Sector Selection / IWM comparison: new Robot call 8/24/2011

[Timothy Clontz]

Since the new Robot and stops have been initiated, Billy has been kind enough to update his algorithm pivots for the sector rotation test.

The previous test seemed to indicate positive synergy with the Robot, but with the new Robot and new stop calculations I'll begin this test as an entirely separate test.

Based on Billy's pivots and my sector selection, then, I've entered a new limit order this morning for:

XLK, buy at 23.10, with a stop set at 21.63.

[Timothy Clontz]

XLK did not hit its entry price.

New order is:

Buy XLK at 23.72, with stop at 22.36.

[Timothy Clontz]

XLK hit its entry price at 23.72 on 08/25/2011.

Current stop is 22.36.

A hypothetical new position would be:

Buy XLK at 23.10, with stop at 21.78.

[Timothy Clontz]

XLK hit its entry price at 23.72 on 08/25/2011.

Current stop is 22.36.

A hypothetical new position would be:

Buy XLK at 23.72, with stop at 22.28.

[Timothy Clontz]

XLK hit its entry price at 23.72 on 08/25/2011.

Current stop is 23.09.

A hypothetical new position would be:

None (per the IWM signal).

[Timothy Clontz]

Now that the new Robot is online and the backtest trade history is available, Billy was kind enough to encourage Pascal to spend ALL DAY LONG analyzing my own backtest sector calls for a compatibility check.

Many many thanks to both of them.

As it stands, the current settings on my sector rotation model are not compatible with the IWM Robot.

I suspected my short selections could underperform the Robot (since my model was primarily designed to be long-only), but the long selections underperformed as well.

At first I thought it had to do with the holding period. My model is designed to hold a trade for months at a time, potentially up to a year. The Robot holds long trades for an average of 23 days and short trades for an average of 14 days.

I ran a test of my own model showing returns of all long selections for a period of 23 days at a time (rather than one day at a time). My long sector selection vastly outperformed all other sectors when using 23 day running returns going back to 2003.

And yet it did not work on the specific days that the Robot made a long signal.

That is, my model selections outperform on the total of 23 day holding periods, and yet still underperforms the IWM Robot.

Billy and I had also speculated earlier that the higher Beta of IWM could outperform my selected sectors because of large cap sluggishness in the ETFs I use, but that also did not seem to be the case. There is something about the specific time triggers of the Robot that (almost perversely) seek out the underperforming periods of my sector selections. Pascal also ran a series of random sector selections and 6 out of 8 random sector selections outperformed my own model's chosen sector for the specific time periods of a Robot's long call.

That is, my selected sector was almost the worst performing sector during the specific Robot timing calls -- much worse than a random selection.

This is a most peculiar and fascinating discrepancy, but also an expensive one. Until I can get to the root of this, I will stop testing with live money once this current market call terminates.

Another perverse factor that I noted is that my sector selections outperformed the Robot when it had a problem with its own signal. In other words, when the Robot works my sector doesn't, and when my sector worked the Robot didn't.

Since they've upgraded the Robot I do not expect that peculiarity to continue.

As it stands I have some puzzles to piece together, and will not resume these tests live in the Algo forum until I can determine the specific causes of the negative synergy between the two models.

The Mousetrap model will continue to trade, however, since it is unrelated to the Robot signals and appears to be beating the S&P within the target range (i.e. S&P +30%).

Tim

[Timothy Clontz]

Just to be specific, these are the numbers for the average return for each sector rank in my model for 23 day rolling returns and 14 day rolling returns:

(NOTE: there are 10 listed instead of 9 because I included the S&P as the "average sector", normally in position 5 or 6)

For 23 days (long)

1 1.16%
2 .81%
3 .25%
4 .52%
5 .52%
6 .51%
7 .88%
8 .25%
9 .11%
10 .54%

For 14 days (short)

1 .69%
2 .63%
3 .11%
4 .34%
5 .26%
6 .32%
7 .51%
8 .07%
9 -.01%
10 .31%

As can be seen, in both holding periods, rank 1 vastly outperformed, while rank 10 performed at or slightly below average.

I'll do some work optimizing these kinds of holding periods, but even if the model were optimized for these short periods, there is the nagging problem of it not working correctly with the listed trigger dates on the IWM Robot.

Tim

[Pascal]

TIM,


It is almost an impossible task to optimize a specific holding period (23 days or 14 days) for your ETF selection method and apply it to the IWM Robot method that is far from these cycles. Below is a small statistics of the market direction calls for the IWM robot (Both strong and weak calls). If you take only the calls that last 25 days or more, you will see that they cover 17 trades (21% of the total number of trades), but that they include 71% of all the trading days. This means that the IWM robot is mainly a trend following method working on longer cycles than 23 or 14 days. Also, when a trade starts, we do not know if it will be a 2 days trade or a 50 days trade. Therefore, you cannot apply a specific cycle at the start of the trade, which is what you are trying to do.

I however believe that your investment method could work very well independently of the robots and I encourage you to post on the VIT forum your findings about the ETF cycles and how to select the best ETF within their cycle.


Pascal

10151




Just to be specific, these are the numbers for the average return for each sector rank in my model for 23 day rolling returns and 14 day rolling returns:

(NOTE: there are 10 listed instead of 9 because I included the S&P as the "average sector", normally in position 5 or 6)

For 23 days (long)

1 1.16%
2 .81%
3 .25%
4 .52%
5 .52%
6 .51%
7 .88%
8 .25%
9 .11%
10 .54%

For 14 days (short)

1 .69%
2 .63%
3 .11%
4 .34%
5 .26%
6 .32%
7 .51%
8 .07%
9 -.01%
10 .31%

As can be seen, in both holding periods, rank 1 vastly outperformed, while rank 10 performed at or slightly below average.

I'll do some work optimizing these kinds of holding periods, but even if the model were optimized for these short periods, there is the nagging problem of it not working correctly with the listed trigger dates on the IWM Robot.

Tim

[Timothy Clontz]

Pascal,

I understand that the Robot has variable holding periods, and that form fitting to past calls is a fool's errand.

Rather, the issue that I'll try to resolve today is the fact that I designed the model as a guide for fundamental investing (as in the Mousetrap), and deliberately formatted it for technical mean reversion over a year.

What gave me the idea was when I analyzed my friend Len Mansky's market timing model and found that automated up and down calls had better performance in a year (% return * 1) time frame than in a quarter (% return *4), a month (% return * 12), or a week (% return * 52). It wasn't much of a stretch for me to incorporate such a model for sector and industry selection for fundamental investments -- which ALSO depend on mean reversion over the course of a year.

Here's the rub: successful technical investors usually trade in faster time frames than typical retail investors; and successful fundamental investors usually trade in slower time frames than typical retail investors. Both HFTs and fundamental investors profit over the unfortunate retailers, who sell to the fundamentalists low and buy from the HFTs high.

The Robot is designed to compete against HFTs, using sophisticated data and even more sophisticated calculation.

My Moustrap and Sector Models are designed to NOT compete with HFTs. I try to find oversold disasters caused by the HFTs so I can pick up the debris they leave in their wake.

These are completely contrary approaches.

Put simply, you make money following effective volume in real time. My model makes money finding pent up non-effective volume that needs to mean revert.

I do have one test that I plan to do today, which is adjusting the measuring period from more than a year to about a month. After measuring that effect on 23 and 14 day holding periods, I'll see if those average values carry-over into the actual dates of the Robot's past calls. If they ARE consistent, then I can adjust my real-time testing and continue. If they are NOT consistent, then the problem doesn't have to do with the holding period, but a fundamental incompatibility between non-effective volume mean reversion and effective volume timing triggers. I was hoping for a positive whiplash effect, but the concept itself could be flawed at the core.

I should have a good idea today.

Tim

[Timothy Clontz]

Models are like people. Sometimes you can have two perfectly great individuals who make a bad couple, and sometimes two unexceptional individuals who make an incredible couple.

Some years ago I tested Value Line's timeliness ranks with Joel Greenblatt's GARP formula to see if they worked together. What I found was that Value Line 1s DID beat Value Line 5s, and good GARP scores DID beat bad GARP scores.

But I ALSO found that the best combination of the two was NOT good Value Line scores and good GARP scores -- but instead BAD Value Line Scores and good GARP scores.

That is, they had negative synergy -- even though they each worked perfectly fine alone.

For the past few months, Billy has been giving me his pivots for nine sector ETFs in order to have the most solid test we could of my sector rotation model using IWM buy and sell triggers. At first it appeared to work well, but then the Robot had to be taken offline and updated. Once that work was completed, Pascal spent a good solid day analyzing my sector selections going back to August 2007 along with the new Robot trigger history.

Pascal also went over and above a simple test and analyzed my sector calls against completely random sector selections.

What he found was that my sector calls were significantly worse than random. 7 out of 9 times that a random sector was selected it performed better with the Robot than my targeted sectors.

This is much like the NEGATIVE synergy I observed between Value Line and GARP scores. Value Line uses positive momentum as part of its formula and GARP selections have negative momentum.

Pascal's model uses effective volume (i.e. volume confirmed with price movements).

My model uses non-effective volume and breadth (i.e. volume and breadth NOT confirmed with price movements).

While my model is designed for investors using some kind of GARP or value investing, Pascal's model appears to work better with the kind of fundamental selection used by IBD. GARP (growth at reasonable price) buys at low P/E and sells at average P/E. IBD buys at high P/E and sells at higer P/E.

Not only would we enter different sectors at different times, we would have different holding periods and different stock selections even within the same sector.

Both of these are perfectly reasonable approaches, with different holding periods. With a good enough model (which Pascal and Billy have), one can make a killing actively trading and shorting the market at precisely targeted times. With my model one can have a good solid return by either staying long or occasionally hedging, and holding these positions for months at a time.

The average holding period for the Robot is (so far) 23 days long and 14 days short.

The average holding period for my model is much longer.

As I've mentioned on a number of occasions, I think Billy and Pascal have an infinitely better model for active traders. My own model is for folks with 401ks who don't want to worry so much about timing.

In any case, this has been a successful failure. The purpose of a test is to determine how something works. Once that determination is made, the test itself is a success. The conclusion is that effective volume cannot be used as a trigger to release excess accumulations long term non-effective volume. Effective volume finds an immediate trend and uses it. Non-effective volume finds an exhausted trend and waits for it to mean revert.

More money is to be made with the Robot, of course.

But Pascal and Billy have been kind enough to suggest that I write in more detail about sector rotation in the VIT forum, and I plan to take them up on it.

Thanks to both of them for their time and help.

Tim

[nickola.pazderic]

I want to read your report, too. Thanks.