I've posted similar questions on Wilmott forum, but I didn't get what I wanted. Maybe ET'ers will know more. Do you know of any trading system (excl. arbitrage) and it's historical hypothetical or real performance on most popular markets (futures, stocks, forex; in 1-30 min. intervals) based on GARCH, ARIMA, Bayesian analysis, Kalman filtering? I wonder whether these tools can outperform classic technical analysis tools. There's a lot of academic research on these models, however their robustness in the real trading isn't described. I found some research by Olsen http://www.olsen.ch/research/workingpapers/319_real-r1.pdf but it only deals with FX 1990-1996 and trading costs aren't specified. Performance is poor when compared to equity futures systems. Somewhere on the internet I found that Mr. Pierre Lequeux made performance analysis on Dax and Nikkei, but I can't find these papers via google, many pdf's are written in french. When academic people apply econometric methods into the financial data series, their conclusions are always related to volatility or many statistical properties. I would like to see simple figures like P&L, drawdowns, Sharpe ratio, profit factor...
Norman Fosback wrote a book back in 1976 that was interesting. He didn't divulge much, prefering that you buy the results of his econometric models via newsletters. I read the book through several times. He has since moved on...http://www.fosback.com/. Alternatively, you might do a search for Arthur Goldberger who wrote several books on econometrics. Box and Jenkins wrote an important book on time series forecasting...probably, you know all this. If you're looking to refresh some things try this link and read the last few chapters...http://books.iuniverse.com/viewgiftoc.asp?isbn=0595142990&page=1 Once you get into the main body of the book instead of the Table Of Contents just plug in the page number where you want to begin reading in your address bar. Sorry I couldn't answer your question.
The very basic reason why stock market time series are reputed to be one of the most difficult arena of forecast is because these classical time series techniques don't work, mathematically these techniques are based on autocorrelation of errors since these autocorrelations are very low in stock market time series they are not worth at least used in traditional way. Now low autocorrelation is not equivalent to independancy, it has been showned for a long time since Mandelbrott that the Market exhibits "long term memory effect" so that the latest kind of stochastic model taking into account that effect is ARFIMA's model. But the performance still is poor. All in all I say it is an error to use stochastic models to do market's forecast because only a deterministic model can do it (ie mine of course ), the problem is to find it and the reason that researchers didn't find it is because they try to extract knowledge from pure datas which is an idiocy from paradigm point of view because the model's knowledge is transcendant to the datas that is to say you cannot deduce it from the datas alone but only if you have the idea of how market really works or you will play with datas and stochastic models much like a monkey see: http://www.elitetrader.com/vb/showthread.php?s=&threadid=28614 ANNs (Neural Net): A Little Knowledge Can Be A Dangerous Thing http://www.secondmoment.org/articles/ann.php ANNs: A Little Knowledge Can Be A Dangerous Thing Posted by Dr. Halbert White
So a new inspirational quote everyday. For Tuesday February 24th:<br><font color="green">"I have been struck recently by the disconnect between the worldview expressed by these economic and finance papers, and the view that I was seeing by standing on trading floors and talking with investment professionals"</font><br>-- Beyond Equilibrium and Rationality [by J. Doine farmer - Santa Fe Institute]
"I have been struck recently by the disconnect between the worldview expressed by these economic and finance papers, and the view that I was seeing by standing on trading floors and talking with investment professionals" I have a similar impression. They're using entirely different languages. Harry, for the first time I understand something from your post Thanks. I've found another paper http://www.if5.com/papers/Constructing_A_Managed_Portfolio_Of_High_Frequency.pdf Analysis uses GARCH and Wavelet Encoding A Priori Orthogonal Network (WEAPON - is this a little marketing trick? ) These methods are beyond my understanding. With 4 ticks per round-trip results are good, but for a very short period - only 6 months. How about >3 years...
DT-waw. Thanks for the link, I'll read it later. The problem I have with all these methods is they involve trying to discern the cyclic behavior of markets. I stopped believing in market cycles a long time ago. I don't doubt market cycles exist, I doubt that you can predict them in advance except in very short term time frames. I've done some study on this...in the end you must trust your prediction and play it or admit...YOU DON'T KNOW!
It can be done in all time frames, however if 4 bar tolerance is ok on 5 min charts and will not have any effect on stop size, the same is not true on daily or weekly time frame. Only well funded traders can play longer timeframe cycle timing game.
Oh, sure, and I'd never lose at Blackjack with deep pockets. What I read here is that you can time the market cycles as long as you have enough cash when you are wrong. This, to me, isn't "timing" and is no better than any other technique of analysis.
I stopped believing in market cycles a long time ago. It's not about believing, it's about showing some hard, statistically significant figures. in the end you must trust your prediction and play it or admit...YOU DON'T KNOW! React, adapt, rather than predict - key rule in trading. If I'm not mistaken, GARCH is used maily to predict the volatility? Neural Nets are used to predict returns. Why not transform these methods into a trading model and show performance stats. It would be nice to compare them with EMA or Jurik MA -based systems.
Could you define what you are asking? Volatility is an antipersitent time series, if it has been declining it is likely to increase and vice versa. If prices have been rising then volatility is likely to decrease at the new price level; however, bad news tends to increase volatility. So, since it is fairly difficult to predict the arrival of bad news with any regular frequency or cycle it is therefore difficult to time market cycles. So, what is the advantage of trying to time the market?