Do you have any suggestions for ETF rotation models?

I’ve looked into various asset class ETF rotation models. They can frequently serve as useful hedges in a traditional stock strategy. I looked at a few here:

The following are the most intriguing:

To mitigate risk, they use momentum as an indication and invest in low-correlated asset classes. On the last trading day of the month, they make a change to the portfolio.

Then I tried to look into some of those that are publicly available on P123.

Those who perform best have a very high turnover rate or a large number of ETFs:

Is there anyone who have a proposal an asset class rotation portfolio that you have found to be effective?

Hi Marchus,

I rebalance an adaptive allocation model today. Showing a backtest of my model would be pure BS seeing as it is a backtest. I could debate how overfitted it is but not whether it is overfitted.

I think your observations about the number of ETFs may be a good one. Using a large number of ETFs works for 2 reasons, I think. First, it provides diversification.

But second, a strategy with too few ETFs is often “over-betting” in the Kelly Criterion sense. Another way to say the same thing is that a strategy with too few uncorrelated assets can cause too much volatility-drag. Kelly over-betting and volatility-drag are the same thing. Or more simply, in the spirit of original discussions of over-betting, even if the rest of the deck is all aces and face cards (making blackjack more likely) you should not bet all of your retirement funds on that one hand even though the odds are with you. Betting no money on that hand would be even more lame. There is a right amount to bet in order to capture the “edge” without to much volatility or “over-betting.”

One way to illustrate what I am getting at is to take a portfolio of 33% each of XLU, XLV and XLP (“provided portfolio” in image). The average return (or “expected return”) each month for this portfolio is less than SPY. And yet the CAGR for this portfolio is better than SPY. This it because a portfolio with less volatility, obviously, has less volatility-drag. Volatility affects returns (and not just drawdowns).

To be sure, there are better ideas that using just XLU, XLV and XLP in a portfolio. This is all just to reinforce that I think you are right in your observation that the number of ETFs makes a difference. And avoiding strategies that will put you in 100% QQQ one month, 100% GLD the next month and TLT the month after, is probably wise.

Some of the adaptive allocation models will underperform SPY just because some individual assets (like TLT) do not have returns as good as SPY. A modest amount of leverage for those underperforming assets may take care of that (no matter what strategy you end up using).


Yes, Seasonal ETF switching works well. Switch between 2 groups of 5 ETFs at the end of April and end of October. So you only have to look at this twice a year.

Summer group from end of April: TICKER(“VIG,XLP,XLK,XLU,TIP”)
Winter group end of October : TICKER(“XLY,XLI,XLB,XLV,VBR”)

Use P123 ranking system “ETF Rotation - Basic” to select fewer than all of them.
See performance for 2 holdings below from mid April 1999 and with 0.1% slippage.
CAGR= 14.8%
Max D/D= -35%
No need to fool around with volatility and other useless criteria.


Seriously, Georg. You should start thinking about volatility drag with you sector SPDR designer model don’t you think?

Here comes Georg with his overfit and cherry-picked models. No surprise here.

The median of your Designer Models with 2 year excess returns has excess returns of -28.91 This is with a minus. A big minus number. That is after you removed the really bad ones. I guess you had to keep a few.

If you want to sell your cherry-picked and overfit models to P123 members [color=firebrick](and P123 wants to encourage this after has Marc left)[/color]) it is fine with me. But looking at volatility-drag would help your designer models. You should try it.

Just sayin’


Hi Jim,
I don’t know where the overfitting is in the 10 ETF model I presented.

Perhaps you can provide some positive input instead.


I did provide positive input: you should look at volatility drag in order to get some some decent Designer Models for P123 members.



Hi Jim,
You can have a look at my latest ETF model on Seeking Alpha which uses six of my market timers to produce risk-on, risk-off, and risk-neutral signals.
It shows a 34% annualized return. To the best of my knowledge nobody else has come up with a market neutral signal in the context of market timing.
1,260 people have looked at this, which is a good advertisement for P123.

Looking forward to your negative comments.


I get that you do not understand what cherry-picked means, or you do and you think the people you are selling your models to do not understand the term.

Either way not really something you should be bragging about exactly.



The reason that a portfolio of XLU, XLV and XLP provides a marginally higher return than SPY is that XLV alone outperforms SPY by a lot. This has nothing to do with volatility-drag. See the two figures below.

Jim, it would have been prudent to do this simple backtest before lecturing us about volatility drag. Others have also found that there is no rational to this. See “The Myth of Volatility Drag”

the author (also an engineer) concludes “Let’s banish “volatility drag” from our vocabularies!”


Why does a portfolio consisting of XLV, XLP and XLV make you feel so insecure? You get that this was not a portfolio that I was recommending or using don’t you?

Seriously, is everything okay?



Jim, you don’t have to be rude.

So what if the AVERAGE RETURN for SPY is higher while the CAGR is lower than that of the Combo XLU+XLV+XLP.
Why don’t you design a model which makes use of this fact and post it so that we can all benefit from your wisdom.

I note that you have not posted any Designer Models presumably because you don’t want to be criticized if your model does not perform well, but do want to have the prerogative to criticize others.


It is you who are rude.

You started it but this quote is the least of it. I find your selling of cherry-picked and overfitted models unethical. I am most offended by your turing this thread into your personal sales pitch.

[b]Marc would have had none of it. He asked you to stop (as he should have).

As I recall he threatened to charge you as a pro if you did not stop. He took offense also. You stopped while he was here. Please expand on that conversation if I got anything wrong.[/b]

Unlike your models Marc’s median model does not dramatically underperform the benchmark which had a lot to do with his view on unethical selling of [color=firebrick]useless[/color] models, I think. Marc never charged for his models.


For most of the models you can create them yourself using the software quantrader by The rules are not the same but the methodology and results are very similar. You can also test out any combination of stocks or ETF. Would love that functionality on P123.


Georg I’ve now seen many iterations of your seasonal ETF switching model and it seems every time it’s a different set of ETFs. Jrinne came on quite strongly but he has a good point about curve-fitting. Why do the ETF options in the baskets change so frequently?

Please list the the different sets that you have seen me use.

You have no response to my challenge to present a strategy that the uses volatility-drag to somehow outperform SPY.
All you can do is to pivot into personally attacking me.

Other than the one you already posted here, these are the other 2 iterations you’ve plugged:

Ticker(“XLY XLI XLB XLv”) vs

Ticker(“XLY XLI XLB XLK VBR”) vs

Note the XLK being swapped and playing a defensive role in one iteration and an aggresive role in the other.

Thank you all for your contributions, albeit I see the argument has gotten somewhat out of hand.

Thank you for linking to the models, Georg, but I, too, am afraid of overfitting, especially in models where I don’t know all of the criteria. Having said that, the seasonal effect (Halloween) on sectors is a well-documented phenomena. So, while I agree that it is not overfitting, you also have some other etf that is unknown to use in a seasonal cycle.

By the way, I’m extremely impressed with your timing model results, and I’ve attempted to understand it. I’m still not sure if it will work out of the backtestperiod. When was this model created?

Jrinne, I agree that a backtest does not always offer much, but it can provide some hints on models that might work. Do you have a link to someone you believe can work and who can be tested on P123?

mv388158, Yes, I attempted to create some of the models in p123. I don’t always get the same results, but I keep on trying.

I don’t want to trade frequently; once a month is about right, and Im not looking for some extreme returns. Its even enough to have the same as the market but but less drawdown. However, I want to use p123 for an asset class rotation portfolio with a specific hedge function for my stock portfolio. So I’m interested in all of the systems offered on p123 that anyone can recommend. Then its possible to altso test them.

For the time being, I’m looking at these models:

A simple take (start) on theese models:

Buy the top tree

CXO Momentum:
Buy the top tree
Ticker(“SHY, TLT, vglt,VNQ, IWM, SPY, GLD, EFA, EEM, DBC voo”) // CXO

Dual Momentum:
Buy the one with highest momentum
In this model VOO and SCZ has to bee in positive terrain to be bought. If not buy TIP or SPLT with the highest one month momentum

Optimum 3:
Buy 3 of the top 6 based on momentum, but choose the tree that is least correlated
My take on the momentum is the same as PAPABear:
I have no idea how to program the pick of the least correlated tree og the top 6.





I was asleep and have not read all of your posts. Not sure that I will. My apologies if I am not responsive to a good question.

This is not a competition.

It is possible that you might be able to recall that Marchus made the point that strategies with more ETFs do better.

I have agreed with this for a while. Going fully into SPY then to TLT then back fully to SPY simply does not work out-of-sample. Nor do other strategies like that (with too-few ETFs) work out-of-sample.

I simply posted that I agreed and gave 2 reasons why that is true: diversification and reduced volatility-drag which are not really separate reasons. You have better reasons I guess?

XLP, XLV, and XLU just took me literally 15 seconds to find as an example of volatility-drag. This actually occurs pretty commonly even within individual sector ETFs and elsewhere. I stated at the time this is not anything I would use as a strategy to invest in. I specifically used this example because it has nothing to do with any of my models and I assumed people would be able to recognize that.

I also said this at the time:

I am not going to start a backtest competition now because you double-dared me. That would be less than meaningless as Marc has pointed out. I think that it is actually unethical–as did Marc–if you are trying to sell a backtested strategy here on P123. Back in the day P123 would have asked you to stop.

Ethical issues aside, one just needs to go to your designer models to see the problem with overfitted backtests. Debating it in this thread will do nothing to change what anyone can find there on their own. What other evidence could one possibly need?

I guess we could pretend that this time is different. Please be my guest everyone. It is fun pretending that you would have known what strategy would have worked best years ago and imagine how rich you would be now.

The real question is why did putting together XLP, XLV and XLU (fixed in equal amounts) after 15 seconds testing–as an example showing that volatility-drag can affect returns–set you off?

[color=firebrick]Are you okay?[/color]

I might see if there are any questions that address the problems that rotation strategies with too-few holdings have later. My appologies if anyone asked a pertinent question of me that I did not answer.


With seasonal timing, slippage of 0.1%, and selecting 2 ETFs with P123 ranking system “ETF Rotation - Basic” and backtest period from mid April 1999:
The first set shows an annualized return of 12.2% with a max D/D= -40%.
The second set shows an annualized return of 13.9% with a max D/D= -35%.
The original set as posted earlier in this thread shows an annualized return of 14.9% with a max D/D= -35%.
Over the same period SPY produced an annualized return of 7.6% with a max D/D= -55%.

What is there not to be liked of the Seasonal Timing strategy, AND WHERE IS THE CHERRY PICKING?
So this is my reply to the question of the thread, and I trust that some members may find this of interest.