- Image via Wikipedia
I love to surf around and discover articles by following Alice down the rabbit hole – meaning – start down a path and just keep going. I was reading an excellent post by Woodshedder over at ibankcoin.com this morning on position sizing. I think the post and Van Tharp articles (part 1 and part 2) it references are a good source for understanding position sizing and the effect on a given system. In fact, I think they should be required reading for any aspiring trader.
The Woodshedder post referred me to a blog I was unaware of – Max Dama – a student/quant. Max is obviously very, very smart (about 10x my brain power judging by his blog). Max was profiled in a piece in Porfolio (also available via Wired) on the rise of quant-based strategies in the financial world. That article referenced Legend Advisory – who offer a quantitative approach to asset allocation. They use an artificial intelligence model for determining the asset allocations between different sectors and styles – think “should I be in financials?” or “will growth outperform value?”.
I did some more searching on Legend and came across an article in Allbusiness that detailed some of their predictions from December 2007. Here’s what I find fascinating – for all their math and study, two of their predictions stood out to me as being so, so wrong (and they had others that were right) that it made me question their whole approach. My comments are [in the brackets].
1. The financial sector should outperform in the second half of 2008. [My comment - wow - stunningly wrong].
2. Corporate earnings growth should continue to decline. [Interesting and correct - but what do you do with it?]
3. P/E multiple expansion will result from lower interest rates. [I don't have the tools to look at this one to know if it right or wrong - my guess is that it is correct but again, what do I do with it?]
4. The US Dollar should continue to weaken throughout the year, but could reverse its course before the end of 2008. [Correct!]
5. After underperforming the broader markets for almost 6 years, healthcare should surprise investors (positively). [Healthcare outperformed the general market, but still declined - I'll give them half-right.]
6. The real US economic growth rate (GDP) will decline in 2008, below its historical trend. [Correct - but then how could you say that financial would do better given that the market is so dependent on the performance of financials?]
7. Growth stocks should outperform value stocks in 2008. [Value actually outperformed growth, at least in the small cap arena, for more than half of 2008]
8. The S&P 500 index should reach 1625 by mid year, 2008. [Stunningly wrong - the index was at 1504 on Dec 7th, the time of the writing of the article]
All of this isn’t to knock Legend specifically – most models (including my own) did not do well in the radically different market that 2008 gave us. They also may be making some predictions here that didn’t make it into their client’s portfolios or were altered significantly during the year.
The point is this: don’t be too impressed by financial firms that throw a lot of whiz-bangy mathematical jargon out there. Hell, I’m not that good at math, and, at first, I thought Legend’s approach sounded very interesting. But in looking at their results, I’m reminded that you can throw all the math you want at the financial puzzle and still end up with a stinker.
It also reinforced, in me, the idea that everyone has to be responsible for their financial future. You can’t put your financial future on autopilot with someone else and not be involved. I am consistently surprised by the number of people, in surveys, that spend more time planning vacations than planning their finances. I am also consistently depressed by the total lack of education prior to graduation from high school and college on these issues.
It sounds all well and good to let a computer model determine everything – but at the end of the day, much of what these firms put out there is mathematical masterbation. Just remember that next time someone tells you they’ve got a model that can solve all your problems.
![Reblog this post [with Zemanta]](http://img.zemanta.com/reblog_e.png?x-id=3418339d-469f-49db-a0ee-9bb2843edef9)
I like the down the rabbit hole approach.
Much of what Max is doing is over my head but it gives me something to aspire to.
Yeah me too Wood!
Nice blog… I’m going to start reading you.
Re: Max Dama – I wish I understood the AI stuff when I was 19. Too me about ten years to catch up and he still has the leading edge of the curve.
This is a great place to start http://ai.arizona.edu/papers/dog93/dog93.html as it covers two of the major types of machine learning: symbolic learning and artificial neural nets.
Personally, I’m not so much fond of ANN because it’s seriously “black box” kind of analysis. Perhaps you have a valid answer but figuring out the byzantine connections between the neurons is maddening and overly obtuse.
ID3 and it’s successor C4.5/5 are superior when it comes to human readability due to the fact that there is a kind of logic to the answers.
I prefer Genetic Algorithms because they’re almost as fast as ID3 and, done right, far more robust.
The one thing to remember about Dama is that he’s more of an AI guy than a trader – in this post he presents the code for his hill climbing code (which is a technique to modify one part of a system until it performs) here:
http://www.maxdama.com/2008/07/trading-optimization-hill-climbing-code.html
The code is in LISP. The original AI language.
I was half expecting his work to be done in Ruby or Python but he’s got props from me for doing work in the holy grail of computer languages.
Re: Legend – meh. They’re using ANN, which means to me there is about an 80% probability they don’t understand the system.
Thanks for the comment. I’ve studied the Neural Net side a whole lot – and done a fair amount of work in it – but I’ve never found anything where I felt I could have some remote level of confidence in the solution. I also felt that that NNs fall apart at times of change in the market. So, you could have a fantastic net going into October and just watch it crumble. Which is to say that they have no resilience.
Now – I agree with you – genetic algorithms is much more interesting – I remember a company called Trading Systems Lab (www.tradingsystemslab.com) that did a ton of work in this area as a commercial product. It was priced at something like $50k so well beyond me. But I thought what he was working on was interesting.
I’m not sure if Max is more trader or AI programmer at this point – I would say he’s a student on his way to being a system designer – which doesn’t necessarily mean a trader.
And LISP – I spent about a year learning it and using it and never “got it” – the power is obvious but beyond me. I expected most of his work to actually be in R or MathLab.
Agreed on Legend – and it’s clear that their techniques suffer from the lack of resilience I mentioned above.
Thanks for stopping by!
Not soley to pimp my own post but, if you’re interested in playing with GA without having to roll your own, there is a semi-nice Java app called ‘Merchant of Venice’ that I wrote about here.
It has GA, GP (which I think are even screwier than NN), and NN options to analyse trades.
http://www.ibankcoin.com/peanut_gallery/index.php/2009/01/26/lets-get-genetic-algorithmically/
Damian,
I just found this post. Looked through the site (this might be the second time, not sure). Liked this one especially: http://skillanalytics.wordpress.com/2009/04/22/final-found-a-document-ive-been-looking-for-the-seven-sins-of-fund-management/
Combine that with Investors Intelligence’s poll of FAs for a nice contrarian indicator: (free source) http://www.schaeffersresearch.com/streetools/market_tools/investors_intelligence.aspx
I’m a normal kid fyi (I study on the weekend though which maybe gives me an edge).
Anyway, none of the many people I’ve talked to about algo trading have found pure machine learning systems to be their holy grail.
dskill – I trade about 1-2 times a week, but never scheduled, based on discretionary views and I also take advantage of two recurring/persistent inefficiencies (looking for more). Besides that I have a long-term/value portfolio and am looking into physical real estate. So far the machine learning systems I’ve tested haven’t produced results I would feel confident trading on, especially intraday. It takes work to clean the data and build a bug-free system but I’m still optimistic. Persistent inefficiencies/arbs have been best in terms of reward/risk for me so far.
Regards,
Max