This is as close as I can come to finding an answer for your question. It shows the actual economic return of several Garch variants compared to historical returns and compared to an exponential moving average. The markets covered include energy and metals, currencies and U.S. equity index and interest rates. Short *.pdf file of 15 slide-like pages. http://cisdm.som.umass.edu/resources/pdffiles/Georgiev CISDM 2003.pdf
Thank you. His analysis shows that forecasting volatility using Exp. Weighted MA can produce higher return with the same St. Dev. than using GARCH. There're no details how he calculated this.
I'll assume you've had your drink for the evening and are enjoying yourself...if a system is predictable i.e. a cyclical market then what does that mean? It means that if you know the variables that control the transitions from one state of the market to another state, in other words if you know what creates the cycles, then the market is just a mechanical process and like a mechanical process, the market can be predicted. There is no choice involved, because market states result from parameter changes in the controlling variables. Well, if that is the case then it is already determined what the market will do tomorrow, doesn't it? Not exactly. The market may be a mechanical process and may in some ways be predictable. However, because it is affected by so many variables it produces random, unexpected behavior from time to time. This wreaks havoc with prediction.
Are you sure? On page 5 he describes how ARCH processes are used. He compares (through regression) the results of a statistical analysis of the historical variance (volatility) of returns to the returns predicted by both techniques. To do the same thing all you or I would have to do is find a set of closing prices, calculate the difference from day to day and use that as our 'return series'. All page 5 is saying is that once you have that series all you have to do is find the average of the series over so many days then compute the standard deviation over the same time frame for all the different models. Do that for all models for a number of different time frames. Once that is done, compare the models' results for all the time frames. Each model will produce a curve based on that model's average and standard deviation for different time frames. Use a linear or other regression calculation to find the difference between the predicted value and the historical value. If the level of significance is fairly low, as it appears to be in most cases, then we can reject the hypothesis that modeling has little economic impact. Looking to the charts...if for each data point you divide the standard deviation by the expected return the result is called the coefficient of variation or volatility at that particular point on the efficiency frontier. Ideally, what you want is "money for nuthin' and chicks for free". But, since you aren't, probably, going to get that it is best to get the highest return for the level of volatility in returns or "sigma" (he mentioned on page 5) you are willing to tolerate. I think I've got that covered. Hopefully, there aren't any mistakes. If you find some, please let me know.
Just wanted to say thanks for this thread in some way. Surprised myself with what I've been able to retain through the years. I've used this link by typing the topic titles into my search engine. http://www.gsia.cmu.edu/pfe/ITM/ffm.html
In one month I will begin to create a dozen short lessons on classical time series so as to demystify them and show when they are suitable or not.
Harry, Great idea! ET is finally going to have a worthy successor to Jack, Bubba & Grob. P.S. Try to squeeze in a few of those flashy colored graphs.
Here's a company which develops econometric models http://www.ht-ag.com/ Go to "trading area" > "performance chart". Commissions for non-eurex members can eat up to 12% of their strategy profits. Company states that there won't be any slippage. Substracting 1 tick per round turn would result in ~3/5 less profits. Click "Mathematical Expectations" - so far, real results differ much from expected... Hmmm I think it won't be any problem with developing a system without econometrics which beats X-ROI assuming zero slippage&commissions.