Skip to content

Rotation Part Deux (Two)

February 24, 2010

In the last post we discussed some of the components of a good rotation model. In our examples we used ETF Rewind’s Rotation tool. http://marketrewind.blogspot.com/2010/02/etf-rewind-announcing-customizable.html
I showed that rotation across major indexes was suboptimal even with a good rotation model, and that adding bonds and a cash proxy had a much bigger impact on returns and risk-adjusted returns. Its now time to see if we can take things to the next level. We want something that is uncorrelated to stocks and bonds but is also diversified and typically delivers high returns relative to volatility. The natural addition to the rotation model is to choose a commodity index, which matches all of these desired characteristics. In keeping with the highly practical model of using ETF’s, we will use ticker DBC which is highly liquid and is a good proxy for a basket of major commodities. As a reminder, included in the previous test were tickers SPY, IWM, QQQQ, SHY, and AGG, for this run we will also include DBC. Lets take a look at the results:

Note: we are selecting the top asset class, LONG ONLY, and rebalancing weekly. This test was conducted from March, 2007 and does not include commissions or trading frictions.

Hypothetical
Top # Performance
1 Model SPY
3-Yr. Gross +106.9% -15.5%
C.A.G.R. +27.4% -5.5%
Sharpe 1.16 0.21
Maximum +8.6% +13.3%
Minimum -11.2% -19.8%
Average +0.5% -0.0%

The returns of the model are a stellar 27% CAGR with a sharpe ratio (risk-adjusted return) over 1. This is a substantial improvement on the previous run which showed a 13% CAGR and .6 Sharpe ratio. Note that both absolute and risk-adjusted returns are higher than in the previous model. The lower correlation of DBC to the others allowed us to achieve higher returns in 2007 and early 2008 while the other asset classes were struggling. Furthermore, the commodity index is valuable in the model because it has no credit risk (unlike individual stocks), and similar long-term return/risk as the major stock indexes.

The question is can we make this even better? The way to improve the model is to remove some of the multicollinearity and noise by eliminating redundant asset classes. In this case, QQQQ, and IWM have a complex relationship with timing the SPY and are best utilized within the context of of a different framework. http://marketsci.wordpress.com/2010/01/27/a-different-spin-on-the-%e2%80%9cperfect-alignment%e2%80%9d-theory/ Another model component- EFA- represents the major developed markets and has a very high correlation with the SPY, making it somewhat redundant. Another redundant asset is AGG, which is the aggregate bond index–this has much less “safety value” than SHY (the 1-3 year T-bond index), and has less upside than some of the other asset classes in the model.

There are also some potential asset classes that are worthwhile adding to improve absolute and risk-adjusted returns. The most logical addition to the model will give us a low or even negative correlation, and a good potential return/vol with low credit risk. Why not have the ability to go short one of the model components? Thus the simple answer is to add an inverse ETF to the model that is not leveraged. In this case we will add ticker “SH” which is simply the S&P500 Short ETF which as the name suggests will simply sell short the S&P500 index. The benefit from a practical perspective is that this ETF is highly liquid, can be traded without having the hassles involved with short-selling and removes the risk of future short-selling restrictions. Another good addition is to have some currency exposure–the Euro also has a negative correlation to many of the model components and has low volatility relative to its prospective return. It also helps to substitute for dropping EFA which is tied heavily to Europe. For this we will use ishares Euro which is ticker “FXE.” Finally lets add a Real Estate component, and in keeping with adding back some international exposure, why not add some Canadian REITs. Note that this also carries a very nice dividend yield that is not factored into performance.This ticker to represent this component is “XRE.TO.” Lets add these asset classes to the model and see how it affects performance. So our final model list of tickers will be: SPY, DBC, EEM, SHY, SH, FXE, XRE.TO

Hypothetical
Top # Performance
1 Model SPY
3-Yr. Gross +195.7% -15.5%
C.A.G.R. +43.5% -5.5%
Sharpe 1.46 0.21
Maximum +19.5% +13.3%
Minimum -11.1% -19.8%
Average +0.8% -0.0%

Voila! This is not only a major improvement in terms of CAGR, but the Sharpe ratio is nearly 1.5 which is much higher than the previous model. Considering that turnover does not occur too frequently, this is a fairly realistic backtest- though obviously testing over longer periods is recommended. Nonetheless, given that the model is logical, there is no reason to believe that this will not perform well in the future. Note that I am using this model and a couple others to trade longer-term capital.
In the final installation, we will discuss some final refinements that are very original and help to improve switches.

29 Comments leave one →
  1. jrudy permalink
    February 24, 2010 6:06 am

    Hi David,

    could you send me the excel file of the model, or at least the description how exactly it is done?

    • david varadi permalink*
      February 24, 2010 8:55 pm

      hi jozef, unfortunately you will need to get ETF Rewind to be able to use this application. The algorithm is undisclosed. I wish I could help you further–there are no options other than that.
      best
      david

  2. Damian permalink
    February 24, 2010 7:22 am

    I like the following combo: SPY, EEM, SHY, FXE, TLT, DBC

    • david varadi permalink*
      February 24, 2010 8:56 pm

      hi Damian, I tested that one out last night ironically. Its certainly a good choice.

      best
      david

  3. prazor permalink
    February 24, 2010 7:39 am

    Great stuff!

    What about if transaction and other cost where to be considered.
    Roghly how much would it affect the stats?
    I know this is very dependent on capital, broker, type of
    pension solution etc. But are we taking half expected cagr or…?

    How many trades a year are there?

    How sensitive are the stats for delay in execution of the
    trades? Eg. what happens if you trade monthly or even quarterly?

    How does your Livermore index compare to Rewinds model?
    I mean if used with the same set of intstruments and only
    selecting the top one.

    Looking forward to next part!

    • david varadi permalink*
      February 24, 2010 9:00 pm

      hi, thanks a lot. the tranasaction costs were not factored in as indicated but at .15% round trip and a typical trade lasting quite a while, this will not hurt the return more than say 5% compounded if that.
      i can’t test the parameter sensitivity unfortunately as it is not my own model. As for the comparison of the livermore to this, they are very different– the livermore is not purely a Relative strength rotational model, it looks for active and “trendy” stocks using a quantitative model that perform well using simple trend-following strategies. that said the hedged top 10 livermore had higher absolute and risk adjusted returns over a longer test period. but they are complimentary, and the correlation between the two models is fairly low.

      best
      david

  4. mike permalink
    February 24, 2010 7:55 am

    with all due respect, anyone can construct a backtest to look really spiffy. people have been doing this for literally decades. you cannot honestly believe thata three yr super optiimized backtest will bear any resemblance to future results, do you? if so, you are naive beyond description.

    • david varadi permalink*
      February 24, 2010 9:05 pm

      hmm, i think you missed the point of the article—which is not to extrapolate performance but rather to show how uncorrelated model components perform better than highly correlated ones. if you believe that my intention was to demonstrate otherwise than you are obviously not familiar with my work or writings. I have always emphasized robustness vs optimization—just take a tour of my past blogs.
      Since i had no control over the backtest length and am using someone else’s product, I think it is a little unfair to make such harsh statements.

      best
      david

  5. Jon Tresslar permalink
    February 24, 2010 9:51 am

    Hi Dave, thanks for the two articles which are great! I think your last model has ticker symbol DBC included as well. At least that is what I nneded to reproduce results similar to yours.

    Jon

    • david varadi permalink*
      February 24, 2010 9:06 pm

      thanks jon, you are correct,
      best
      david

  6. eber terandst permalink
    February 24, 2010 11:48 am

    This model, like all rotationals, is strongly dependent to the basket of vehicles to trade. It is clear that the less correlation between these vehicles, the better.
    The problem is that these correlations have drift and volatility. How do you measure correlation, over what time span, and how often you change the basket due to changes in correlation ?
    eber

    • david varadi permalink*
      February 24, 2010 9:11 pm

      absoluted right eber. there is a distinction between “inherent” correlations and current 20-day correlations. “inherent” correlations are logical relationships that are based on structural relationships. ie T-bonds will always have a low correlation to stocks because one has less credit risk than the other (although that is strangely debatable at this point—who is more creditworthy, JNJ or the US Govt?)
      a more dynamic model could us optimization procedures to allocate based on covariance etc every day or every week overlaid on this model.

      best
      david

  7. Albert permalink
    February 24, 2010 1:15 pm

    David,

    Did you measure the peak drawdown?

    Cheers
    Albert

    • david varadi permalink*
      February 24, 2010 9:11 pm

      hi Albert, the output does not show the peak drawdown, but it seems to hover between 10-15% just eyeballing it.

      best
      david

  8. February 24, 2010 6:55 pm

    A 3 year backtest with Canadian REITs thrown in – wow, I’m impressed!

    • david varadi permalink*
      February 24, 2010 9:12 pm

      another sarcastic comment, WOW thanks for the contribution……….please say something useful next time.
      best
      david

      • February 25, 2010 2:05 am

        Apologies! I didn’t read thru to note that you were using the ETF Rewind tool.

  9. Dave Svilar permalink
    February 24, 2010 7:53 pm

    Thanks for sharing your results. It seems like you were shooting for the standard asset classes, then you throw in XRE.TO. Interesting. For what it’s worth, I played with the model and did not get anything near the returns you posted today. Could be operator error, but seems difficult to mess up. Looking forward to Part III.

    • david permalink*
      February 24, 2010 8:53 pm

      hi dave, i tested it in the rewind application a couple times and copy and pasted to the blog—that is strange, did you double check the tickers? i also know that you need to test it at night for it to be accurate vs during the day—can’t remember the reason for that.

      best
      david

  10. February 24, 2010 11:21 pm

    JB: FYI, removing the Canadian REITs from the equation still leaves a CAGR of 40.5% and a Sharpe of 1.33. The two points that I am left with are: 1) It is conceptually correct to employ noncorrelated assets in a trend-following rotation model, as they will generate the highest return when adjusted for risk, and otherwise, and 2) The ETFRewind Model is “sound,” to put it mildly. For you to make such a statement indicates you have never read Mr. Varadi’s body of work, as laid out in his blog.

    Mike: Your comment about this being a “super optiimized (sic) backtest” might hold water if a) the asset classes were highly specific (e.g. the Polish zloty or the Belize dollar), or b) the model was optimized just for this test. However, the asset classes are the broadest imaginable, are represented by very liquid ETFs, and the model cannot be tweaked by anyone other than its creator (not Mr. Varadi).

    David: Great job, as always.

    • david varadi permalink*
      February 25, 2010 1:01 am

      thanks K, much appreciated.
      cheers
      dv

  11. Aristotle permalink
    February 24, 2010 11:26 pm

    David, as I’ve mentioned before, you have a great blog. I’m familiar with a model very similar to this one tested over a few populations and a very long time frame. It holds up well, great work!

    • david varadi permalink*
      February 25, 2010 1:02 am

      thanks aristotle, glad to hear you can confirm that this is a robust model over longer periods.

      cheers
      dv

  12. George permalink
    February 25, 2010 6:38 am

    Hi David,
    1.
    Would you consider modifying this line of yours in the post?:

    So our final model list of tickers will be: SPY, EEM, SHY, SH, FXE, XRE.TO”
    to
    “So our final model list of tickers will be: SPY, EEM, SHY, SH, FXE, XRE.TO, DBC”
    As other commenters and now me can verify the DBC is to be there to reproduce the 43% CAGR you mentions.

    2.
    Could you plz. explain why you picked specifically the Canadian REIT?
    There are other REITs in the stock universe. Why did you pick exactly this one?
    One reason against it is that it is on another exchange in another currency.
    And I am not sure that the USD/CAD currency rate doesn’t affect the performance of it.
    Instead of this XRE.TO what other REIT (preferable US) would you pick?

    3.
    Just as a contribution to the discussion:
    I noticed while I backtested myself that the %Invested column is zero for SPY, SHY and FXE.
    This means that the rotation model doesn’t use this 3 ETFs.
    So, basically, they can be eliminated here.
    So, the 4 ETFs what is left (EEM, SH, DBC, XRE.TO) gives the same performance charasterics (CAGR, sharpe, etc.)
    than the strategy with the proposed 7 ETFs (SPY, EEM, SHY, SH, FXE, XRE.TO, DBC).

    Backtesting even more, a comment here said that killing XRE.TO from the list doesn’t decrease the CAGR too much.
    So, I did run the backtest.
    With only these 3 ETFs (EEM, SH, DBC)
    It shows CAGR: 40% with Sharpe: 1.29.

    I dare not eliminate the DBC from the list.šŸ™‚
    But if I do…
    It reverts back to a simple go long (EEM) when bulls are in charge and go short (SH) when bears are in charge.
    Would you interpret this in any other way than me?

  13. Brad permalink
    February 25, 2010 6:54 am

    I think something that most are missing here is the time frame that the backtest was run on. The stellar results were a function of the massive bear market in 98. The model’s ability to produce positive returns in that market are mostly at work here. If you break down the time periods, the returns for the data set from the bottom on 3/2/09 were 25.4%, vs SPY at 64.7%. As David has intimated in many other posts, RS strategies like this desperately need a bear market to look good, and often lag early in a bull market.

    • david varadi permalink*
      February 25, 2010 4:40 pm

      george, i agree with #1 that was a typo and i will change that, DBC is in the backtest. to your next point i picked the Canadian REIT because I am Canadian, so that was just a personal bias–would rather own Canadian Real Estate vs US any day. replacements could be VNQ in the US, but I would caution that it has much less diversification benefit bc it has a very high correlation to the S&P500. to your next comment, just because the model did not use those components in the past, does not mean it will not use them in the future. you are correct that the model could be stripped down in this backtest, but the future is unknown and there is no reason why cash (shy), the euro(fxy) or spy may not outperform the others at some point. as for your final comment I would agree with you over the last several years this has been true. however this may not be true in the future if China runs into serious problems like Japan did. in other words, its always hard to tell which will outperform in the future.

      best
      dv

      • George permalink
        March 1, 2010 5:31 am

        David,
        I could see one possible reason ‘why cash (shy), the euro(fxy) may not outperform the others’ (as you said) at some point in the future.
        For example, if the ETFRank is constructed in a way to favorize ‘bigger’ movement.
        We don’t know how ETFRewind’s EtfRank is calculated, but I guess a very stupid ETFRank calculation like this:
        ETFRank = %priceGain(lastweek); // (I told, it will be stupid)
        Let’s suppose cash (shy) end Euro (fxy) has much less volatility (like 4-10 times less) than stocks (spy).
        If SPY has a high volatility, it is very rare to see that we have a constant SPY during a week(like not going up and not goind down in the short term).
        So, either SPY (bulls) or SH (bears) will probably have a higher ETFRank (%priceGain(lastweek)) than cash and Euro.
        If we include both SPY (or the emerging equivalent) And SH (the short ETF) into the tested ETFs set,
        the cash (SHY) has almost no chance to be selected as the Top ETF based on the aforementioned ETFRank in the future.
        I don’t say I am right that cash (SHY) will be never selected with ETFRewind’s EtfRank; (never say never)
        Let’s say: I just have this suspicion.
        compare the volatility of SPY to the cash SHY:
        http://finance.yahoo.com/echarts?s=SHY#chart3:symbol=shy;range=6m;compare=^gspc;indicator=sma+volume;charttype=line;crosshair=on;ohlcvalues=0;logscale=on;source=undefined
        Thank you for replying. It is great to have this conversation here.
        Everybody puts his little piece to the puzzle (I really like other’s commens, warnings) and the collaboration
        I am sure will result a fabolous strategy. The utmost thanks of course goes to the author of the blog,
        since he invest the most time into this. Great work.

    • david varadi permalink*
      February 25, 2010 4:53 pm

      hi brad, actually i will have to disagree here—-the only bear component of this RS model was SH, and during this period there weren’t as many good long opportunities prior to last year. In contrast if we moved this backtest to 2005, it would have done 60% or more since emerging markets EEM and commodities DBC were on a tear and had virtually no standard deviation during that time frame to boot. I do agree that RS strategies struggle when the market transitions in both directions (from bull to bear or from bear to bull), but they certainly do not do well on average in bear markets. The livermore had its best performance from 2003-2007, even the hedged version.

      best
      david

  14. Brad permalink
    February 25, 2010 7:44 am

    I meant 2008, not 98 LOL. To be clear, RS strategies that include data set’s such as the one David selected look great when they have a bear market in the backtest. I think it is really important to look at other time frames to see the strengths and weaknesses of an asset class rotational approach. That said, I have seen a ton of different takes on asset class rotation and Market Rewinds approach here is stellar in every respect, and really should be the bedrock for any portfolio as it really will keep you out of trouble in a bear market. Marry it with a few alpha generating strategies that work outperform in a bull market and you have great portfolio.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: