building a synthetic currency to match your strategy - Great idea or flawed?

G

GoldStandard

The idea is that if you create a synthetic currency by buying or selling various portions of other currencies, to create a synthetic that matches the characteristics of your trading system better than any one currency could. For example if you have a mean reversion trading system, you create a synthetic currency that has stronger mean-reversion tendencies in the timeframe you want to trade than an individual currency would.

An example of someone experimenting with this idea can be found here: http://www.forexfactory.com/showthread.php?t=262827

I suppose you could try this with stocks as well as currencies.

I'm hoping some members of the ET community will comment on the theoritical validity, or lack thereof, of this approach. Are there any glaring flaws you see in the idea? Has anyone else experimented with it?
 
Quote from GoldStandard:
I'm hoping some members of the ET community will comment on the theoritical validity, or lack thereof, of this approach.
More...
I trade nothing but similarly structured synthetics on the forex side. There is a good lot of theoretical work in the literature, but the FF posters are going at it on a trial and error basis, with no apparent heuristic. Totally hit or miss.


Quote from GoldStandard: Are there any glaring flaws you see in the idea?
More...
No major flaws. FF posters just miss, but are in the ballpark.
 
Thank you very much for the reply Kevin. If you don't mind me asking; do you use both trend-following and mean reversion techniques with your synthetics? And does this stuff work on lower timeframes?
 
Thank a lot for your reply, Kevin, its interesting to know that someone is using these techniques with success.

If you don't mind me asking a couple questions: Do you use this technique with trend trades as well or is it more suited for mean-reversion strategies?

Does this stuff work on lower timeframes at all?

Thanks!
 
One small flaw in this strategy is that it is incredibly easy to "curve fit".
The number of potential synthetic currencies you can build is pretty much limitless.
As a result, the fact that you can create a synthetic currency that displays either strong trending or strong mean reverting properties on any given data set shouldn't be a surprise.
Whether those properties hold outside of that data set is the question.
 
Quote from Kevin Schmit:

I trade nothing but similarly structured synthetics on the forex side. There is a good lot of theoretical work in the literature, but the FF posters are going at it on a trial and error basis, with no apparent heuristic. Totally hit or miss.



No major flaws. FF posters just miss, but are in the ballpark.
More...

Could you please point to some papers that you found interesting?

Thanks

Ninna
 
Quote from nLepwa:

Could you please point to some papers that you found interesting?
More...

I've PM'ed you.

Basic method is to do some kind of factor decompositon (PCA, ICA, Projection Pursuit, etc.) of all the liquid pairs. Then look for structure in the scatterplot matrix. Attached jpeg is an Excel scatterplot matrix of the first five factors of a Maximum Kurtosis (Roweis) ICA projection. If you see structure (check out factor 1 vs factor 3 in the attached image) then try custom loadings, project the original factors into the orthocomplement of the space spanned by your candidate factor ( (I-XX')A(I-XX') where X is your vector of loadings, make sure the the squares of your loadings sum to 1), iterate until the structure in the scatterplots goes away.

Interestingly, the raw factors are reasonably predictive of SPX strong up (green dots in the scatterlots) and strong down (red dots).

No, I am not channeling Jack Hershey, this is the real thing.
 

Attachments

Quote from GoldStandard:

Thanks for the additional explanation Kevin. If you could also PM me those research references I would be most grateful.
More...
PM sent.

FWIW, I have not been able to synthesize a momentum/mean-reversion factor in the FX market that shows a consisten enough bias to justify a core long (momo) or short (MR) exposure.

On the other hand, if you know what you are looking for, if you have a stronly informative prior model of what is going on under the hood, you can often find useful timeseries decompositions that automated methods miss.

I have attached a paper I have found particularly useful in the area.
 

Attachments

You must log in or register to reply here.
Back
Top