July 22, 2015

Multiple Time Frames for Scoring ETF Rotational Strategies

Today we have a guest post from David Weilmuenster who I worked with while at Connors Research.

A widely applied technique for scoring assets in rotational systems is to rank those assets by their price momentum, or return, over a given historical window and to rotate into the assets with higher momentum. This approach seeks to capitalize on the well-demonstrated tendency for price momentum to persist. But, it begs some questions:

  1. “What is an appropriate historical period for measuring price momentum?” Clearly, the momentum of a given asset can rank quite differently compared to the tradable universe over 1 month, 3 months, or 6 months.
  2. “Is one historical period sufficient?” If relative momentum can vary widely depending on the historical window, would it be better to consider multiple slices of history?
  3. Is higher momentum always preferable to lower momentum, especially if the system rules filter the tradable universe before scoring the ETFs for rotation?

Let’s see what we can learn about these questions from a straightforward system that rotates monthly among ETFs that represent the following 4 categories of worldwide assets:

  • Equities (pre-dominantly stocks of operating companies)
  • Fixed Income (bonds, preferred stocks, etc.)
  • Commodities (broad markets, precious metals, energy)
  • Real Estate

 

The universe we will trade is:

150722a

And, the trading system is:

  • Test Dates: 2007-01-01 through 2014-12-31
  • Rotate Monthly, taking signals from last trading day of the month, and entering/exiting on the Open of the first trading day of the subsequent month.
  • 3 maximum positions, with equal size, i.e., 33.3% of Equity. (Results for 2 & 4 position portfolios are shown in a spreadsheet that you can request below).
  • Filtering (each of these is tested separately):
    • No filtering.
    • ETFs are candidates for the next rotation only when trading at prices above 90% of their weekly closes for the past year.
  • Rotational Score: A weighted average of the ranks of 1 month, 3 month, and 6 month returns.

The highest returns are ranked as 1 for each rotation, and lower returns have progressively higher ranks, so a negative weight favors higher returns. If only one of the weights is non-zero, then we are scoring only on the return from the related historical period.

For example, weights of -20, -40, & 40 favor higher returns for 1 and 3 months, and lower returns for 6 months. Weights of 0, -100, & 0, favor ETFs with higher returns for the previous 3 months, regardless of 1 and 6 month momentum.

We aren’t attempting here to design the best possible portfolio, but to generate a credible system that allows us to analyze the effects of different approaches to rotation scoring.

Here’s the performance of the variations that use no filtering and only a single timeframe to score assets for rotation (again, all variations are available in a spreadsheet that you can request below.) Remember that negative scores favor higher returns. “Adjusted CAR (Compound Annual Return)” is calculated by substituting the median 12 month rolling return for the highest 12 month return. Adjusted CAR helps us to avoid a bias toward variations with high CAR that results mostly from a single, exceptionally high 12 month return:

150722b

Even though the results aren’t attractive due to the high drawdowns, performance definitely improves when favoring higher returns (negative weight) for a given historical timeframe. The 1 month historical period is generally the best option in this case.

Next are the “Top Performers” when using no filtering and employing either a single timeframe, or multiple timeframes, for rotation scoring. “Top Performers” are in the Top 10 for every one of the following 4 metrics:

  • CAR
  • Adjusted CAR
  • CAR / Maximum Daily Drawdown (CAR/MDD)
  • Sharpe Ratio

150722c

Note that all of the top performers incorporate multiple timeframes into the rotation score. And, some of the variations emphasize lower price momentum (positive weights) for certain timeframes.

The filtering method described above can reduce the drawdowns. Here are the variations that use filtering and only a single timeframe for scoring:

150722d

Notice we now have an “anti-momentum” effect. The variation scored only by lowest 3 month momentum has the 2nd highest CAR (10.7%, practically identical to the highest CAR), and the highest CAR/MDD, Sharpe Ratio, and Adjusted CAR. So higher momentum isn’t always the most effective choice depending on how one filters the universe of assets and one’s preferred performance metrics.

Next are the top performers when using filtering and either single or multiple timeframes for scoring.

150722e

Now (for my money) the best weighting yields a 12.8% CAR by favoring higher 1 and 6 month returns, and lower 3 month returns. You might prefer another variation, but all of the top performers incorporate multiple timeframes in the scoring.

So, how do we answer the questions at the start of this post:

  • “What is an appropriate historical period for measuring price momentum?”

For this system, the most recent month’s momentum does best when using only one historical timeframe for ranking. But, we had to evaluate test results to learn that. No theory could have predicted that 1 month was better than 3 months, or 6 months. Testing different possibilities is essential.

  • “Is one historical period sufficient?”

For this system, it’s generally better to incorporate more than one historical timeframe in the ranking algorithm. This doesn’t prove that multiple timeframes are better for all systems, but experience suggests that it usually pays.

  • Is higher momentum always preferable to lower momentum, especially if the system rules filter the tradable universe before scoring the ETFs for rotation?

Filtering in this system clearly has a substantial impact on whether to favor higher or lower momentum, and in what timeframe. Again, this isn’t proof for all systems, but a good tip to remember when creating your own systems.

 

How does this example compare to your experience in creating rotational systems?

Backtesting platform used: AmiBroker. Data provider:Norgate Data (referral link)

David Weilmuenster

Fill in for free spreadsheet:

Contains yearly breakdown, different weightings and more.

spreadsheeticon

 

 
Visited 16 times, 1 visit(s) today

Click Here to Leave a Comment Below

Thomas Musselman - July 22, 2015 Reply

I tested using a universe of Fidelity Select Funds from the start of 1985 (when 18 such funds existed; more were added later) to the present, holding for 21 trading days (i.e. a rough month) whichever 3 of such funds had highest momentum, weighting the momentum by adding the total return of (2 x 1 month) + (2 x 3 month) + (1 x 6 month). I.e. doubleweight most recent 1 and 3 month, but normal weight 6 months. Result was a CAGR of 15.3%, gsd of 25.7, Sharpe of .6, Treynor of 13.8, beta of .96, max drawdown -53%

From end of 2006, 10.4% CAGR, sharpe .53, Trynor 11.8, beta .96, max drawdown -53%.

In comparison, using a simple 200 day return top 3 gave 18.86% CAGR, gsd 24.4, .75 Sharpe, and .94 drawdown since end of 1984. From end of 2006: 9.1% CAGR, 24.3 gsd, .48 Sharpe, .87 beta.

So for the Fidelity Select Fund database with its longer test period 200 day momentum is superior.

Quantocracy's Daily Wrap for 07/22/2015 | Quantocracy - July 22, 2015 Reply

[…] Multiple Time Frames for Scoring ETF Rotational Strategies [Alvarez Quant Trading] Today we have a guest post from David Weilmuenster who I worked with while at Connors Research. A widely applied technique for scoring assets in rotational systems is to rank those assets by their price momentum, or return, over a given historical window and to rotate into the assets with higher momentum. This approach seeks to capitalize on the well-demonstrated tendency f […]

IMHO BEST LINKS FROM QUANTOCRACY FOR THE WEEK 20 JUL 15 — 26 JUL 15 | Quantitative Investor Blog - July 23, 2015 Reply

[…] Multiple Time Frames for Scoring ETF Rotational Strategies […]

David Weilmuenster - July 25, 2015 Reply

Thomas, thanks for sharing your results. It shows again that no one approach fits all situations. Doing the research is the only way to know.

When calculating returns in your first example, did you “normalize” 1 month, 3 month, and 6 month returns to a similar timeframe? For example, did you annualize those returns before combining them in your final score? If not, you might give that a try to see how it affects outcomes. The reason is that a 5% return over 1 month is obviously much different than a 5% return over 6 months, so adding them without scaling them to a common timeframe is hard to interpret.

Multiple Time Frames for Scoring ETF Rotational Strategies | SAMUELSSONS RAPPORT - July 27, 2015 Reply

[…] Källa: Multiple Time Frames for Scoring ETF Rotational Strategies | Alvarez Quant Trading […]

Multiple Time Frames for Scoring ETF Rotational Strategies | SAMUELSSONS RAPPORT - July 28, 2015 Reply

[…] Källa: Multiple Time Frames for Scoring ETF Rotational Strategies […]

Tom Musselman - July 30, 2015 Reply

I tried to normalize; if I did it right the 1985 to present CAGR your combo version in the Fidelity Select universe was 15.4% CAGR for top 3 funds each month (using 22 trading days as the “month” not days of the calendar), 23.6 gsd, 15.5 downside deviation, .62 sharpe, .91 beta, -53.8 max drawdown.

From end of 2006 10.0% cagr, 21.8 gsd, 15.4 dd, .56 sharpe, .99 beta, -52.1 max drawdown.

So not really different that way.

aya - April 15, 2016 Reply

Hi Cesar,

Do you always trade on the same day of a month? Suppose, market crashes soon after the last re-balance and your current allocation quickly gets out of whack. Any contingency plan for this?

Just imagine sitting on your hands the rest of the month, waiting patiently for the next re-balance date, amidst the market turmoil! Oct’1987 or May’2010 or Aug’2015 are good examples of rapid price drops without early warning:-(

Thanks in advance.

    Cesar Alvarez - April 15, 2016 Reply

    There is no plan for this. The problem is that you are trying to solve a problem that only has happened a couple of times. What happens if you make a change and your test results are worse? Would you keep it, even if the concept makes sense? This is one of the hard parts of creating strategies is how to handle these situations.

Leave a Reply: