Trendfollowers: When oh when are we going to start making some $$$?

[asiaprop]

So, according to your definition of curve fitting, you should not trade at all. ANYTHING you ever tested and seemed to work is "fitted to the curve" and may stop working tomorrow. Why do you trade then? I think you have a slight misunderstanding of when one can conclude something is curve fitted and when not.

If you test a strategy and it works over a very large sampling space of price data for one stock but does not work for another stock, would you conclude this is curve fitted even though the standard error under the number of samples and assumed level of confidence clearly point to the fact that the test results are statistically significant? What you are hinting at is that just because the 2 assets under investigation are both stocks it cannot be that one time series exhibits trending properties while the other does not. So then let me ask you, if I can prove to you that with 99% confidence the return correlation between 2 stocks I pick, on the basis of daily data for 10 years, lies within a coefficient of -0.05 and 0.05 you would claim I made a mistake? Or are daily data over 10 years not enough? ;-)

Quote from Ash1972:

Thinking that some markets trend and others are mean reverting is a classic mistake of curve fitting. It's really the most basic one you can make.

 

[mizhael]

Quote from Ash1972:

Thinking that some markets trend and others are mean reverting is a classic mistake of curve fitting. It's really the most basic one you can make.

To be honest with you, everybody told me that this selecting the trend-friendly ones and discard the non-trend-friendly ones approach is precisely curve fitting.

To avoid that, I just throw both the trendies and non-trendies all into the pool.

But he made money and we lost money.

So maybe what he's doing is okay?

 

[ProfLogic]

Quote from asiaprop:

Did the same individual not stress a multiple times here at ET that he DOES NOT HAVE ANYTHING TO SELL? Funny, the course where he unloads his snake oil costs an easy 3000 US dollars. But I guess you can sell anything to some (stress "some" not all) dumb farmers in the country side.

Read his background on that pdf link, most amusing. Also, how can a guy of his age still argue on the lowest verbal level with teenagers/young adults is beyond my comprehension, no matter his background.

You really can't comprehend or read English can you. This was the last class I taught at the University . . . over a year ago before I retired from teaching. You know the class at the University the toothless Brit said didn't exist.

What about my background do you find amusing, that I am educated or that I successfully worked in a variety of fields before I found something that I really enjoyed?

So you think the Brit and Surf are teenagers/young adults? Maybe that is because you are so young yourself but Surf is over 50 and bent on discrediting me in this forum because he is a failing hedge fund promoter that can't trade with any consistency. The Brit simply can't trade.

You really have a propensity of pushing your unfounded opinions as fact don't you.

 

[mizhael]

Quote from asiaprop:

So, according to your definition of curve fitting, you should not trade at all. ANYTHING you ever tested and seemed to work is "fitted to the curve" and may stop working tomorrow. Why do you trade then? I think you have a slight misunderstanding of when one can conclude something is curve fitted and when not.

If you test a strategy and it works over a very large sampling space of price data for one stock but does not work for another stock, would you conclude this is curve fitted even though the standard error under the number of samples and assumed level of confidence clearly point to the fact that the test results are statistically significant? What you are hinting at is that just because the 2 assets under investigation are both stocks it cannot be that one time series exhibits trending properties while the other does not. So then let me ask you, if I can prove to you that with 99% confidence the return correlation between 2 stocks I pick, on the basis of daily data for 10 years, lies within a coefficient of -0.05 and 0.05 you would claim I made a mistake? Or are daily data over 10 years not enough? ;-)

Thank you for this argument.

Could you please tell me how to draw conclusion that

" test results are statistically significant? " at certain level?

I am probably missing something, but there is an assumption of the underlying distribution when you say something is "statistically significant" isn't it?

 

[Equalizer]

Quote from asiaprop:

So, according to your definition of curve fitting, you should not trade at all. ANYTHING you ever tested and seemed to work is "fitted to the curve" and may stop working tomorrow. Why do you trade then? I think you have a slight misunderstanding of when one can conclude something is curve fitted and when not.

If you test a strategy and it works over a very large sampling space of price data for one stock but does not work for another stock, would you conclude this is curve fitted even though the standard error under the number of samples and assumed level of confidence clearly point to the fact that the test results are statistically significant? What you are hinting at is that just because the 2 assets under investigation are both stocks it cannot be that one time series exhibits trending properties while the other does not. So then let me ask you, if I can prove to you that with 99% confidence the return correlation between 2 stocks I pick, on the basis of daily data for 10 years, lies within a coefficient of -0.05 and 0.05 you would claim I made a mistake? Or are daily data over 10 years not enough? ;-)

I think Ash displays some extremely healthy skepticism that's all. Especially with correlation. That does not mean you have made a mistake if all you use is a statistical measure, no more than someone whose predictive model of the S&P500 is based on butter production in Bangladesh.

Ever read this classic?
nerdsonwallstreet.typepad.com/my_weblog/files/dataminejune_2000.pdf

Yes, 10 years helps a lot, but it not eliminate the existence of spurious relationships or abrupt regime chages that the model does not capture.

 

[Ash1972]

Well, the term "curve fitting" is actually short for invalid (i.e. statistically non robust) optimisation. Obviously some back fitting will have been done, otherwise you will have no system at all.

Assigning properties to individual markets is statistically a type of prediction and a major mistake. Frankly, any amateur can create truly fantastic back tested results by taking 20 or so markets and fitting a couple of parameters on a *per market* basis. No system can be robust unless every market is traded in the EXACT SAME WAY.

What makes multiple markets useful is you can boost the risk reward ratio by using the fact that on the whole some markets will be trending whilst others aren't. This is meaningless if you've decided in advance which ones are going to trend. A trend following system by definition makes money by attempting to follow a trend WHEREVER it may occur.

WHAT IF your nicely mean reverting (for the last 20 years) market suddenly, totally out of the blue, became a trending market for the next 5 or 10 years? Your backfitted system will fail to capitalise on this. You will be in quite a bit of trouble too as no doubt you will have based your leverage on the smooth, attractive equity curve you have back fitted.

It's amazing how few people understand this.

 

[Equalizer]

Quote from ProfLogic:

You really have a propensity of pushing your unfounded opinions as fact don't you.

Now there's the pot calling the kettle...

Why is it that when I hear "FrothLogic" I think of the words "Pathological Liar, Charlatan, Clown, Fraud, Crook, BS artist..."? I wonder?

 

[ProfLogic]

Quote from asiaprop:

lol, exactly my thinking when I read that. Too funny!!!

Have you every seen a cow? Hand milked a cow? Slopped pigs? Orphaned corn? Bailed hay? Ran a grinder. Pushed grain in a silo?

I'd pay good money to see that happen.

 

[asiaprop]

mizhael,

you are right if you backtest all names over a short period of time, weed out the ones that did not perform well and believe you will achieve the same performance metric going forward. However, if you run out of sample tests and get for several different forward tests the same statistically significant results then I dont see how this has anything to do with curve fitting.

P.S. Please do not forget I did not select any names based on past performance. I select assets based on their trending properties which I defined mathematically. I dont want to bore you with theory go and see for yourself and tell me what you see when you observe the price pattern of AAPL over the past years versus a name such as RIM. And those two names are not even the most extreme examples and far from being outliers.

Quote from mizhael:

To be honest with you, everybody told me that this selecting the trend-friendly ones and discard the non-trend-friendly ones approach is precisely curve fitting.

To avoid that, I just throw both the trendies and non-trendies all into the pool.

But he made money and we lost money.

So maybe what he's doing is okay?

 

[asiaprop]

stats 101, here some references (I already told you some work also needs to be done by yourself).

* Econometric Analysis, Greene
* Mathematical Statistics and Data Analysis, John A. Rice
* Probability and Statistics, DeGroot, Schervish

Those are mostly the ones I went through in college before I delved into stochastic calculus and derivatives pricing in grad school (not MBA)

Quote from mizhael:

Thank you for this argument.

Could you please tell me how to draw conclusion that

" test results are statistically significant? " at certain level?

I am probably missing something, but there is an assumption of the underlying distribution when you say something is "statistically significant" isn't it?