| Index | Recent Threads | Who's Online | User List | Search |
|
|
 |
Portfolio123 » List all forums » Forum: General Comments » Thread: BLOG: A Change-Of-Pace Growth Strategy |
|
Total posts in this thread: 36
|
[Request a Feature]
[Post new Thread] [Add To My Favorites] [Watch this Thread] |
| Author |
|
|
mgerstein
Advanced Member
|
EXCERPT Growth investing is often seen as dicey at times when markets struggle. But this time around, things have been seemed upside down in terms of what works and what doesn't. Indeed, we see some of that in the growth arena, where it is possible to develop a successful strategy by focusing on a big picture, as opposed to using the sort of here-and-now approaches that are typically preferred. FULL ARTICLE. BY: Marc Gerstein CATEGORY: What's Working |
||
|
|
olikea
Advanced Member
|
This is a truely excellent article- thank you! I like the way you show what has been working recently vs. what hasn't. However, I have a comment to make: I found that in the 6 month period before the end of 2007, "growthy" portfolios seemed to perform the best. I did a test on your "recent growth" ranking system over the 2nd half of 2007, and it clearly outperforms my "value" system over same period. I hypothesize that this was due to the fact that at the time, it was considered the "credit crunch" was going to be limited to specific industries, in particular it was considered that tech (read high growth) was immune. However, in the 1h of this year, notice how the situation appears to have reversed, with the poor performance of the growthy factors (as you mention) but the better performance once again of value factors. One wonders to what extend one needs to adopt a contrarian approach rather than be constantly "chasing one's tale". ---------------------------------------- 
growthfactors2h07.png
(25177 bytes)
(Download Count: 212)
----------------------------------------
valuefactors2h07.png
(23798 bytes)
(Download Count: 210)
----------------------------------------
growthfactors1h08.png
(23854 bytes)
(Download Count: 208)
----------------------------------------
valuefactors1h08.png
(21641 bytes)
(Download Count: 208)
|
||
|
|
jerrodmason
Advanced Member UNITED STATES Joined: Jan 14, 2005 Posts: 599 Status: Offline |
A philosophical comment about ranking systems: To my mind, the purpose is more akin to Discriminant Analysis than it is to Regression. That is, it's only the performance at the extreme that really matters. I don't particularly care if the return/rank curve is smooth or even monotonic (ie always increasing or decreasing) all the way from 0 to 100. Which would you rather have (or be more comfortable with)? A nice smooth curve, or one that has a spike at the highest ranks? For me the answer is clear. Until my pports/sims/screens start selecting stocks with ranks below 90, I'm only interested in the performance of the 90+ set. ---------------------------------------- The smart money was on Goliath. "He's not the Messiah. He's a very naughty boy." |
||
|
|
o806
Advanced Member
|
I agree with your emphasis on just the very top buckets. I think the reason people like to see a smooth slope for most or all buckets is to build confidence that the ranking system is responding to a valid signal rather than to just noise. I think there is a better way to get this confidence than by looking for a consistent slope over all the buckets. Here is my approach. First, I like to use buckets that will contain 1/2 to 1/4 of the stocks I will eventually use in a portfolio based on the ranking. Since I like to have portfolios with 15-20 stocks, this generally means I want 5 to 10 stocks in a bucket. Because the maximum number of buckets is 200, this means I want to run tests on universes of no larger than 2,000 stocks (such as the Russell 2000). Second, I have made a number of custom universes which have 2,000 stocks and another set with 1,000 stocks (these divisions are based on market cap after stocks pass a basic liquidity and price filter). Ideally I want to see that the top 2 to 5 buckets of a 200 bucket test all do well. I am especially watchful for a top bucket that under performs the 2nd to 5th buckets. That makes me suspicious that the ranking system is finding stocks that are "too good to be true". Also I want to see that the top 5 buckets all do well for several universes: Such as microcap, smallcap, midcap and largecap (each 1,000 stocks as I define them for testing purposes). Often largecap will not do well compared to the other universes, but even so, I am pleased if the top 5 buckets of large caps do better than the average of the remaining 195 buckets. Also I like to rerun the tests on the NYSE and NASDAQ universes. Usually the NASDAQ universe outperforms the NYSE for the rankings I develop, but sometimes it is the other way around. Again, what I am looking for is to see if the top 5 buckets do better than the average of the other 195. If so, I conclude that that ranking has validity. In short, my testing method tries to see if a ranking system has "consistency". I think this desire for consistency also is behind the desire of many to see a smoothly rising slope of returns by all the buckets in a test. I understand the desire, but I think rerunning the tests for 4 or 6 different market universes is a better way to go for several reasons. One of those reasons is that it lets me focus on the top 2-5 buckets from each universe without requiring a smooth slope for the rest of the buckets. Brian ---------------------------------------- [Edit 1 times, last edit by o806 at Jul 31, 2008 1:36:44 AM] |
|||
|
|
o806
Advanced Member
|
Marc: I agree with olikea that you have given us though provoking article. Your use of the YTD 2008 test period illustrates that a ranking system that gives good returns over a 5 or 7 year period can "misbehave" when the time frame is less than a year. I like to run several time segmented tests for a ranking system. This lets me see how a ranking system handled past adversity. So in addition to the overall 7 year test result, I want to see several shorter period results. I am especially interested in seeing how a system behaved (or misbehaved) in the March 2001- March 2003 period which had a couple bear rallies followed by declines that led to lower lows. This helps me not to panic if one of my real money systems struggles over the past year. Actually one of my systems is down 6% for 2008 YTD (but it is still doing a few percent better than the general small cap market). This "negative" performance does not concern me since before I started this method I knew from tests on 2001-2002 that it would not make money during down markets. So its 2008 YTD returns do not make me think the system is broken because I can see that it is just doing what it naturally did during previous bear markets. So, I just keep rebalancing this portfolio for with the confident expectation that the ranking system will do explosively well then the current bear period is over (just like it did explosively well in 2003-2004 following the 2001-2002 bear). Oh, it tests decently well for 2005-mid 2007 so it like all types of bull markets. My other two real money systems are ahead of the market by 15-22% so far this year and that also makes it easy to stick to the rebalancing of third system which is (temporarily) struggling. Testing multiple periods takes time since P123 does not provide any way to automate this. Even the selection of dates in the performance test page takes several more mouse clicks than is required to do the similar date changes in the Simulation test page. So I can understand why many P123 users do not bother with time segmented ranking tests. However, I consider testing over narrow time frames to be well worth the effort. Regards, Brian ---------------------------------------- [Edit 1 times, last edit by o806 at Jul 31, 2008 2:27:52 AM] |
||
|
|
Stittsville123
Advanced Member
|
In this environment it is best to K.I.S.S. One good ranking system is ET http://www.portfolio123.com/rank_details2.jsp?rankid=54051 Steve ----------------------------------------
ET Ranking System.GIF
(41260 bytes)
(Download Count: 180)
|
||
|
|
o806
Advanced Member
|
Steve: Good point about keeping the ranking system simple. And it is a nice ranking system for tough times as well as the good ones. I put the ranking system though my own quick mini slice and dice "stress" test. Here are the results. I ran with a min price of 3. Results of top 5 of 200 buckets (Aug 2007-present) Top bucket is listed 1st (reverse order from how they are displayed on the P123 report chart.) R1K..........12, 22, -22, -23 (5 stocks per bucket) R2K..........10, 40 -35, 25 (10 stocks per bucket) R3K..........17, -2, -3, -15 (approx 15 stocks per bucket) Nasdaq....15, 6, -25, -8 (approx 12 stocks per bucket) NYSE........12 -20, 0, 1 (approx 10 stocks per bucket) Even during the last 12 tough months, it looks like one would have done OK with the top 10 or top 15 stocks, but a top 20 or top 30 would not have done so well. Brian |
||
|
|
mgerstein
Advanced Member
|
Simplicity does have its virtues. Also, EV-based systems do seem to be working. Olikea cited another system along those lines a few weeks back. This system is particularly intriguing since you used the CIAC item instead of a number from higher up on the income statement, like EBITDA. So your numerator was fairly close to the E in a conventional earnings yield calculation, with the bigger change being in the switch from price/market cap to EV. That means you were subject to alot of the accounting issue people cite when expressing discomfort with EPS, yet your system works. It would also be interesting to look into the role played by the path from some thing like CIAC (income available to common shareholders) to EPS, where share averaging and dilution factors figure in. I think I'm going to dig further into this. Has anybody else tried to de-construct PE to find the source(s) of the shortcomings |
||
|
|
mgerstein
Advanced Member
|
re: the discussion of positive spikes in the best bucket(s), they tend to make me a bit nervous, especially given the factor reversals we've been seeing lately (the subject of another fascinating thread). A smooth bucket-to-bucket slant is certainly no guarantee against a reversal, but as noted by someone else in this discussion, it does suggest there may be something more solid about the system as a whole. |
||
|
|
grokkalot
Advanced Member
|
I'm willing to take the other side of many of the ideas being supported here. The reality of the market is that there is nothing to stop it from being a non-stationary system, so any ranking that worked well in the past can stop working in the future. Everyone agrees on this point. By being active in P123, presumably everyone also agrees that ideas which hold up to empirical testing are more likely true than those that don't. But since we can't test what will happen in the future, we are drawn to engage in non-empirical arguments about what practices one believes are most likely to be robust against the extrapolation into the future known as "predictio"n. So with the above in mind, I'll play too... Q1) Is a strategy that backtests well against 5 years of data likely to perform better than a strategy that backtests well against 1 year of data? A1) Other things being equal, yes. But other things are never equal, and, in particular, if the 1 year strategy performs better on the most recent year of data than the right answer is far from clear even at the level of the vagaries. Q2) How relevant is the rank order correlation of a P123 ranking system in the ranks below where one would normally buy and sell? A2) If ranking systems were real time, if "normally" meant always, and if the bucket graph was continuous rather than discrete, then I'd have to say "not relevant" - at least I'd invite someone to do a better job of explaining why if they disagree. But because all of those caveats are false, there is some relevance. In practice, people do sometimes wait to sell a stock until it falls to a lower rank, and people may even buy lower rank stocks if they distrust all the top ranked ones or if their sector/industry concentration constraints are causing them to buy lower than they think (or not buy at all). And because the data the rankings are based on isn't updated in real time, real time rankings may be different than simulated ones. And because buckets are discrete, the immediately lower buckets probably say something about what's happening at the lower end of the bucket that is just above any given cutoff. Q3) Are simple ranking systems - e.g. ones with fewer factors - more likely to be robust to future change? A3) No - or, at least, the reverse is true in theory. I suspect that some of the support for the "simple is robust" idea is a mis-analogy from the regression phenomena where the variability and overfitting tendencies of model do increase monotonically with their number of effective parameters. But that is a mis-analogy because the weights in P123 models are all required to be positive and because the weights are a partition of unity; in fact, a multi-factor system constructed out of factors that are individually positively correlated with stock performance is actually very similar to a boosting or ensemble modeling technique in statistical prediction modeling, and such techniques are known to *reduce* variance. That's the formal/theoretical story, but what can go wrong with a more complicated model is that it is less well understood. So if some of the predictive performance of simpler models tends to reflect not only there past statistical performance but also domain understanding that is captured by the model creator, then a more complicated model that has been put together without an accompanying increase in time spent thinking through the implications of the various terms may be more error prone in practice. Q4) How special was the past year? A4) IMO, it was pretty special. For one thing, the last time small cap value took such a beating was in 1998, and that year falls outside of the P123 database. For another thing, the fundamental problems with the U.S. economy and the negative wealth effects today are in a different category than anything that was happening in 1998 or more recently. My investing lifespan doesn't allow me to compare this in detail to 1991 or 1974, but one seemingly would have to go back at least that far for analogs. Q5) Does the market behave now as if it was less interested in value ratios than in previous bear markets? A5) I don't know. Would be interested in looking at data on this if anyone has it. I suspect that the answer is no for past bear markets that were equivalently bad. |
||
|
|
[Show Thread Printable Version] [Post new Thread] |
