All, We released two new models that were the winners in an Stanford University class contest. You can see them in the P123 models section. They are called “Stanford MomValue” and “Value Sentimentum”.
Make sure to check the Knowledge Base articles they wrote about how they designed the systems. The articles are [url=http://www.portfolio123.com/kbase.jsp]http://www.portfolio123.com/kbase.jsp[/url]
Dear Standford students
Thanks for this model the results look amazing. Could you pleaze let me know what the exact formulas are for the ranking system
Thanks
Stefan
Just click on the system and go to Trading System.
Also check out the About page for the system.
Finally, you can find papers about the methodology here:
[url=http://www.portfolio123.com/kbase.jsp]http://www.portfolio123.com/kbase.jsp[/url]
I added my market timing rules discussed here to the 2 new P123 Ports. The results are:
For the Value Sentimentum Port; the annual return has increased from 23% to 29%, and the max drawdown was reduced from 56% to 23%. The mod Sim can be seen here.
For the Stanford MonValue Port; the annual return was reduced from 31% to 26%, but the max drawdown was reduced from 57% to 25%. Considering that the new Sim was out of the market for over 1/3 of the time the tradeoff of 5% reduced annual return for 32% reduced drawdown seems a no brainer. The mod Sim can be seen here.
Denny 
Stanford MomValue was renamed to ‘Kasa Model’ as requested by the author.
After giving the Kasa model (its new name) a quick look when was first announced, I promptly set it aside because it seemed to be nothing but a recombination of a few concepts used in P123 circles. The only thing slightly unusual on first glance was using a 2 month relative strength look back rather than a more standard 6 month or a year. However after seeing this week’s performance report email gives Kasa a 1 year 127% excess performance vs the SP500, I decided to give it a closer look. I did not like what that deeper look revealed.
Since the Kasa model works with a 3 month rebalance (presumably something imposed by the professor in the MBA course it was designed for), the model only has 3x9 = 27 data points. Now that number of 27 may seem close enough to the 30 data point rule of thumb for statistical dependability but 27 is also low enough that it can be easy to “curve fit” results with just a few rules. But I am starting to ramble. Let’s get to the point.
If the rebalancing is left at every 3 months but the starting date is moved 1 week later results are about the same (around 28-30%/year). But change the starting date by 1 more week (ie start the sim 2 weeks later than the MBA students did) annual gain drops to 0.9% and maximum drawdown increases to 83%. Ouch. Start one week later still and returns rebound a bit to 18%/year. I did not bother shifting the start date any more since it is obvious that success or disappointment with a 3 month rebalance is mainly dependent on random luck in picking a lucky starting date.
Let’s set aside that 3 month rebalancing rule which was likely imposed by the professor. When I tried rebalancing weekly, every 2 weeks and every 3 weeks, I saw annualized gains of 4%, 8%, and 13%. Nothing to get excited about here. I stopped testing at this point. (Oh, I also did some performance testing on the Kasa ranking system before doing the sim tests described above. The Kasa ranking system is remarkably weak as ranking systems. I used 5 different custom universes of different market caps but all with reasonable liquidity requirements for my ranking performance tests.)
It looks to me that the Kasa model is a prime example of curve fitting and just had a lucky past 12 months to get into the winners circle in the P123 email.
Am I missing something?
Regards,
Brian
Brian, I had exactly the same thought, and ran the same experiment on Denny’s variation with the same kind of result.
That brings up an interesting problem: It doesn’t seem possible to impose timing rules that act mid-balance interval, and so Denny’s timing only comes into play on four days each year.
This limitation reminds me of the several longstanding feature requests to allow sell stops to be simulated as they occur mid-session rather than next day.
All,
Brian’s results are a very big difference. I wondered how much of that difference was caused by bad timing relative to the market recession and corrections. So I tested the Kasa Model with the market timing rules mentioned above and by shifting the start date 1 week at a time for 13 weeks. The results were:
0 week shift; Annual return = 26.9, Max Drawdown = -25.4
1 week shift; Annual return = 27.0, Max Drawdown = -25.4
2 week shift; Annual return = 14.4, Max Drawdown = -39.1
3 week shift; Annual return = 25.6, Max Drawdown = -34.1
4 week shift; Annual return = 25.4, Max Drawdown = -34.1
5 week shift; Annual return = 18.0, Max Drawdown = -36.1
6 week shift; Annual return = 21.8, Max Drawdown = -30.4
7 week shift; Annual return = 21.9, Max Drawdown = -30.4
8 week shift; Annual return = 17.2, Max Drawdown = -29.4
9 week shift; Annual return = 17.1, Max Drawdown = -29.5
10 week shift; Annual return = 15.7, Max Drawdown = -33.9
11 week shift; Annual return = 15.8, Max Drawdown = -23.7
12 week shift; Annual return = 16.0, Max Drawdown = -27.4
13 week shift; Annual return = 15.3, Max Drawdown = -27.4
The annual return varied from 14.4% to 27.0%, and the max drawdown varied between 23.7% and 39.1%. That’s not nearly as bad as Brian’s results without market timing, but still not nearly as robust as I would be willing to follow.
Next I re-ran the 14 Sims with 1 week rebalance. The results were: annual return varied between 21.7 and 22.8%, and the max drawdown was -31.7% for all 14 Sims. Now that’s a consistency that I could trade.
I think the bottom line is DON’T USE 3 MONTH REBALANCING!
Denny
Technical question (error?): This model uses “2mnPctRet” = close(0)/close(60). Since “close” is defined as “bars ago” (?), shouldn’t it be “3mnPctRet” (2mn would be close(40) instead of close(60)?
In other words, I’ve never seen “bars” for Sat or Sun.
[fwiw1: The students made this mistake in their pdf discussion.]
[fwiw2: technically, I believe the “avg” trading days, ergo bars, is 21 per month, so it should be 42 instead of 40)
OldQuant,
I can confirm that close(0)/close(60) is approximately three month momentum, not two month. Actually a more precise estimate for three month momentum is close(0)/close(63).