How to research and verify trading ideas
- Thread starter talontrading
- Start date
- Status
http://localweeklytrading.blogspot.com/2009/12/longed-gld-at-close.html
disclosure: all the stuff there is papertrading research ...
disclosure: all the stuff there is papertrading research ...
well the ETF-tracker GLD is above 50% percentile and below 10% percentile, we had three down candles, so it might be a good long. Just found that interesting, but then maybe it wasn't. And no I haven't done the full research on this, somebody else is doing some testing for me and I am hoping to get some results this weekend, as I don't have the software to do this properly.
Talon- On some old posts here on ET, a poster shared his methodology for handling correlation risk among systems. Basically, he measured the correlation between historical monthly cashflow streams from various strategies looking for uncorrelated systems. Once he had a few that were <.2 or so, he would solve for the optimal weighting among the systems to maximize his monthly sharpe ratio - suggesting that the net result was basically the same return with lower drawdown allowing him to take larger size. More or less standard portfolio theory as I understand it. Assuming they are truly structurally different, achieving diversification across systems seems plausible given that you can go with trend, mean reversion, multiple timeframes, etc.. I don't know how to improve on this, despite it being 'optimized' to past performance.
I wanted to ask what you make of this approach? If it is sound, does it make sense to use the same approach on the stock portfolio traded within a given system? Weighting positions differently for each stock based on historical performance feels like a different form of curve fitting. That said, perhaps its just a matter of sample size - if there are enough trades to be statistically significant for each name in the portfolio, then adding/deleting names or upsizing/downsizing positions based on historical performance would be valid?
Completely unrelated, you mentioned something in an earlier post about scalping based on reading the orderbook...any books or resources you can recommend to learn a bit more about it?
Sorry if these are elementary questions, but they've been rattling around for a while. Thanks again for all your help.
I wanted to ask what you make of this approach? If it is sound, does it make sense to use the same approach on the stock portfolio traded within a given system? Weighting positions differently for each stock based on historical performance feels like a different form of curve fitting. That said, perhaps its just a matter of sample size - if there are enough trades to be statistically significant for each name in the portfolio, then adding/deleting names or upsizing/downsizing positions based on historical performance would be valid?
Completely unrelated, you mentioned something in an earlier post about scalping based on reading the orderbook...any books or resources you can recommend to learn a bit more about it?
Sorry if these are elementary questions, but they've been rattling around for a while. Thanks again for all your help.
I gotta call you out on this one! 
The software you need is available free... you can download a great statistics package called R with tons of good documentation for free... you can also do a lot of good stuff with VBA in Excel.
What it requires is some real dedication in learning to program and use these packages. I did not have a programming background so I certainly understand how difficult this is... however I find there is no substitute for manipulating the data myself, even if I can work with someone who does it 100 times better than I do... I still have to get my hands dirty because that's how I learn.
The software you need is available free... you can download a great statistics package called R with tons of good documentation for free... you can also do a lot of good stuff with VBA in Excel.
What it requires is some real dedication in learning to program and use these packages. I did not have a programming background so I certainly understand how difficult this is... however I find there is no substitute for manipulating the data myself, even if I can work with someone who does it 100 times better than I do... I still have to get my hands dirty because that's how I learn.
1) No one on the entire planet hates "Modern Portfolio Theory" (don't get me started... this is 1960's stuff... why is it still called Modern... geeze...) more than me. No one. My main issue is that asset or asset class returns cannot be predicated based on past performance nor can correlations between asset classes.
Having said that, the approach may have more validity with systems that are, as you say, structurally different. We don't do any kind of portfolio optimization like that here, but we do definitely make an effort to keep correlation between systems and just things we do in general to a minimum.
I really haven't done the work or research on this to give you any kind of answer... I will say it is an interesting idea. I suspect the fluxing correlations between markets and asset classes could also cause some issues with this approach, but I'm not sure how serious those issues would be without doing a lot more investigating.
2) There is almost nothing available in any format to teach this. I do not know of any good books or instructional material on this topic... it has always been taught person to person... and the rules change constantly so anything written more than 3 years ago probably isn't valid anymore.
It's also worth mentioning that this is absolutely the most difficult kind of trading there is. Beyond a doubt. Get profitable with something else first would be my advice.
Having said that, the approach may have more validity with systems that are, as you say, structurally different. We don't do any kind of portfolio optimization like that here, but we do definitely make an effort to keep correlation between systems and just things we do in general to a minimum.
I really haven't done the work or research on this to give you any kind of answer... I will say it is an interesting idea. I suspect the fluxing correlations between markets and asset classes could also cause some issues with this approach, but I'm not sure how serious those issues would be without doing a lot more investigating.
2) There is almost nothing available in any format to teach this. I do not know of any good books or instructional material on this topic... it has always been taught person to person... and the rules change constantly so anything written more than 3 years ago probably isn't valid anymore.
It's also worth mentioning that this is absolutely the most difficult kind of trading there is. Beyond a doubt. Get profitable with something else first would be my advice.
Hmm I respectfully disagree (to an extent). I think there is value in portfolio theory, but I must add the caveat that what the other poster suggests is quite dangerous. Suggesting to optimize based on current correlations and leverage up is a recipe for disaster when you have, say, a liquidity crisis.
In such situations, suddenly everything goes to correlation of 1 and you can now bend over and kiss it goodbye if you are too juiced. Just ask LTCM, countless other hedge funds, or just look at the crazy perverse market behavior of 2008.
However I submit that this could be usable from the opposite viewpoint. As your systems/positions start to correlate more, you reduce size in those systems/positions to give you the same net result as your "benchmark".
You can manage risk in this manner or you can abuse it and blow up. My suggestion is no credit spreads, no leverage unless you really know what you're doing and even then stuff happens (counterparty risk, exchange closures, other unforseeable black swans, etc etc etc).
-TD80
In such situations, suddenly everything goes to correlation of 1 and you can now bend over and kiss it goodbye if you are too juiced. Just ask LTCM, countless other hedge funds, or just look at the crazy perverse market behavior of 2008.
However I submit that this could be usable from the opposite viewpoint. As your systems/positions start to correlate more, you reduce size in those systems/positions to give you the same net result as your "benchmark".
You can manage risk in this manner or you can abuse it and blow up. My suggestion is no credit spreads, no leverage unless you really know what you're doing and even then stuff happens (counterparty risk, exchange closures, other unforseeable black swans, etc etc etc).
-TD80
yes, I see you are right on this one, now I just need to find myself some time cause lately I have been juggling quite a few things. Keeping everything in the air without breaking anything seems to be an art in itself. It's on my todo list from now on ...Quote from talontrading:
I gotta call you out on this one!
The software you need is available free... you can download a great statistics package called R with tons of good documentation for free... you can also do a lot of good stuff with VBA in Excel.
What it requires is some real dedication in learning to program and use these packages. I did not have a programming background so I certainly understand how difficult this is... however I find there is no substitute for manipulating the data myself, even if I can work with someone who does it 100 times better than I do... I still have to get my hands dirty because that's how I learn.
There are techniques that people use to correct this issue. Basically you take the view that the past performance and correlations have some info about the return structure mixed with a bunch of noise, and the problem is that basic MPT optimization on raw historical data overfits the noise. So these methods adjust either the input data or the weights that come out of the optimization to get a less extreme portfolio.
If you want more info google bayesian shrinkage or the Black Litterman model. IMHO the idea is good but the formal techniques are overkill for most of us - you could get most of the benefit by using the basic MPT approach and then adjusting the extreme weights towards something more reasonable without the complications. For example, you could take the average of the optimized weight and an equal weight for each system. Or just equal weight the systems and find something better to spend your time on!
The prior post also mentioned the problem of correlations increasing in times of stress - this is often true but does not trash the benefits of the approach. You still get higher compounded returns if you can lower the portfolio vol during normal conditions.
If you want more info google bayesian shrinkage or the Black Litterman model. IMHO the idea is good but the formal techniques are overkill for most of us - you could get most of the benefit by using the basic MPT approach and then adjusting the extreme weights towards something more reasonable without the complications. For example, you could take the average of the optimized weight and an equal weight for each system. Or just equal weight the systems and find something better to spend your time on!
The prior post also mentioned the problem of correlations increasing in times of stress - this is often true but does not trash the benefits of the approach. You still get higher compounded returns if you can lower the portfolio vol during normal conditions.
- Status