In Part 3, we examined the effect of weighting two different RSI lengths, and using those weights to rank ETFs. After backtesting, the resulting statistics and equity curve appeared to be curve fit, and I acknowledged that the weights chosen were optimal. I ended part 3 by noting that the next step is to test a moving average filter vs. using short side ETFs. We will examine those two factors in this installment.
First, a note about the ETF portfolio. I began tweaking the portfolio a bit based on the recommendations made by a reader in the comments section here. I think his recommendations are important because the portfolio has some redundancies. The redundancies will mean that it is possible to be over-weighted in certain sectors or countries. For example, if the system chose [[EWZ]] and [[ILF]] , it would have 50% of its capital invested in highly correlated ETFs. For this reason, I believe it is necessary to weed out highly correlated ETFs. I also propose removing [[USO]] and [[UNG]] due to the contract rollover shenanigans.
In terms of adding ETFs, there are more that could be added to give broader exposure.
The point of this discussion is that the portfolio is almost as important as the factors of the system itself. After making changes to the portfolio, the results dropped fairly significantly from what was reported in Part 3 (another sign of curve-fitting?) I decided for purposes of continuity and consistency to use the same portfolio in Part 4 as in Part 3 so that we could have an apple to apples comparison of the moving average filter, but I think we should spend more time refining the portfolio. My gut feeling is that it is too large, and because of that, it makes it more difficult to isolate what effects (if any) the factors have on performance.
On to the tests…
Lets Remove the Short ETFs
I removed [[SH]] , [[PSQ]] , [[DOG]] , and [[RWM]] and then ran the test with the system using the same parameters as in Part 3. Here are the results:
The Annual Return % drops over 3% and the Max System % Drawdown increases ~2% when the short ETFs are removed.
Add Short ETFs Back to Portfolio and Add a Moving Average Filter
Now I will add the four short ETFs back to the portfolio and add a 200 day simple moving average filter. The filter will not let the system rotate into new positions if the SPX has closed beneath its 200 day SMA. As soon as the SPX closes above the 200 day SMA, the system will jump in long on the next open.
Note that the 200 day moving average filter does not allow the system to rotate into the short ETFs, so even with the ETFs added, we have an apples to apples comparison of the system without the short ETFs versus the system with a moving average filter.
The Results:
Summary
We now see that the performance without short ETFs is very similar to the performance with a 200 day SMA filter, in terms of the compound annual return. However, the max system drawdown is ~7% less when using the moving average filter. That is a ~20% improvement in the drawdown.
The moving average filter also improves other metrics, such as average trade, profit factor, and the Sharpe ratio. And since the filter means the system makes fewer trades, if commissions were added, there would be even greater improvement shown in the moving average filter results vs. no filter and no short ETFs.
A Few More Thoughts…
I ran an optimization to determine the best moving average length for the filter. I graphed two metrics, Compound Annual Return (CAR) and CAR/Maximum System Drawdown (CAR/MDD). The results show that there are many moving average filter lengths that will work, but it appears that lengths greater than 170 bars give fairly smooth results with lengths greater than 220 resulting in improved performance.
What Have We Learned?
I’m not sure. There are still a lot of variables here, with portfolio composition being a formidable one. I would say that short ETFs helped results more than I thought they would, and I am always surprised how well a simple moving average filter performs. I have learned that RSI seems to be sub-par as a ranking tool, compared to Rate of Change. More than anything, these tests have created more ideas for future tests. Ultimately, I hope to be able to synthesize all that I have learned into a robust rotational system. I do not believe that this system is a good final product and therefore I will take the good, leave the not-so-good, and push forward.
Where Do We Go From Here?
Long story short, I’m done messing around with RSI as it was used in the tests as a tool for ranking ETFs in a rotational system. I’m positive that Rate of Change is a better ranking tool. However, I am interested in using RSI in a different way, based on the comments left by Ruschem:
“I was thinking about your approach to ranking ETFs. While I like RSI based system (RSI is one of my favorite indicators) there are a few points that I am not quite sure about. First, using blended RSI(65) and RSI (30) is somewhat redundant because the entire RSI (30) set of data is already included in RSI(65). Adding RSI(30) to RSI(65) simply adds some more weight to the last 30 days regardless in what proportion the two are used. Instead, I suggest using three independent periods. For example, X*RSI(21) today + Y*RSI(21) 22 days ago + Z*RSI(21) 43 days ago. In all, this system will use 65 days worth of data but split into three independent (or almost independent because of the way RSI is calculated) periods. What it does, it measures the consistency of outperformance or underperformance in each of the last three months. I did RSI ranking both ways (yours and mine). Most numbers were very close but there were notable differences as well. I didn’t run the backtest because I don’t have good backtesting software and lack skills to do it in excel.
Another problem that I see with RSI ranking is that it’s a measure of the internal strength and it really doesn’t tell the whole story. As a result, an ETF may have a high RSI rank but only a very low absolute return. This is why it seems that some kind of a combined RSI and ROC ranking should do the best. Ideally, I would filter ETF pool using RSI ranking (maybe 8 best) and buy four with the highest ROC out of these eight. What do you think?”
I will begin playing around with Ruschem’s idea, to see if it has any merit.
I also want to dig deeper in the following areas:
1. Portfolio construction
2. Shorter minimum hold times
3. Better weighting and smoothing of the ranking metric
Thanks for reading and contributing ideas.
Hey Wood,
Do you subscribe to the Premium Data for Futures? Have you considered backtesting against something like this? The reason I ask is that it would provide a diversified portfolio of investments with a longer historical period than ETFs? Just a thought.
I don’t subscribe to the Futures at PD. I would love to have the data but just haven’t had the time to expand my testing to that area yet.
What would you use to switch among RedShark, single stock futures?
I second that question Red, but I’m guessing your thinking Wheat, Corn, Gas, etc.
Sorry for being vague. I was thinking more along the lines of Wheat, Corn, Gas as Wood suggested but even financial futures as well (Stock Indexes, Interest Rates, etc). The reason for my suggestion is Premium Data offers also offers continuous market contracts. For more information -> (http://www.premiumdata.net/support/futurescontinuous.php)
The reason I suggested this is since ETFs have such a limited amount of historical data and largely track these Futures contracts. So perhaps you could use the Futures Data as a proxy for the ETF data since the ETF data is limited. Also, then you would not have to worry about which ETFs should be backtested (i.e. inverse, double inverse, etc).
Wood: i wonder on what basis do you believe that ROC might be a better tool than RSI? From your tests as well as performance evaluation of different ET’s during different market stages (bull/bear) it seems that portfolio composition plays more important role than tools to measure RS. If you include bear market instruments, bonds or money market instruments is ROC (especially long term ROC) going to be an appropriate tool as these seem to have rather low volatility compared to indices or commodities? Regardless thanks for an interesting study.
Hi ji, a friend sent me some research comparing ROC and RSI for this type of rotational system. That, coupled with my own research leads me to believe that ROC is superior. I should be more specific and state that for a system with these general parameters, ROC seems to work better. I’m sure there are examples of rotational systems with certain parameters where RSI outperforms ROC.
You didn’t ask for it, but my gut feeling is that ROC is a more “pure” ranking of momentum.
You make some very good points though, re: ranking volatility oninstruments that are not very volatile. I’ll have to think about that and make some observations.
Wood –
I’ve emailed Jeremy with some specific questions regarding your system — looking forward to hearing back from him regarding that.
Working Man,
Check your email. I sent you one just before I saw that you left this comment.
Running the Fidelity Select Rotation system here. I have FSHCX in the #1 spot followed by FSAIX. I am rolling with 50% into each.
Woody
Have you tried to rank the ETFs with a Sharp Ratio or 2 Sharp Ratios ie Return/Vol?
Have you tried t put trailing stops?
How about leverage?
Also what tool do you use for your simulations thank you
Andy, the FSF system ranks on return/volatility. The ETF system did not do very well when I penalized volatility, but I did not test it exhaustively. Your idea of ranking based on Sharpe is interesting, although Sharpe is used to describe returns relative to risk, and not the return of an individual security. Still interesting though.
Woody
Thank you for your answer. Also do you have the possibility to test with trailing stops/ or stops at all?