.html

Bill

 

 

 

 

 

 

 


schmitty


Total Posts: 57
Joined: Jun 2006
 
Posted: 2008-03-25 07:13

IDBIl wrote:

> "I'm looking at a program called APS Automatic Pattern Search"

That program is a very crude version of the Floating Hypercubes method of Norm Packard and then grad student Tom Meyer at U of I at Champaign-Urbana around twenty years ago. You are better off with the original.

Harris brings about one tenth the brain power of Packard and Meyer to the problem.


starchild


Total Posts: 90
Joined: Nov 2007
 
Posted: 2008-03-25 07:52

I guess you meant this work of Meyer and Packard:

Local Forecasting of High-Dimensional Chaotic Dynamics
http://www.ccsr.uiuc.edu/TechRep/1991/CCSR-91-1.pdf

Will take a more careful look at that and the other paper on candlesticks.

 

On a related note, I've been finding these lecture notes on

Statistical Learning Theory by Rob Nowak to be very readible and interesting:

http://www.ece.wisc.edu/~nowak/SLT07.html


When I see a bubble, I buy it, because that's how I make money. -- G. Soros

temnik


Total Posts: 204
Joined: Dec 2004
 
Posted: 2008-03-25 16:15
Technical analysis does work, just not in the ways most academics imagine it does.

Their analysis seems to lack context in which all those patterns appear. At the simplest level, three-day rally described in the original article could be suggesting a buy or a sell trade depending on what has happened longer-term, and what the existing position (if any) is.

Mon métier et mon art, c'est vivre.

svquant


Total Posts: 113
Joined: Apr 2007
 
Posted: 2008-03-26 16:56
While the Meyer and Packard work is interesting and good - it really didn't pan out for the Prediction Company from what was the scuttlebutt back in the day. Of course this could have been due to the input space used, the signal vs noise in financial data, or the algorithm itself.

Yes, technical analysis works and buy and large it appears simple is a better principal (Occam's razor sort of thing) if you look at the successful traders that acknowledge use of technical analysis.

If you are interested in machine learning one of the most successful managers that publicly admits to running their fund based on such an approach is QIM in Virginia.


intradaybill


Total Posts: 110
Joined: Mar 2008
 
Posted: 2008-03-28 09:59

svquant,

QIM performance is very good:

http://www.quantitative.com/performance.html

Bill

 

 


pepper_john


Total Posts: 45
Joined: Jun 2006
 
Posted: 2008-03-29 03:03

Yeh, QIM looks impressive. How much is under management?


nikol


Total Posts: 576
Joined: Jun 2005
 
Posted: 2008-03-31 10:47

Searching patterns is pretty easy.

You should define in formula what a pattern is. Let's make example of "head-and-shoulders" (H&S). Surely, I am not 100% correct, but lets try. 

Define time window, find a maximum within window (find respective time_max). time_max will split window into 2 sub-windows (up-slope an down-slope).

We can have difficulty on how to define "shoulders". I propose this: build 2 historgrams of prices within 2sub-window, hist#1 (prices(up-slope-window)) and hist#2(prices(down-slope-window)).

Maximums from two histograms whould give indication for shoulders. To confirm this, you have to validate those shoulders, that they are result of price swing in up- or down-slope.

Check some found patterns with an eye (although subjective, eye is the most powerfull pattern recognition tool) and voi la - you have your pattern recognition defined mathematically. Acquire some number of time series from Yahoo, Reuters, Bloomberg and run analysis.

Quastion to answer - is it true that when H&S pattern has been formed, the price will drop in value? And if you find reverse H&S will price go UP? Is there any correlation?

If an answer is yes, I would start searching an investor rather than a publisher... Because if you publish it, the correlation will go away.... But you may find an investor (e.g. employer).. who will belive that you know how to do pattern recognition technique in general...

 

 


svquant


Total Posts: 113
Joined: Apr 2007
 
Posted: 2008-03-31 17:37
Yes, QIM is very good which is why I pointed them out. One can be successful with what appears to be from all public documents and interviews a pattern based machine learning and data mining approach.

With almost $3B under management this is very impressive. One thing I did noticed is that they make their money in many of the same markets as RenTec does but using (from what I can tell are) different methods.


gnarsed


Total Posts: 87
Joined: Feb 2008
 
Posted: 2009-12-16 03:48
the funny thing is if you look at the org chart they've posted on their website, you dont see any people purporting to have expertise in either machine learning or data mining.. and the technical team seems really small.

SimJimons


Total Posts: 198
Joined: Aug 2007
 
Posted: 2009-12-16 08:39

To my knowledge it's one guy (the boss, GW) that has come up with every piece of the system they (QIM) use. I didn't know it was relying on machine learning, but I guess that could be the case (even if I believe that it's more basic/classical stuff involved). Anyway, one cannot compare them to Rentec. QIM is a short-term (about 6 days on average) CTA, Rentec is a liquidity providing GodSmiley A very, very different thing. Having said that, QIM is really good, with low corr to most other CTAs. They manage north of $5 bln today (I even saw one number saying $7+). Massive!

/Sim


Sounds great...keep me out!

intradaybill


Total Posts: 110
Joined: Mar 2008
 
Posted: 2010-02-09 15:43

How is QIM doing nowadays, anyone knows? The link to their fund performance below is not active ny longer and they have password protected their website.

http://www.quantitative.com/performance.html

Update about pattern search and APS : I have purchased the APS program and I can tell you it was one of the betetr investments in software I have ever made. Since then the company came out with a new version that includes an indicator some people I talked to claim it is quite interesting and novel. I will think about upgrading after I here some opinions about it. Here is the link to their article:

http://www.tradingpatterns.com/The_p-Indicator.pdf

 

 

 

 


SirAppleby


Total Posts: 182
Joined: Dec 2006
 
Posted: 2010-02-10 01:13
Like almost all systematic CTAs, QIM is in draw-down (approx -8%) at end of January. It seems that everyone is struggling with financials, and commodity trading is not offsetting the losses. Perhaps 2010 will test the models that performed so well in '07-'08?

Patience is necessary, and one cannot reap immediately where one has sown.

Maltese


Total Posts: 51
Joined: Aug 2005
 
Posted: 2010-03-11 23:54
Any more news on how QIM is doing?

Thought is only a flash between two long nights, but this flash is everything. Henri Poincare.

amateur


Total Posts: 147
Joined: Mar 2010
 
Posted: 2010-03-24 14:28
QIM was down last month and still down in March (about 0.6%).

“unnecessary complex models should not be preferred to simpler ones. However . . . more complex models always fit the data better”

svquant


Total Posts: 113
Joined: Apr 2007
 
Posted: 2010-03-31 19:42
Did a quick search for the documents I had that talked about QIM and "machine learning" type of pattern development but couldn't easily find it since it was an older document - well before QIM hit the $1B AUM mark.

Did pull a version of a 2010 pitchbook and here are some tidbits:

So you can sort of see how I got to my conclusion abouth their techniques.

Other interesting tidbits include that they have expanded the number of models from 1500 to 5000 (in mid 2009), average trade length appears to have increased a few days from what I remember (it is now ~7.5 days), and they appear to have reduced the number of markets traded actively which makes sense due to the large increase in AUM.

As for the number of models I know many shops that say they have "150 models" if they really have 3 base ideas/models on 50 markets. Other firms call this 3 models... unclear what QIM is doing but with less than 50 active markets - this would still imply 30-100 patterns per instrument.

gnarsed


Total Posts: 87
Joined: Feb 2008
 
Posted: 2010-04-01 00:10
to me, one guy writing 99% of "predictive code" is a red flag, no matter how brilliant that guy might be... especially with billions aum. but what do i know.

amateur


Total Posts: 147
Joined: Mar 2010
 
Posted: 2010-04-08 01:29
http://www.finalternatives.com/node/12042

QIM Takes A Hit In March, First Quarter

Quantitative Investment Management has not started the new decade off on the right foot.

Three of the firm’s four hedge funds finished the first quarter of 2010 in negative territory. Its flagship Quantitative Global Program, with $4.6 billion in assets, is down 5.74% after shedding 1.02% in March. Its Quantitative Global Fund 1x is a little further down, with a 5.81% drop in the beginning of 2010 following a 1.06% loss in March.

Unsurprisingly, its three-times levered version of the latter fund is really down in the dumps. The $348 million fund fell 3.07% in March and is down 16.76% on the year.

The funds’ interest rates investments took the biggest hit in March, accounting for nearly three-quarters of the decline. Grains and soft commodities also hurt the funds.

But, according to QIM, the bad news is actually good news.

“On March 12, the Global Program reached a 10% decline from its September peak, triggering a third reduction in risk which brought our exposure to 25% of its normal level,” QIM wrote to investors. “This is the second time in our track record that the 10% drawdown level has been hit. Despite the last severe drawdown, which occurred in March 2004, the Global Program still managed to return over 23% that year.”

The first quarter was not all bad for QIM: The firm’s two-year old, $482 million Quantitative Tactical Aggressive Fund rose 1.09% in March, and is up 4.95% on the year.

“unnecessary complex models should not be preferred to simpler ones. However . . . more complex models always fit the data better”

amateur


Total Posts: 147
Joined: Mar 2010
 
Posted: 2010-05-10 01:26
FIN Alternatives
07-05-2010
Things improved—marginally—for Quantitative Investment Management in April, and just barely, as
well.
The firm’s flagship Global Program “surged” 0.91% in the last week of last month, bring its April return
to 0.06%. Unfortunately for investors in the $4.4 billion fund, it’s still down 5.68% on the year, thanks in
no small part to its worst-ever drawdown in March.
“Strong sell-offs in the US equity markets, which also saw a surge in volatility, drove profits,” QIM said
in its monthly performance report. “The modest gain of 0.06% for the month ends a stretch of four
straight losing months, the longest in the program’s history.”
The struggles certainly extended to the other versions of the QIM flagship. It’s 1x fund rose 0.08% in
April (down 5.7% year-to-date) and its three-times levered 3x fund added 0.25% (down 16.52% YTD).
By contrast, QIM’s Quantitative Tactical Aggressive Fund soared 9.33% on the month and is up
14.83% on the year.

“unnecessary complex models should not be preferred to simpler ones. However . . . more complex models always fit the data better”
Previous Thread :: Next Thread 
Page 1 of 6Goto to page: 1, 2, 3 ... 5, 6 Next