UnThreaded | Threaded | Whole Thread (69) | Ignore Thread Prev Thread | Next Thread
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: of 251809  
Subject: Blending at a Whole New Level Date: 7/22/2008 12:37 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 59
Introduction

Although some may think that moving between blends based on the market's environment is too prone to curve-fitting, in my view the opposite is the case. It only makes sense to me to have the blend match the environment. My greatest concern all along as it relates to curve-fitting was the tendency of most to arrive at a blend using the Solver and all of the screen's history with the thought that what was a perfect match with the past would be a perfect predictor of the future. In my book, that is the premier definition of curve-fitting. Of course, with the work last fall a new standard was set for arriving at a blend. A standard based on using all the history up to the date of the new blend's selection (end of each year) for the next year's investment. Then to take that idea and backtest it going forward from the earliest available data right up to the present to see if it actually worked, not just a hope and a prayer, but a system built on concrete evidence that the method had worked in the past and should continue to work going forward.

It only makes sense that a measure that is the most predictive for bullish periods will not be the measure that is most predictive for bearish periods. How could that not be the case? Obviously, it must be. It is for this reason that I set out on this backtesting journey to determine if it would be possible to find a way of doing this that would have worked in real-time historically.

Determining the Bullish and Bearish Periods

Some may have the thought that this is the most important element in the mix. I don't think so. I've tested exponential moving averages on various indexes and the results are pretty consistent. All outperform a random selection of screens for a blend and most definitely outperform an equal weight of all screens by a huge margin. I've also tested other timing systems and found the results to be consistent. The primary component, therefore, is not the means for identifying the periods, but rather, the consistency in doing so.

So I thought I'd pull out one of my simplest methods for determining bullish and bearish periods. It simply takes the new highs and new lows of the index and subtracts the one from the other and applies a weighted moving average to the result to remove excessive whipsaw. When it is above, say zero, you are bullish, and when below, you are bearish. I've used values I've tuned a bit, but anyone is welcome to get the data from Pinnacle and go through the same process. This results in more trades than with the EMA crossovers, but I prefer it anyway simply because it is something that is tried and true and has a very long history of out-performance. From 1978 the Nasdaq Composite has returned a CAGR of 25.25% during the bullish phases and -8.07% during the bearish compared to a long-term buy and hold of 10.89%. Using this indicator the signal turned to bearish on 10/19/2007 and has remained so till now. Can't hardly beat that, now can you?

Treynor/GSD for the Full History as a Benchmark

As a benchmark I'm using the Treynor/GSD which was selected last November as the absolute best measure when used for the full history -- that is, without dividing up between bullish and bearish periods.

Benchmark  Treynor/GSD
CAGR 31.43%
GSD 16.21
Sharpe 1.57
UI 5.16%
Drawdown -26%

Parameters Used

Begin Date:      1/2/1989
End Date: 11/9/2007
Universe: Value Line
Screens Held: 5
Ranks: 1-4
Total Stocks: 20
Switch: NH-HL
Index: Nasdaq
Smoothing: Yes
Weighted

All Measures Compared

The following table is the fruit of a test where all measures are used for the period up to 7/18/2008. The test compares using Ascending and Descending sort. I'm sorting the results based on the highest return during the bullish period using a descending sort. Here we see that the Sharpe Ratio has finally gotten its revenge since it comes out on top. In fact, Sharpe/GSD descending produces the highest return for the bearish period for the descending sort measures -- some are higher for ascending.

                           ASC      ASC        ASC         ASC       DESC      DESC       DESC        DESC
Screen Bullish Bearish Bull Win % Bear Win % Bullish. Bearish. Bull_Win % Bear_Win %
Sharpe 34.83% 25.83% 59% 74% 54.28% 15.58% 78% 63%
Sortino 35.65% 26.48% 63% 74% 51.49% 22.89% 78% 72%
Ulcer Performance Index 37.76% 31.27% 59% 76% 50.25% 15.31% 80% 65%
CAGR/(UI^x) 36.75% 27.23% 59% 76% 49.91% 15.92% 81% 67%
Alpha 41.04% 27.64% 69% 76% 49.31% 11.57% 63% 69%
Calmar 35.95% 29.16% 65% 78% 48.29% 17.16% 78% 67%
GSD 31.40% 24.70% 67% 76% 47.21% 14.45% 57% 72%
Downside Deviation 30.46% 24.76% 67% 72% 46.93% 15.96% 57% 70%
Jensen 37.30% 27.10% 70% 74% 46.41% 12.74% 59% 69%
Upside Potential Ratio 30.59% 26.29% 67% 76% 45.91% 25.32% 74% 70%
CAGR/(GSD^x) 35.68% 26.65% 59% 72% 45.83% 20.91% 74% 72%
CAGR/(SF^x) 36.05% 26.34% 59% 78% 45.70% 14.84% 81% 65%
Upside Potential 27.89% 21.35% 65% 78% 45.65% 16.40% 57% 69%
Normalized Trough Count 35.04% 30.43% 69% 83% 45.54% 15.68% 57% 72%
Beta 35.00% 25.13% 70% 69% 40.10% 4.57% 61% 69%
Treynor 37.86% 24.50% 69% 74% 39.14% 22.01% 70% 72%
GSD Ratio 37.25% 29.08% 61% 80% 37.77% 29.13% 72% 80%
Ulcer Index 33.74% 27.33% 76% 81% 37.72% 26.70% 63% 76%
Sleep Ratio 31.27% 25.88% 69% 80% 36.99% 24.05% 61% 76%
Sharpe/(GSD^x) 36.16% 25.94% 61% 74% 35.29% 30.44% 74% 83%
CAGR/(UPI^x) 32.24% 23.69% 70% 80% 35.24% 29.58% 63% 78%
Correlation 43.65% 31.38% 70% 78% 32.75% 13.96% 59% 72%
Treynor/(Beta^x) 40.48% 26.46% 65% 78% 32.19% 24.82% 67% 74%
Treynor/(GSD^x) 36.83% 23.76% 61% 74% 30.09% 25.01% 61% 76%

Nothing comes close to the Sharpe for the bullish period whether you use an ascending or descending sort. But for the bearish period you can also see that under ascending sort these three do quite well: Correlation, Ulcer Performance Index, and Normalized Trough Count. I'll be showing the results of these below.

Improved Blends -- Raising the Blend to a Whole New Level

Here is a table showing the results from above which I will follow with some comments and observations:

Begin Date:     1/2/1989                                                                             
End Date: 11/9/2007

Bullish: Sharpe Sharpe Sharpe Sharpe
Desc Desc Desc Desc
Bearish: Sharpe/GSD Correlation Normalized Trough Count Ulcer Performance Index Russell 2000
Desc Asc Asc Asc
CAGR 45.59% 46.10% 46.19% 46.07% 9.19%
GSD 18.44 23.25 18.4 24.44 18.18
Sharpe 2.03 1.7 2.05 1.63 0.34
Ulcer Index 5.10% 7.60% 5.16% 7.87% 13.85%

Bullish:
Reb Win % 78% 78% 78% 78%
Screen Win % 68% 68% 68% 68%
Bearish:
Reb Win % 83% 78% 83% 76%
Screen Win % 71% 70% 76% 73%

Yrs >= Index 89.47% 89.47% 94.74% 94.74%
Yrs >= 0 94.74% 94.74% 94.74% 100.00% 68.42%
Drawdown -23.92% -38.46% -24.91% -38.21% -46.04%
Bullish 54.28% 54.28% 54.28% 54.28% 9.52%
Bearish 30.44% 31.38% 30.43% 31.27% -3.45%

Years ROI ROI ROI ROI ROI
1989
18.08% 23.50% 17.39% 18.73% 14.22%
1990 -0.02% -0.32% -2.42% 10.59% -21.45%
1991 90.80% 117.82% 99.15% 111.33% 43.68%
1992 67.15% 39.42% 46.01% 41.50% 16.36%
1993 45.87% 66.41% 33.70% 56.17% 17.00%
1994 28.07% 45.14% 24.54% 52.97% -3.18%
1995 91.62% 74.60% 90.28% 72.06% 26.21%
1996 39.90% 42.84% 37.09% 37.37% 14.76%
1997 41.36% 47.19% 48.15% 61.50% 20.52%
1998 36.26% 11.03% 37.70% 14.51% -3.45%
1999 144.16% 151.31% 142.16% 135.41% 19.62%
2000 77.02% 51.61% 87.86% 73.27% -4.20%
2001 37.17% 57.99% 49.53% 30.48% 1.03%
2002 35.77% 27.87% 42.84% 7.62% -21.58%
2003 36.15% 29.30% 32.95% 61.92% 45.37%
2004 43.54% 53.87% 46.59% 43.08% 17.00%
2005 56.08% 66.58% 65.94% 57.85% 3.32%
2006 15.58% 3.42% 19.63% 5.91% 17.00%
2007 13.03% 43.13% 14.69% 57.44% -2.75%

Observations and Conclusions

I personally see this as an excellent way to apply all the research at our disposal for blend selection. This increases Sharpe to levels above two while also adding 13 to 15 points of CAGR. That is very impressive IMO. Last fall I didn't think it would be possible by any means to get the Sharpe above 2.0.

I would choose the Sharpe for the bullish periods and the Sharpe/GSD or Normalized Trough Count for the bearish side. Both of these produce a low Ulcer Index which I consider important.

What would my forward expectation be for a blend like this -- I'd be very happy with a CAGR in to the 20-25 range, but not the least bit surprised if it resulted in a 25-35 CAGR over the next twenty years. There is absolutely no question that this method beats out a toss of a coin, or just throwing darts, as the means to select the screens one uses to build a blend. Could it be just a statistical fluke? Sure, but to no more degree than anything and everything else done on this board.

Now for those who just love to have something to torpedo -- have at it! I challenge anyone to come up with a method for building a blend that surpasses what is being shown here. If a challenge can get a passenger rocket built, it certainly ought to be able to get a blend built that is worthy of this board and its long history. :)
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211386 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 3:11 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Zee,

A couple of basic questions about the method and data...

When you are backtesting a blend based on "Sharpe ASC", this means that each year, you choose the five screens (ranks 1-4) which have, up to that point in time, the lowest historical Sharpe ratios among all of our VL screens. Is that correct? If so, then based on the "All Measures Compared" table, choosing the five screens with the worst Sharpe ratios each year would have produced a bullish CAGR of 34.83% and a bearish CAGR of 25.83%. Is that right?

How many VL screens are in the universe that you're drawing from -- is it all VL screens that are in our current rankings?

Thanks so much,

Todd

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211387 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 3:52 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 5
Todd asked:
When you are backtesting a blend based on "Sharpe ASC", this means that each year, you choose the five screens (ranks 1-4) which have, up to that point in time, the lowest historical Sharpe ratios among all of our VL screens. Is that correct?

Yes, that's right. Sharpe ASC is a good example of a measure that I consider useless. It ranks in position #17 out of 24 for both the bullish and bearish. That means it produces the same rank position in both periods. In contrast, Sharpe DESC ranks #1 in the bullish period, and #17 in the bearish period. That is a diff of 16. There are only a few measures that show such a contrast between the two. Correlation is another measure with a huge contrast between the two -- -21 and -20. Even greater than Sharpe. Here, why don't I share the whole table I built of these values:
                         ASC   ASC   DESC   DESC                  
Screen Bull Bear Bull. Bear. Bull.Diff Bear.Diff
Correlation 1 1 22 21 -21 -20
Ulcer Performance Index 5 2 3 18 2 -16
Normalized Trough Count 15 3 14 16 1 -13
Calmar 12 4 6 12 6 -8
GSD Ratio 7 5 17 3 -10 2
Alpha 2 6 5 23 -3 -17
Ulcer Index 18 7 18 4 0 3
CAGR/(UI^x) 9 8 4 15 5 -7
Jensen 6 9 9 22 -3 -13
CAGR/(GSD^x) 13 10 11 11 2 -1
Sortino 14 11 2 9 12 2
Treynor/(Beta^x) 3 12 23 7 -20 5
CAGR/(SF^x) 11 13 12 19 -1 -6
Upside Potential Ratio 22 14 10 5 12 9
Sharpe/(GSD^x) 10 15 20 1 -10 14
Sleep Ratio 21 16 19 8 2 8
Sharpe 17 17 1 17 16 0
Beta 16 18 15 24 1 -6
Downside Deviation 23 19 8 14 15 5
GSD 20 20 7 20 13 0
Treynor 4 21 16 10 -12 11
Treynor/(GSD^x) 8 22 24 6 -16 16
CAGR/(UPI^x) 19 23 21 2 -2 21
Upside Potential 24 24 13 13 11 11

Anything over sixteen I consider a great measure, because the different sort completely and totally changes its rank position. Screens with a very low correlation to the S&P 500 do great for both the bullish and bearish periods, while highly correlated screens do horrible in both. I consider this a reflection of the fact that correlation as a measure is more highly robust than many other measures. The UPI is similar in that for the bear period the diff between them is 16. UPI is good in ascending sort for the bearish period, but horrible in descending sort. The greater the contrast, the more robust.

If so, then based on the "All Measures Compared" table, choosing the five screens with the worst Sharpe ratios each year would have produced a bullish CAGR of 34.83% and a bearish CAGR of 25.83%. Is that right?

Yes, that is right.

How many VL screens are in the universe that you're drawing from -- is it all VL screens that are in our current rankings?

I decided to use the all screens not marked as duds in the VL Pre-Post Discovery thread. My primary reason for this was simply to save time, not that I thought it would necessarily improve results.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211388 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 4:33 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Zee,

Thanks for your helpful answers to my previous questions!

Just a couple more things I was curious about...

What exactly do the Bull Win % and Bear Win % mean? A couple of cases seemed odd with these values... Just one example, "Beta DESC" had a Bullish CAGR of 40.10% and a Bearish CAGR of 4.57%. However, its Bear Win % was higher than its Bull Win % (69% vs. 61%, respectively). That was confusing to me. Also, what does Reb Win % mean, and how does it differ from Screen Win % ?

I used datahelper.com and Google to try to find a definition of "Normalized Trough Count", but I couldn't seem to find one. What does the term mean, and how is it calculated?

I apologize, I'm sure you've answered this before, but how did you choose blends for the very earliest years of the backtest (1989, 1990, 1991, etc.), given that before 1989, many of the screens don't have any historical data to examine?

Best,

Todd

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211389 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 5:40 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 9
Todd - first, thanks for the questions. They've caused me to find an error which is very helpful.

Todd asked:
What exactly do the Bull Win % and Bear Win % mean?

It means the percent of the screens selected that resulted in a winning result. Keep in mind, that my indicator produces more trades than an EMA crossover, so there are many more opportunities to fail. Anything over 79% is quite impressive.

A couple of cases seemed odd with these values... Just one example, "Beta DESC" had a Bullish CAGR of 40.10% and a Bearish CAGR of 4.57%. However, its Bear Win % was higher than its Bull Win % (69% vs. 61%, respectively). That was confusing to me.

Just because the one had a higher CAGR means nothing about its success rate. These values at this level are more noise than anything else.

Also, what does Reb Win % mean, and how does it differ from Screen Win % ?

The Reb Win % refers to the number of rebalance points that resulted in a win for the combined five screens held. So for example, on a particular date the signal went bullish and five new screens were selected. Of these five three returned a positive result and two failed, but the latter two failed more than the better three, resulting in the combined return being negative. This would count again the Reb Win % since it did not come out with a positive return. The Screen Win % refers to the individual screens win rate. All of the screens held are added up and the number with a positive return are used to determine the percent positive.

I used datahelper.com and Google to try to find a definition of "Normalized Trough Count", but I couldn't seem to find one. What does the term mean, and how is it calculated?

I used datahelper searching for ALL words using trough and count and found one of my explanations here:
http://boards.fool.com/Message.asp?mid=25278360

Here is one case where I employed it and mentioned its usefulness at the bottom of the post. Actually, I've run many a test using this measure. It is one of my own measures, not something I have ever read about elsewhere.
http://boards.fool.com/Message.asp?mid=26231522

I apologize, I'm sure you've answered this before, but how did you choose blends for the very earliest years of the backtest (1989, 1990, 1991, etc.), given that before 1989, many of the screens don't have any historical data to examine?

Good question -- this question is what made me realize my error. I reran all screen tests before running these tests on measures. I did it by grouping them into their Olympic medal standard. Unfortunately, I made the mistake of having these tests start in 1989 rather than 1986 as they should have. That is why the returns are so low for 1989 and 1990. You would need at least two years of data to have enough for the sorting measure to work. If I limit the test to just 1991 to present the results are all above 50% CAGR and the Sharpe is at 2.24 for Normalized Trough Count (NTC) and 2.18 for Sharpe/GSD. It is interesting that my home-grown NTC beats out everything else as to predicting the best screens to hold going forward in time. I actually like it because I think it is an excellent pain metric that accurately reflects the typical investor's psyche.

Anyway, I'm rerunning all the tests and will share the results with the extended data once they are complete. Again, thanks for the question that led me to track down this foolish mistake.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211390 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 9:18 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 19
Well, that took some time, but here are the results. Needless to say, fixing that mistake resulted in quite an improvement.

Begin Date:     1/2/1989                                                                             
End Date: 11/9/2007

Bullish: Sharpe Sharpe Sharpe Sharpe
Desc Desc Desc Desc
Bearish: Sharpe/GSD Correlation Normalized Trough Count Ulcer Performance Index Russell 2000
Desc Asc Asc Asc
CAGR 48.58% 47.92% 48.85% 49.81% 9.19%
GSD 19.17 22.42 18.83 24.09 18.18
Sharpe 2.07 1.8 2.12 1.76 0.34
Ulcer Index 5.48% 7.33% 4.94% 7.23% 13.85%

Bullish:
Reb Win % 78% 78% 78% 78%
Screen Win % 68% 68% 68% 68%
Bearish:
Reb Win % 83% 78% 83% 76%
Screen Win % 71% 70% 76% 73%

Yrs >= Index 89.47% 89.47% 94.74% 94.74%
Yrs >= 0 94.74% 94.74% 100.00% 100.00% 68.42%
Drawdown -27.57% -41.06% -22.75% -34.97% -46.04%
Bullish 54.28% 54.28% 54.28% 54.28% 9.52%
Bearish 30.44% 31.38% 30.43% 31.27% -3.45%

Years ROI ROI ROI ROI ROI
1989
67.99% 68.38% 68.38% 69.21% 14.22%
1990 0.84% -4.94% 0.64% 5.74% -21.45%
1991 93.46% 121.44% 92.09% 117.26% 43.68%
1992 69.87% 43.25% 48.39% 43.38% 16.36%
1993 42.28% 56.07% 30.40% 48.68% 17.00%
1994 30.27% 45.81% 26.68% 51.10% -3.18%
1995 91.62% 77.48% 90.28% 74.66% 26.21%
1996 39.90% 37.59% 37.09% 30.25% 14.76%
1997 41.36% 47.19% 48.15% 56.14% 20.52%
1998 36.26% 3.74% 37.70% 18.91% -3.45%
1999 144.16% 151.31% 142.16% 135.41% 19.62%
2000 77.02% 51.61% 87.86% 73.27% -4.20%
2001 37.17% 57.99% 49.53% 37.45% 1.03%
2002 35.77% 27.87% 42.84% 15.71% -21.58%
2003 36.15% 29.30% 32.95% 64.13% 45.37%
2004 43.54% 53.87% 46.94% 39.87% 17.00%
2005 56.08% 63.13% 63.10% 62.68% 3.32%
2006 15.58% 16.54% 18.13% 20.49% 17.00%
2007 13.03% 34.59% 13.14% 40.64% -2.75%


Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211392 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 11:03 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Another bit of information on the blend -- the Normalized Trough Count blend to be exact. Over the full history, 1989 to 7/19/2008, here are the # of times that the various olympic screens were selected, and the average gain.

Olympic Level    Begin Year  Avg ROI   #
Silver & Bronze 1997 3.81% 8
Bronze 1986 3.84% 81
Silver 1986 8.40% 71
Gold 1986 9.17% 380

I was pleasantly surprised to see that the measures select the Gold screens so often, even though they are a small representation of the whole potential screen list.

Print the post Back To Top
Author: KBGlenn Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211393 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 11:03 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Zee,

How are you calculating this? I don't come remotely close to 25%/-8% using Pinnacle AM, ND and NY High-Lows data and ^IXIC from Yahoo. If you don't want to share a trade secret how about something slightly different that I can attempt to independently confirm? I should be able to do this (and optimize) using my software no problem.

Tks KG

Print the post Back To Top
Author: klouche Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211396 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 11:19 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
So I thought I'd pull out one of my simplest methods for determining bullish and bearish periods. It simply takes the new highs and new lows of the index and subtracts the one from the other and applies a weighted moving average to the result to remove excessive whipsaw. When it is above, say zero, you are bullish, and when below, you are bearish. I've used values I've tuned a bit, but anyone is welcome to get the data from Pinnacle and go through the same process

Are you using the NASDAQ count of new highs/lows for bullish/bearish? If so, how does it (your Blending Method) work using the S&P?

KL

Print the post Back To Top
Author: mungofitch Big gold star, 5000 posts Top Favorite Fools Top Recommended Fools Feste Award Winner! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211399 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 4:11 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 20
Like KBGlenn, I didn't understand your high/low signal construction at first.
But, incomprehension can be good too. This spurred me to try out
several different ways to do this, and they all work to some extent.

All figures using Nasdaq composite 1978-present, both for the highs
and lows and for the trading on those highs and lows.
Signal at close, assuming trading the index that moment at close, no friction.
i.e., not very realistic, but probably OK for signal development if
the number of signals per year is small enough.


EMA of daily high-low number > 0
Bull: 20.8%
Bear: -6.6%
Signals/year 9.62
Bullish 66% of the time


Do two EMA's of the daily high-low number, a longer one and a shorter one.
shorter EMA > longer EMA
Bull: 26.6%
Bear: -5.6%
Signals/year 18.29
Bullish 55% of the time


MACD of high-low number
(take EMA1, EMA2, difference of those, EMA of difference, and check diff>EMA(diff))
Bull: 22.4%
Bear: -8.4%
Signals/year 8.83
Bullish 66% of the time


Trickier one:
Build a high/low line. Start at 0 on day 1. Value on day 2 =
value on day 1 + highs on day 2 - lows on day 2.
Then, create an SMA on that line.
Then, bullish if the high/low line is above SMA(high/low line)
Bull: 20.4%
Bear: -6.0%
Signals/year 8.7
Bullish 67% of the time


And, just for fun, YATS on all the above: take a vote on the sub-signals.
Simply sum the four signals above. If any are bullish be bullish, else bearish.
Bull: 21.3%
Bear: -17.6%
Signals/year 10.0
Bullish 76% of the time

Alternative YATS: bullish if 1 or more of the sub-signals are bullish.
Bull: 22.0%
Bear: -9.7%
Signals/year 9.36
Bullish 68% of the time


And lastly, as above, but with whipsaw reduction to reduce signal count.
If 3 or more of the 4 are bullish today, go bullish.
If 0 are bullish today, go bearish.
Otherwise, stay with previous day's state.
Bull: 22.8%
Bear: -10.5%
Signals/year 6.41
Bullish 67% of the time

Just for interest's sake, here are the last few signals for the last version.
Bullish at close 2006-08-18, Nasdaq at 2163.95
Bearish at close 2007-03-02, 2368
Bullish at close 2007-03-22, 2451.74 (prior bear signal was wrong)
Whipsaw: one day of bearishness triggered at close 2007-06-27, loss .12%
Bearish at close 2007-07-20, 2687.6
Bullish at close 2007-09-04, 2630.24
Bearish at close 2007-10-19, 2725.16
Index now is 2285. Ignoring the one-day whipsaw, that's 5 out of 6 good.
If you could have traded the index with this long & short without friction,
you'd have been up around 46% instead of 5.6% since August 2006.

Here's a little table of how it did by year. This shows the Nasdaq
index return for the year, what fraction of the time the signal was
bullish, the Nasdaq return during the bullish fraction of the year (not annualized),
and the Nasdaq return during the bear periods of the year (not annualized).
Biggest problem is the very poor market performance during bull
periods in 2001 and 2002. The rise in the market during the bearish
period in 1999 is also interesting historically, since market breadth
dropped long before the [megacap] market peaked. If you weren't
in megacaps at that time, it probably was a bear market for you.

Year    Nasdaq    % bull    During bull    During bear
1978 19% 79% 24% -4%
1979 26% 86% 23% 2%
1980 37% 79% 44% -5%
1981 -4% 59% 8% -11%
1982 18% 51% 36% -14%
1983 20% 73% 27% -5%
1984 -11% 34% -3% -9%
1985 32% 91% 34% -1%
1986 9% 64% 11% -2%
1987 -4% 59% 14% -16%
1988 12% 73% 9% 3%
1989 21% 77% 20% 1%
1990 -19% 30% 2% -20%
1991 58% 85% 43% 10%
1992 15% 76% 8% 6%
1993 15% 91% 6% 8%
1994 -4% 50% -5% 2%
1995 42% 93% 26% 13%
1996 21% 78% 16% 4%
1997 23% 72% 9% 13%
1998 40% 59% 41% -1%
1999 87% 69% 54% 21%
2000 -45% 40% 12% -50%
2001 -14% 60% -12% -2%
2002 -30% 54% -18% -14%
2003 45% 86% 39% 4%
2004 7% 80% 10% -2%
2005 4% 72% 0% 5%
2006 8% 75% 8% 0%
2007 8% 62% 8% -1%
2008 YTD -13% 0% N/A -13%


This isn't meant to be a trading signal by itself, but it's nice
to see the degree of consistency of the signal over time, which
is pretty good. I like that you can get quite good results with
surprisingly few signals per year, which I have come to appreciate
is a very important thing in timing signal development.

My only question for Zee: how many signals per year did your example signal produce,
and what percentage of the time was it bullish?

Jim

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: KBGlenn Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211400 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 4:39 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
OK,

What am I doing wrong? I get nowhere anything close to Zee or Jim's work? My results are worse than buy and hold.

Pinnacle Data Elements Used:
A) AM-HIGHS
B) ND-HIGHS
C) NY-HIGHS

D) AM-LOWS
E) ND-LOWS
F) NY-LOWS

Yahoo Data Used: ^IXIC 6/1/1978 - 7/1/2008

Variable Used: WMA length

Formulas Used:
G = (A+B+C) - (D+E+F)
WMA of G
Buy/Long when WMA of G > 0
Sell/Short when WMA of G < 0
Open/Close position on the next day's open

Any help would be greatly appreciated.

Tks K

Print the post Back To Top
Author: eboller Big gold star, 5000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211401 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 5:24 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
What am I doing wrong? I get nowhere anything close to Zee or Jim's work?

My impression was that they were doing the new high and lows from one index, not the sum of the three. Try with just NYSE or just nasdaq?

Eric

Print the post Back To Top
Author: buckaroobonzai Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211402 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 5:29 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Using this indicator the signal turned to bearish on 10/19/2007 and has remained so till now. Can't hardly beat that, now can you?

Hmmm, wonder if that is a record length of time in bearish mode.

Buckaro

Print the post Back To Top
Author: mungofitch Big gold star, 5000 posts Top Favorite Fools Top Recommended Fools Feste Award Winner! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211405 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 7:45 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Any help would be greatly appreciated.
...
My impression was that they were doing the new high and lows from one
index, not the sum of the three. Try with just NYSE or just nasdaq?


Yes, I used Nasdaq only for this test, per Zee's suggestion.

I didn't try a WMA, I tried EMA and SMA.


Jim

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211407 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 8:59 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 10
Wow! You guys all focus on the thing that is least important, at least in my view. It is not so much what signal you use that is important, but the fact that it is possible to use it, and raise the blending concept to a whole new level. At least that was my take away message. I certainly wasn't intending to set up a class on timing!

Kev asked:
How are you calculating this? I don't come remotely close to 25%/-8% using Pinnacle AM, ND and NY High-Lows data and ^IXIC from Yahoo. If you don't want to share a trade secret how about something slightly different that I can attempt to independently confirm? I should be able to do this (and optimize) using my software no problem.

As some have said, I only used one exchange's values, not three or a combination of any. That is the first and most important point. The second point is that now Jim has given many ways to accomplish this which really drives home my whole point. The signal used is not what is important here.

Klouche asked:
Are you using the NASDAQ count of new highs/lows for bullish/bearish? If so, how does it (your Blending Method) work using the S&P?

Yes. Do you mean how does the timing method work on the S&P? I'm not clear what you're asking.

Jim wrote:
Signal at close, assuming trading the index that moment at close, no friction. i.e., not very realistic, but probably OK for signal development if the number of signals per year is small enough.

I'd never do this even for signal development. What I share is always based on a delay of one day and entering/exiting at close prices. Of course, open would probably improve it, but this sort of indicator is not so brittle / sensitive.

This isn't meant to be a trading signal by itself, but it's nice to see the degree of consistency of the signal over time, which is pretty good. I like that you can get quite good results with surprisingly few signals per year, which I have come to appreciate is a very important thing in timing signal development.

Yes, that is right. I'd never use it as a stand-alone trading signal. My purpose was to choose an indicator that was simple to make and update (maybe failed on that one, but to be fair Jim, I don't recall being able to replicate even one of your indicator's either) and captured the general trend where perfection was not needed. Remember, we are 100% invested either way. In fact, it is possible that you remain in the same screen from one signal position to the next.

As to the "few signals per year" comment -- that is most definitely my own findings as well. It is a very rare thing that I would consider a signal with a high turnover rate as being good. The lower the better in almost all cases.

My only question for Zee: how many signals per year did your example signal produce, and what percentage of the time was it bullish?

The signal I developed without a weighted moving average is bullish 72.34% of the time, and has 5.1 signals per year on average. The weighted average signal is bullish just 60% of the time, and has 5.2 signals per year on average.

I broke the results of this indicator into three baskets -- 1978 to 1989, 1990 to 1999 and 2000 to 2008. The results showed out-performance in the first and the last period, but slight under-performance in the middle period. Here is the Long and Cash results of this indicator minus the LTBH of these indexes. Positive means it won.

                                                              1/3/1978  12/31/1989          1/1/1990  12/31/1999          1/1/2000  7/15/2008     
Diff CAGR GSD Sharpe CAGR GSD Sharpe CAGR GSD Sharpe CAGR GSD Sharpe
Combined 3.98% -6.56 0.52 6.87% -4.54 0.80 -4.15% -5.13 -0.13 9.32% -10.27 0.71
^SPX S&p 500 Ruel 2.41% -6.55 0.27 5.59% -6.21 0.46 -5.11% -4.75 -0.23 6.58% -8.95 0.40
^N225 Tokyo Nikkei 5.29% -6.89 0.21 1.83% -4.74 0.48 4.25% -9.72 0.00 11.06% -7.77 0.35
^GSPC S&p 500 2.26% -6.55 0.25 5.23% -6.20 0.43 -5.12% -4.75 -0.23 6.54% -8.96 0.40
^DJI Dow Jones Industrial Average 2.24% -6.55 0.23 6.00% -6.62 0.44 -5.43% -4.77 -0.28 5.92% -8.40 0.40
^DJT Dow Transport 3.27% -8.40 0.27 8.12% -6.84 0.59 -3.83% -6.74 -0.25 5.69% -12.09 0.40
^DJU Dow Utilities 2.63% -6.04 0.20 3.76% -4.43 0.28 0.07% -3.82 -0.11 4.30% -9.84 0.43
^DJA Dow 65 Composite 2.34% -6.15 0.27 5.71% -5.97 0.47 -4.14% -4.43 -0.26 5.49% -8.22 0.49
^VLIC Value Line 7.56% -6.33 0.68 9.73% -5.50 0.91 2.23% -4.65 0.21 11.00% -8.91 0.75
^NYA Nyse Composite 2.49% -5.98 0.30 5.89% -5.80 0.51 -4.09% -4.38 -0.23 5.61% -7.93 0.51
^IXTR Nas - Nms Transportation 4.67% -7.29 0.43 8.61% -5.14 0.73 -0.87% -5.63 -0.10 6.50% -11.39 0.55
^IXFN Nas - Nms Financial 6.50% -7.25 0.56 8.85% -4.18 1.50 -0.91% -3.95 0.10 11.24% -13.96 0.55
^IXIC Nasdaq Composite 6.86% -9.16 0.57 10.36% -5.26 1.15 -4.07% -6.00 0.05 13.16% -16.10 0.48
^IXID Nas - Nms Industrial 8.22% -9.05 0.64 11.62% -6.30 1.00 -0.94% -7.25 0.17 13.69% -13.89 0.66
^IXIS Nas - Nms Insurance 1.39% -5.13 0.26 3.80% -3.85 0.38 -3.00% -4.75 -0.10 3.24% -7.38 0.43
^IXBK Nas - Nms Bank Ndx 3.09% -5.46 0.46 9.21% -3.00 1.16 -2.27% -4.52 0.08 1.35% -8.97 0.03
^IXUT Nas - Nms Telecommunications 7.34% -9.66 0.33 7.66% -3.96 0.29 -4.71% -7.67 0.01 17.65% -19.52 0.47
^OEX S&p 100 2.43% -6.88 0.23 5.30% -6.51 0.36 -5.34% -4.94 -0.20 7.03% -9.53 0.36
^XAX Amex Composite 5.65% -5.30 0.68 9.79% -6.30 0.91 0.92% -3.51 0.18 6.06% -5.65 0.79
^RUT Russell 2000 6.36% -7.12 0.65 10.01% -6.13 1.17 0.02% -4.75 0.15 8.58% -10.25 0.56
^RUA Russell 3000 2.56% -6.70 0.30 5.16% -6.39 0.52 -4.62% -4.68 -0.20 6.77% -9.12 0.44
^RUI Russell 1000 2.12% -6.76 0.26 4.42% -6.56 0.45 -5.03% -4.75 -0.22 6.65% -9.10 0.42

It does must better on some than on others. When used for the purpose I'm using it to move between various screen blends, the indicator does the job perfectly. If it wasn't so difficult to test every timing method I have available I would test them all, but I think this accomplishes the purpose and does so with elegance.

Buckaro wondered:
Hmmm, wonder if that is a record length of time in bearish mode.

It is close to the maximum for my way of doing it. The max # of days is 281 and we are now at 276. The max # of days in bullish mode is 413. The median is 42 days in bullsih mode and 29 in bearish.

I wonder if anyone at all noticed the blend work! Seriously, it seems to me that this work is pretty revolutionary on the blending side and all that gets commented on is a simple little timing indicator. A tad frustrating you might say....

What a contrast to the past. Three or four years ago the timing would be shot down and the focus and interest would be toward screens. Now, it is just the opposite.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: fred04 Three stars, 500 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211409 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/22/2008 10:17 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
I have the blender for the period to end of 2006. Is there a blender available that covers the periord to end of 2007 and includes Treynor, etc.?

Print the post Back To Top
Author: sencoman One star, 50 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211411 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 1:01 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
I can email you a version which I updated with SIPRO to 12/07, and VL 73 screens to 10/07 ; Treynor, Treynor/GSD, and Sharpe/GSD ranking.

senco

Print the post Back To Top
Author: KBGlenn Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211412 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 1:14 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
As some have said, I only used one exchange's values, not three or a combination of any. That is the first and most important point. The second point is that now Jim has given many ways to accomplish this which really drives home my whole point. The signal used is not what is important here.

I'm not focused on the timing aspect per se, other than I can't duplicate it...even remotely. With Jim coming in in the same ball park that tells me I'm doing something wrong. My first run through was with only the Nasdaq data and a WMA vs. a SMA should not make that much difference relative to the numbers I'm getting.

Print the post Back To Top
Author: Foolfromfool Two stars, 250 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211413 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 5:52 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Hi zeelotes,
Thank you for providing great information.

So, are you saying that instead of choosing one blend for the whole year, the results could be a lot better if we chose one blend during bullish period and a different blend during bearish period, within the year?

Is the blend for both kind of periods decided at the beginning of the year, or do you find it just when the switch signal occurs?

And, you are saying that the measure to use for identifying the best screens to use is Sharpe during bulling period and Sharpe/GSD during bearish period?

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211414 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 6:40 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 4
Fool asked:
So, are you saying that instead of choosing one blend for the whole year, the results could be a lot better if we chose one blend during bullish period and a different blend during bearish period, within the year?

Yes, although the "within the year" bit is far from certain. There can be long periods where you stay in the same blend, but also times when you will switch more often.

Is the blend for both kind of periods decided at the beginning of the year, or do you find it just when the switch signal occurs?

Just when the switch signal occurs.

And, you are saying that the measure to use for identifying the best screens to use is Sharpe during bulling period and Sharpe/GSD during bearish period?

Yes, that is one of the best. The others that I've posted all do well. The absolute highest Sharpe comes from Normalized Trough Count, but that is probably a bit difficult for most to calculate. Plus, the difference is so minor between the two that they can be chalked up as no more than noise.

Print the post Back To Top
Author: fred04 Three stars, 500 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211417 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 9:01 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Just when the switch signal occurs

When the switch occurs, is the blend changed the next day or wait until the monthly rebalance dsy?

Print the post Back To Top
Author: SteveAl Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211422 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 11:55 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 27
I wonder if anyone at all noticed the blend work! Seriously, it seems to me that this work is pretty revolutionary on the blending side and all that gets commented on is a simple little timing indicator. A tad frustrating you might say....

Let me try to help you, here. I may not be representative of even a small minority, but here goes.

You seem to assume that I have a regularly updated database of all (100?) VL screens from which I can calcaulate Treynor/Alpha/Calmar/Jensen and the rest of the family. Not so.

Do I have the Pinnacle data for NH-NL? No.

Does it make sense that the Treynor/(GSD^x) ratio gives the same results whether you use it ASC or DESC? No.

I guess that the short version is this: You produce another complicated method for achieving an astonishing set of results (CAGR = 54%, no problem) almost daily. It's overwhelming.

I can read and re-read the whole of the first post in this thread and it doesn't get me one inch closer to the screens I need to invest in for the next year.

You obviously put a lot of effort into your research and your early work on blends was very helpful - thankyou, but don't be surprised if more recent stuff doesn't elicit learned questions from the likes of me.

Why the focus on the timing indicator? Well I could calculate that with some effort. Then the 20% a year would make me rich by just investing in the Nasdaq. How good is that?

Keep up the good work.

Steve

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: mali6491 One star, 50 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211423 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 12:15 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
SteveAl:
I guess that the short version is this: You produce another complicated method for achieving an astonishing set of results (CAGR = 54%, no problem) almost daily. It's overwhelming.

I can read and re-read the whole of the first post in this thread and it doesn't get me one inch closer to the screens I need to invest in for the next year.

You obviously put a lot of effort into your research and your early work on blends was very helpful - thankyou, but don't be surprised if more recent stuff doesn't elicit learned questions from the likes of me.

Why the focus on the timing indicator? Well I could calculate that with some effort. Then the 20% a year would make me rich by just investing in the Nasdaq. How good is that?



My thoughts exactly. I couldnt have said it better.

Print the post Back To Top
Author: mungofitch Big gold star, 5000 posts Top Favorite Fools Top Recommended Fools Feste Award Winner! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211431 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 1:10 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 18
Wow! You guys all focus on the thing that is least important, at least
in my view. It is not so much what signal you use that is important, but
the fact that it is possible to use it, and raise the blending concept to a whole new level.


Hey, the thread is yet young! I'm just askin' stuff one step at a time.

As I understand it, the number of trading days per year would roughly
equal 12 plus the number of signals per year. Trading days from signals
would probably involve higher turnover than trading days from monthly
cycle, because you'd probably be switching screens entirely. So, the
signal frequency is pretty important from a friction point of view.

In fact, I've been looking at a similar approach.
Note 3rd paragraph here
http://boards.fool.com/Message.asp?mid=26819376

One of the things that interested me is this: as you say, a simple
timing signal is all you need, anything that works "well enough" and
doesn't have very many signals. However, strictly speaking, it
doesn't matter whether the signal distinguishes up markets from down
markets. In principle, any signal which distinguished which kinds of
screens would do well could work just as well. Obviously rising and
falling markets are the biggest discriminators, but in theory you
could slice things differently, for example if you had a signal which
told you whether large or small caps were in vogue, or whether
or not momentum was working lately, or whether growth stocks were
in the ascendency.

In fact, it's possible that part of the reason that your method works
so well is that the signal you chose (being breadth based) may be
a better distinguisher of successful screen styles than it is of
rising and falling markets. The fact that the market falls
on average during your bear signals is not really necessary.
Along those lines, it would make sense to try a wide variety of
different simple timing signals for this, even if they were all of
similar merit based on frequency and accuracy in predicting an index,
since they may vary quite widely in their ability to predict which
sets of screens are likely to do well in the next month.
Given your great success with the breadth model, this may not be
worthwhile, since you may already have hit upon the one that works best.
But, throwing a few other signals at it might be worth the time.

Another thought along the same lines: given that there is no real
need to distinguish up and down markets, this drives home the point
that the number of states does not have to be two. As an example,
it might work well to have a blend for each of "flattish bond market",
"rising bond market", and "falling bond market".

Jim

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: ilmostro Big red star, 1000 posts Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211435 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 4:30 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
Zee,

your definition of simple doesn't mesh at all with my definition of simple.

How long did it take you to develop your backtester?

I understand your frustration but hope you won't let it get the better of you.

Bryan

Print the post Back To Top
Author: wmhays Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211439 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 8:37 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 5
does this come close to your switch signal ?

http://stockcharts.com/h-sc/ui?s=$NAHL&p=D&yr=0&mn=11&dy=0&i...


thanks

wmhays

Print the post Back To Top
Author: jj307 One star, 50 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211440 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 8:51 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
I think Steve is making a very important point here. I don't doubt for a second that Zee gets the 50% percent returns he describes. However I doubt if many others do, and I know I can't. We've seen many big backtest numbers over the past (too many) years but very few people report anything like them when they describe their real money results.

But as Steve says, the really great thing is that it doesn't matter. Chasing the 100% CAGR is a wonderful intellectual challenge, like chasing the perfect wave. But if I could reliably get a measly little 20% per year after commissions, I would have more money than I could ever use, and I would start saying "Good morning Warren" when I looked in the mirror.

Maybe the really smart people here can spend a tiny bit of their time on a fun exercise of developing simpler, lower risk methods for those of us with these very modest goals. thanks, Jeff

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211442 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 10:05 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 9
wmhays asked:
does this come close to your switch signal ?

http://stockcharts.com/h-sc/ui?s=$NAHL&p=D&yr=0&mn=11&dy=0&i......


Yes, I'd say that it would do the trick just fine. Very simple... even for the whiners who desire to be spoon fed. :)

Jim raises the key point I've been working on here:
As I understand it, the number of trading days per year would roughly
equal 12 plus the number of signals per year. Trading days from signals
would probably involve higher turnover than trading days from monthly
cycle, because you'd probably be switching screens entirely. So, the
signal frequency is pretty important from a friction point of view.


I found the high level of whipsaws to be a major hurdle to me personally adopting this approach so I wanted to see what would happen if you did not allow a switch within one month. These are the dates of the trades that resulted using this approach. You'd actually make the switch the next day.

 Weighted                    
Date Signal # of Days
12/9/1988 BULL 203
6/30/1989 BEAR 186
1/2/1990 BULL 30
2/1/1990 BEAR 34
3/7/1990 BULL 30
4/6/1990 BEAR 31
5/7/1990 BULL 51
6/27/1990 BEAR 211
1/24/1991 BULL 306
11/26/1991 BEAR 162
5/6/1992 BULL 36
6/11/1992 BEAR 30
7/11/1992 BULL 44
8/24/1992 BEAR 52
10/15/1992 BULL 131
2/23/1993 BEAR 281
12/1/1993 BULL 118
3/29/1994 BEAR 65
6/2/1994 BULL 30
7/2/1994 BEAR 39
8/10/1994 BULL 56
10/5/1994 BEAR 97
1/10/1995 BULL 273
10/10/1995 BEAR 267
7/3/1996 BULL 118
10/29/1996 BEAR 30
11/28/1996 BULL 109
3/17/1997 BEAR 51
5/7/1997 BULL 174
10/28/1997 BEAR 64
12/31/1997 BULL 140
5/20/1998 BEAR 42
7/1/1998 BULL 30
7/31/1998 BEAR 94
11/2/1998 BULL 42
12/14/1998 BEAR 30
1/13/1999 BULL 28
2/10/1999 BEAR 29
3/11/1999 BULL 140
7/29/1999 BEAR 30
8/28/1999 BULL 30
9/27/1999 BEAR 32
10/29/1999 BULL 140
3/17/2000 BEAR 30
4/16/2000 BULL 30
5/16/2000 BEAR 50
7/5/2000 BULL 30
8/4/2000 BEAR 30
9/3/2000 BULL 30
10/3/2000 BEAR 101
1/12/2001 BULL 40
2/21/2001 BEAR 57
4/19/2001 BULL 61
6/19/2001 BEAR 120
10/17/2001 BULL 128
2/22/2002 BEAR 251
10/31/2002 BULL 49
12/19/2002 BEAR 91
3/20/2003 BULL 413
5/6/2004 BEAR 30
6/5/2004 BULL 33
7/8/2004 BEAR 48
8/25/2004 BULL 203
3/16/2005 BEAR 64
5/19/2005 BULL 144
10/10/2005 BEAR 30
11/9/2005 BULL 189
5/17/2006 BEAR 49
7/5/2006 BULL 30
8/4/2006 BEAR 30
9/3/2006 BULL 183
3/5/2007 BEAR 30
4/4/2007 BULL 111
7/24/2007 BEAR 42
9/4/2007 BULL 45
10/19/2007 BEAR 276
7/21/2008 BULL

Here is the table of stats comparing the methods:

           Weighted                 No Weight                 Weighted            
BULL Min 2 BULL Min 1 BULL Min 28
Max 413 Max 509 Max 413
Median 42 Median 55 Median 61
BEAR Min 2 BEAR Min 1 BEAR Min 29
Max 281 Max 272 Max 281
Median 29 Median 29 Median 45
1/3/1978 Total # 160 Total # 156 Total # 118
7/21/2008 # / Year 5.2 # / Year 5.1 # / Year 3.9

Weighted No Weight Weighted
BULL Min 2 BULL Min 1 BULL Min 28
Max 413 Max 509 Max 413
Median 40 Median 53 Median 58.5
BEAR Min 2 BEAR Min 1 BEAR Min 29
Max 281 Max 272 Max 281
Median 29 Median 28 Median 49.5
1/1/1989 Total # 108 Total # 100 Total # 76
7/21/2008 # / Year 5.5 # / Year 5.1 # / Year 3.9

So you drop from 5.5 trades per year to 3.9. I can live with that.

Now, the question is, how does this work compared to the more active switching? I ran a new test with these trade dates using the Sharpe / Sharpe/GSD as the basis of screen selection and got the following results:

                          No Sig in
One Mth
Begin Date: 1/2/1989 1/2/1989
End Date: 11/9/2007 11/9/2007

Bullish: Sharpe Sharpe
Desc Desc
Bearish: Sharpe/GSD Sharpe/GSD
Desc Desc
CAGR 48.58% 47.48%
GSD 19.17 19.61
Sharpe 2.07 1.99
Ulcer Index 5.48% 5.33%

Bullish:
Reb Win % 78% 84%
Screen Win % 68% 76%
Bearish:
Reb Win % 83% 79%
Screen Win % 71% 71%

Yrs >= Index 89.47% 89.47%
Yrs >= 0 94.74% 100.00%
Drawdown -27.57% -27.57%
Bullish 54.28% 59.23%
Bearish 30.44% 29.89%

Years ROI ROI
1989
67.99% 67.99%
1990 0.84% 4.45%
1991 93.46% 92.93%
1992 69.87% 67.97%
1993 42.28% 44.85%
1994 30.27% 25.97%
1995 91.62% 72.75%
1996 39.90% 41.37%
1997 41.36% 43.55%
1998 36.26% 30.93%
1999 144.16% 127.36%
2000 77.02% 81.13%
2001 37.17% 37.41%
2002 35.77% 35.77%
2003 36.15% 36.15%
2004 43.54% 41.12%
2005 56.08% 57.70%
2006 15.58% 12.73%
2007 13.03% 19.01%

You'll note that the return in the bullish period actually increases from a CAGR of 54% to 59%, while it drops the Ulcer Index from 5.48% to 5.33%. Granted that is not much, but my fear was that the opposite would happen.

Take home message: You can remove a whole lot of the whipsaw by simply requiring that the trade be limited to no more than one a month. Many times this means that a movement back to the other side is simply removed and the previous signal stays in force. Very nice when that happens.

Jim added:
Obviously rising and falling markets are the biggest discriminators, but in theory you could slice things differently, for example if you had a signal which told you whether large or small caps were in vogue, or whether or not momentum was working lately, or whether growth stocks were in the ascendency.

I like your thinking here. Any ideas on concrete signals to act on. Post them or send via email and I'll test it. Be sure they are very KISS lest we cause a revolt from the masses. :)

Given your great success with the breadth model, this may not be
worthwhile, since you may already have hit upon the one that works best.
But, throwing a few other signals at it might be worth the time.


I doubt very much that this is the case. I've not tried much simply due to the fact that one run of this whole method takes about 14 hours. There are probably better methods out there that we have not tried. And I'm convinced that a method that got the signals down to once a year or less would work even better.

Another thought along the same lines: given that there is no real
need to distinguish up and down markets, this drives home the point
that the number of states does not have to be two. As an example,
it might work well to have a blend for each of "flattish bond market",
"rising bond market", and "falling bond market".


I love this idea even more, but only if we are still talking few signals a year. That is the key point. Too much whipsaw and the results become not worth the hassle. I can try the signal I devloped last year related to your suggested example. Can't recall the # of signals on that.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211443 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 10:28 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 25
Steve wrote:
Let me try to help you, here. I may not be representative of even a small minority, but here goes.

First off, Steve, this is raw research. It is a work in process. I don't expect the vast majority of board members to say much about it since they don't have the tools or experience to contribute anything value-added. But there are some who can make contributions... my comments were addressed to them. The ones having the ability to assist with bringing this to a higher level.

You seem to assume that I have a regularly updated database of all (100?) VL screens from which I can calculate Treynor/Alpha/Calmar/Jensen and the rest of the family. Not so.

Someone like Eric, Robbie, Gritton, or others would possibly see the benefit in this research and add a tool online that would enable this. Hold your horses. At the raw research stage it is not yet ready for full implementation.

Do I have the Pinnacle data for NH-NL? No.

wmhays has already posted a simple way to obtain the signals. I'm sure that although you don't have a Pinnacle subscription, you do have a mouse that can click on a link. :)

Does it make sense that the Treynor/(GSD^x) ratio gives the same results whether you use it ASC or DESC? No.

It gives the same on the bearish side, but vastly different on the bullish side. This simply shows that the indicator is not good for selecting the best screens to invest in during bearish markets. For the Descending sort though, it still ranks #4 out of 24. I'd say that is pretty good. And that is why it is still an excellent sort measure for all markets.

I guess that the short version is this: You produce another complicated method for achieving an astonishing set of results (CAGR = 54%, no problem) almost daily. It's overwhelming.

I'm on summer holiday where I finally have time to put to work all my ideas. I can post it, or not. Which do you prefer?

I can read and re-read the whole of the first post in this thread and it doesn't get me one inch closer to the screens I need to invest in for the next year.

Again... don't expect the raw research stage to give you tools to work with. That comes later. Just sit back and watch the process. Give it a few months to see whether it produces something you can actually use. Edison wasn't passing out light bulbs a few days after his invention.

You obviously put a lot of effort into your research and your early work on blends was very helpful - thankyou, but don't be surprised if more recent stuff doesn't elicit learned questions from the likes of me.

I'm not the least bit surprised. I don't expect any questions from those lacking the tools and ability to raise this to the next level. My hope is that those with the experience will step up to the plate and raise questions and issues that will improve the method, or trash it.

Why the focus on the timing indicator? Well I could calculate that with some effort. Then the 20% a year would make me rich by just investing in the Nasdaq. How good is that?

The return in the CAGR period does not translate in a full period return. You understand that, right? A strategy using this indicator invested long during bullish periods and safe during bearish would result in a CAGR of 16.47% with a Sharpe of 0.94 compared to a LTBH on the Russell 2000 of 10.12% with a Sharpe of 0.29. If you invested long and short, you'd get 15.92% and a Sharpe of 0.65. So it is not 20% CAGR, although it still beats out the LTBH of the index. If that is what you want, go for it.

Jeff added:
I think Steve is making a very important point here.

.....

Maybe the really smart people here can spend a tiny bit of their time on a fun exercise of developing simpler, lower risk methods for those of us with these very modest goals.


There are tons of ideas shared here that do what you are suggesting. There are times, however, that some of us like a bit of a challenge.

If you really want it simple here is my suggestion:

Use Robbie's backtester at the end of each year to get the daily portfolio values for each of the major screens. You can start by using what I've just posted to get the Gold, Silver and Bronze screens. Then calculate the Treynor/GSD on each of them using the spreadsheet on my blog, or that others have shared. Choose the top five screens and buy ranks 1-4 from each for the whole year. Repeat this process at the end of each year. I'd guess it would take about 5-10 hours to accomplish this. But you'd only do it once a year. I wouldn't think that is too much time to invest for your future.

During the year do not visit this board. Just come to get the screen picks once a month. Don't bother reading all this mumbo-jumbo. The method above will accomplish your very modest goals with minimum hassle and frustration.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211445 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 10:55 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Zee,

Not a terribly monumental issue, but I notice that the new trading method raises bullish CAGR by 5 points, lowers bearish CAGR by half a point, but lowers overall CAGR by a little more than a full point. How can that be?

Todd

Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211447 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 11:04 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 5
Use Robbie's backtester at the end of each year to get the daily portfolio values for each of the major screens. You can start by using what I've just posted to get the Gold, Silver and Bronze screens. Then calculate the Treynor/GSD on each of them using the spreadsheet on my blog, or that others have shared. Choose the top five screens and buy ranks 1-4 from each for the whole year. Repeat this process at the end of each year. I'd guess it would take about 5-10 hours to accomplish this. But you'd only do it once a year. I wouldn't think that is too much time to invest for your future.

Hey, why are you giving away all my secrets?

:-)

Actually I ended up using Sharpe/GSD because Treynor seems to have kind of a strange behavior when ranking screens. For some reason it occasionally rewards screens that don't look very good to me.

I think that if I tried choosing screens with Sharpe for bullish periods and Sharpe/GSD for bearish periods I would end up with fairly similar blends for both periods. i.e. Some of the screens would appear in both blends, reducing total turnover.

Elan

Elan

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211450 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 11:18 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Todd asked:
Not a terribly monumental issue, but I notice that the new trading method raises bullish CAGR by 5 points, lowers bearish CAGR by half a point, but lowers overall CAGR by a little more than a full point. How can that be?

Why? Because updating these spreadsheets from one computer to the other using Remote Desktop sometimes results in me forgetting what I updated and what I didn't. Those bits were some that I obviously didn't. Got the main bits, but not the details. I'm happy though to see that someone is paying attention to the details. I sure do.

                          No Sig in
One Mth
Begin Date: 1/2/1989 1/2/1989
End Date: 11/9/2007 11/9/2007

Bullish: Sharpe Sharpe
Desc Desc
Bearish: Sharpe/GSD Sharpe/GSD
Desc Desc
CAGR 48.58% 47.48%
GSD 19.17 19.61
Sharpe 2.07 1.99
Ulcer Index 5.48% 5.33%

Bullish:
Reb Win % 78% 84%
Screen Win % 70% 76%
Bearish:
Reb Win % 83% 79%
Screen Win % 72% 71%

Yrs >= Index 89.47% 89.47%
Yrs >= 0 100.00% 100.00%
Drawdown -27.57% -27.57%
Bullish 59.79% 59.23%
Bearish 30.43% 29.89%

Years ROI ROI
1989
67.99% 67.99%
1990 0.84% 4.45%
1991 93.46% 92.93%
1992 69.87% 67.97%
1993 42.28% 44.85%
1994 30.27% 25.97%
1995 91.62% 72.75%
1996 39.90% 41.37%
1997 41.36% 43.55%
1998 36.26% 30.93%
1999 144.16% 127.36%
2000 77.02% 81.13%
2001 37.17% 37.41%
2002 35.77% 35.77%
2003 36.15% 36.15%
2004 43.54% 41.12%
2005 56.08% 57.70%
2006 15.58% 12.73%
2007 13.03% 19.01%

Elan wrote:
Actually I ended up using Sharpe/GSD because Treynor seems to have kind of a strange behavior when ranking screens. For some reason it occasionally rewards screens that don't look very good to me.

Do you mean Treynor/GSD or just Treynor? I can post all of the historical screens selected if interested.

I think that if I tried choosing screens with Sharpe for bullish periods and Sharpe/GSD for bearish periods I would end up with fairly similar blends for both periods. i.e. Some of the screens would appear in both blends, reducing total turnover.

Yes, that is very possible. It has a 50% turnover rate for this combination. Which means that there is certainly not as high a turnover rate as would be the case with many other combinations.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211451 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/23/2008 11:52 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Elan wrote:
Actually I ended up using Sharpe/GSD because Treynor seems to have kind of a strange behavior when ranking screens. For some reason it occasionally rewards screens that don't look very good to me.

Do you mean Treynor/GSD or just Treynor? I can post all of the historical screens selected if interested.


I meant Treynor/GSD. Treynor alone favors very aggressive screens.

Elan

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211452 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 12:02 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
I'm trying to find a way to post a table that has sorting at the top. The following is the table I'm trying to share using Google Docs. The problem is that I can have sorting on the site as I work on it, but once published there is no sorting.

http://spreadsheets.google.com/pub?key=pdqjYSr9aNmg68Kt4dfq7...

I also tried Zoho Sheet but without success. Any ideas?

NOTE: The table above shows the same sort of tests I've been doing but against the SIPRO universe using the NH-HL dates with no changes within one month. The ranks are 1-4 also, and the blend holds five screens.

As you can see, the best sort measures are not the same when SIPRO is employed.

Bullish: Downside Deviation descending
Bearish: Ulcer Index Ascending

Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211453 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 12:43 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 4
I can post all of the historical screens selected if interested.

Zee,

I'd be interested in seeing this!

Thanks,

Todd

Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211454 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 12:59 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Use Robbie's backtester at the end of each year to get the daily portfolio values for each of the major screens. You can start by using what I've just posted to get the Gold, Silver and Bronze screens. Then calculate the Treynor/GSD on each of them using the spreadsheet on my blog, or that others have shared. Choose the top five screens and buy ranks 1-4 from each for the whole year.

Zee,

I have forgotten -- when ranking and selecting screens on a measure like Sharpe or Treynor/GSD, what number of stocks do you generally base the ranking on? For some reason, I recall that the ranking was based on 1-10 -- is that correct, or is it 1-4 (or something else)?

Todd

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211455 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 1:45 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
Todd wrote:
I'd be interested in seeing this!

Elan's point was that he opted to use Sharpe/GSD over Treynor/GSD due to some issues he felt existed in Treynor/GSD. Both of these two measures were found to produce nearly the same results based on my various tests last November.

In thinking more on this I believe this is a mute point now. It is quite obvious to me that the better blending method per historical backtest is the use of one blend for bullish periods and a different blend for bearish periods. Treynor/GSD works great for both, but it does not come to the top of the list for either when the market is divided in this way.

Furthermore, as I look over the screens held I see a tremendous amount of cases where the same screen is held from the bullish to the bearish. That is why the turnover is just 50%. That is a whole lot lower than I expected.

I have forgotten -- when ranking and selecting screens on a measure like Sharpe or Treynor/GSD, what number of stocks do you generally base the ranking on? For some reason, I recall that the ranking was based on 1-10 -- is that correct, or is it 1-4 (or something else)?

I've tested based on 1-4 and 1-10. The conclusions came out the same. Refer to all my posts on this theme last November to see. All tables were built based on finding the results on the lower # of ranks and the higher # of ranks combined. But frankly, I never found that the one produced something significantly different than the other.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211471 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 1:35 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 10
In thinking more on this I believe this is a mute point now. It is quite obvious to me that the better blending method per historical backtest is the use of one blend for bullish periods and a different blend for bearish periods. Treynor/GSD works great for both, but it does not come to the top of the list for either when the market is divided in this way.

Furthermore, as I look over the screens held I see a tremendous amount of cases where the same screen is held from the bullish to the bearish. That is why the turnover is just 50%. That is a whole lot lower than I expected.


This flashes my curve fitting warning light.

Let's say you optimize a blend selection process without segmenting between bull and bear markets, as you did earlier. You find that a blend optimized on Sharpe/GSD works best, or a close second best to Treynor/GSD.

Then you segment the test period into bull/bear periods and allow a different measure for blend selection in each sub-period. You find that the measures selected in each sub-period are different, but they select very similar sets of screens. What happens here in effect that the number of degrees of freedom has been increased. If the differences are small they may very easily be an artifact of curve fitting. It goes without saying that as the degrees of freedom increase the opportunities for larger excursions from the "real" result increase, and the optimal result is bound to be better than the earlier result for purely statistical reasons. In other words, curve fitting.

This is essentially the same objection I had to Jamie's attempts at screen switching about eight years ago.

Elan

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: mungofitch Big gold star, 5000 posts Top Favorite Fools Top Recommended Fools Feste Award Winner! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211475 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 6:16 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
This flashes my curve fitting warning light.

I agree that this process adds a bit more room for curve fitting, but there aren't
really a lot of degrees of freedom, since this is (as I understand it)
entirely a step-forward test. At each month you're using only screen
performance data for periods prior to that date. This doesn't increase
the degree to which the blend is overfit, because it isn't an
optimization on the whole test era. (two master blends, one bear and
one bull, created with the optimizer and hindsight would definitely have that problem).
Instead, subdividing the past and doing separate bull and bear blends
is really reducing the statistical support for the blends which are
created at each trading period. They will fit their smaller training
sets better, but that doesn't matter, as the results don't use those
numbers, they use the results in the period after that training.
You would expect that to reduce their predictive power. But, in this
case, it seems there was still enough information to create useful
blends for each case, as the fits were predictive enough to give good results.

The main degrees of freedom rest only in (a) the choice and
tuning of the bull/bear signal, and (b) the sort metric used to pick
the screens at each period, which implicitly includes things like the
lookback period for the screen results, the screen depth to use, and the
number of screens to use. That doesn't really seem like a whole lot to me.
To me it seems that by far the biggest issue is that when you're looking
at "only" the results up to 1998, you're using a screen whose 1999
results were already taken into account when the screen was being tuned.

FWIW, as for the timing signal, which is the other big tuning risk,
I have duplicated a signal almost identical to Zee's, and it has
nothing but a lookback length and a single cutoff value, so it's not a
particularly overwrought timing model as these things go.

So, as always I'm on the side of the dubious, but this procses doesn't
seem particularly overfit to me. Obviously I wouldn't expect the same
results going forward because the screens themselves are mostly in their
backtest periods, but the screen selection process seems very sensible to me.

Jim

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211476 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/24/2008 8:06 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
This flashes my curve fitting warning light.

I agree that this process adds a bit more room for curve fitting, but there aren't
really a lot of degrees of freedom, since this is (as I understand it)
entirely a step-forward test. At each month you're using only screen
performance data for periods prior to that date.


What you're describing is the lack of a look-ahead bias, i.e. crystal ball, which is true.

What I'm describing is a different issue. Let's say you have two screens A and B, and A performs slightly better over the whole test period. Now you perform some random process to split the test period into alternating subperiods (simulating a timing signal). You call the two sub-series "bull" and "bear". Then you test both screen A and B in the bull and bear periods separately, and pick out the one that performed better in each period. You have created two new opportunities to outperform A alone - one where A is "bull" and B is "bear", and the other where they are reversed.

Do you think this is totally legitimate? Now what if I revealed to you that the original A and B were two instances of the same screen run with different starting dates within the month. Clearly, therefore, the difference in performance between A and B was due to chance. And the timing exercise only gave luck an opportunity to work harder.

This is of course not exactly what Zee did. It's done for illustration only, and I hope nobody wastes time telling me how the metaphor is different from reality.

Maybe Zee is onto something, but as you say, I'm dubious.

Elan

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211498 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/25/2008 8:12 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 23
I'm getting ready to wrap up this new research related to blending. Before I do, I thought I'd post what I would consider the best that I have to offer on the subject of what is the optimal way to sort screens for bearish and bullish periods. This is based on millions of data points, and hundreds of tests. The tables that follow are the result of taking all of that and finding what measure had the best return. I did this by taking the 78-88 VL test, the 89-08 VL test and the 99-08 SIPRO test. I found the ranking for each of these tests for each measure, then I averaged the ranks. I did this first for all three, and then again, just for the 89-08 VL test and the 99-08 SIPRO test. I personally put more weight on the latter two since the earlier test has so few screens to sort off from.

What these tables will answer is this:

Which sorting measure has produced the highest forward return on a blend when used to select five screens holding ranks 1-4 from each for a total hold of twenty stocks?

My conclusion from all of this testing is not so much that one particular measure most definitely stands out in the crowd, but rather that quite a few are consistent winners. Choosing nearly any of these from the top wouldn't hurt. In fact, it may also be possible to choose to blend the results that would come from the top five. Say five stocks from each measure resulting in choosing five screens for each measure so going with the top four measures would result in actually holding twenty stock positions.
    Bullish Periods            All 3  VL-SIPRO
All Combined Sort Avg Avg.
Alpha Desc 4.0 2.5
Downside Deviation Desc 5.0 5.0
Jensen Desc 5.7 5.0
Sortino Desc 7.7 6.0
Upside Potential Desc 11.0 7.0
Calmar Desc 13.7 7.5
Sharpe Desc 10.3 8.0
Ulcer Performance Index Desc 13.0 8.0
GSD Desc 7.0 8.5
Normalized Trough Count Desc 7.0 9.0
CAGR/(UI^x) Desc 15.3 11.5
Treynor Desc 11.0 12.0
Upside Potential Ratio Desc 10.3 12.5
Treynor Asc 23.0 14.5
CAGR/(SF^x) Desc 18.0 15.5
Treynor/(GSD^x) Asc 26.7 16.5
Treynor/(Beta^x) Desc 15.7 17.0
CAGR/(GSD^x) Desc 17.3 18.5
Beta Asc 19.7 21.0
GSD Ratio Desc 28.3 21.5
Correlation Asc 18.7 22.0
Ulcer Performance Index Asc 29.7 24.0
Sleep Ratio Desc 17.3 25.0
Beta Desc 21.7 25.5
Sharpe/(GSD^x) Desc 23.3 26.0
Sortino Asc 29.0 26.0
GSD Ratio Asc 26.3 29.5
Calmar Asc 32.3 30.0
Treynor/(GSD^x) Desc 23.3 30.0
Alpha Asc 35.3 31.0
CAGR/(UI^x) Asc 36.0 32.5
CAGR/(UPI^x) Desc 31.3 32.5
Ulcer Index Desc 22.0 32.5
Sharpe/(GSD^x) Asc 38.0 33.0
CAGR/(SF^x) Asc 35.3 33.5
Sharpe Asc 34.3 34.0
Treynor/(Beta^x) Asc 38.0 34.5
CAGR/(GSD^x) Asc 39.0 35.5
Sleep Ratio Asc 33.3 36.5
CAGR/(UPI^x) Asc 31.7 37.0
Jensen Asc 37.3 37.0
Correlation Desc 36.7 39.0
Ulcer Index Asc 36.3 40.5
Upside Potential Ratio Asc 35.3 42.0
Normalized Trough Count Asc 39.7 43.0
Downside Deviation Asc 40.3 43.5
GSD Asc 41.3 46.5
Upside Potential Asc 41.3 47.0

So this tells us that GSD Ratio with a descending sort is the absolute best for the bearish periods. Ulcer Index comes in second, and the Treynor/GSD is fourth.
    Bearish Periods            All 3  VL-SIPRO
All Combined Sort Avg
Avg.
GSD Ratio Desc 15.0 5.0
Ulcer Index Asc 8.7 5.0
Sleep Ratio Asc 13.7 8.5
Treynor/(GSD^x) Desc 21.0 9.0
Correlation Asc 18.3 10.5
CAGR/(UPI^x) Asc 18.3 13.5
Ulcer Performance Index Asc 12.7 13.5
Sharpe/(GSD^x) Desc 24.3 14.5
CAGR/(GSD^x) Asc 13.7 15.5
Calmar Asc 15.3 16.0
Jensen Asc 17.3 16.0
Treynor/(Beta^x) Desc 23.0 16.5
Treynor/(GSD^x) Asc 16.7 16.5
Alpha Asc 18.3 17.5
CAGR/(UI^x) Asc 12.0 17.5
Sortino Asc 18.0 18.5
Treynor Asc 19.3 19.0
GSD Ratio Asc 18.3 20.0
CAGR/(SF^x) Asc 14.0 20.5
Upside Potential Ratio Asc 17.3 21.5
Upside Potential Ratio Desc 30.3 21.5
Sharpe/(GSD^x) Asc 19.0 22.0
Ulcer Performance Index Desc 25.3 22.5
Treynor Desc 28.0 23.0
CAGR/(UI^x) Desc 25.7 23.5
CAGR/(UPI^x) Desc 17.0 23.5
Treynor/(Beta^x) Asc 24.0 23.5
CAGR/(SF^x) Desc 30.3 24.0
Normalized Trough Count Asc 17.3 24.5
Calmar Desc 29.0 25.0
Sharpe Asc 23.0 26.0
Correlation Desc 28.7 26.5
Downside Deviation Asc 25.7 27.0
Sortino Desc 33.0 30.0
Ulcer Index Desc 25.0 31.5
GSD Asc 24.3 33.0
CAGR/(GSD^x) Desc 35.7 34.0
Sleep Ratio Desc 25.3 34.0
Beta Asc 25.7 36.0
Sharpe Desc 38.7 37.0
Upside Potential Asc 27.3 38.0
Jensen Desc 39.3 38.5
Alpha Desc 42.7 40.5
Downside Deviation Desc 36.0 40.5
Normalized Trough Count Desc 37.3 41.5
GSD Desc 39.0 42.5
Upside Potential Desc 37.7 43.5
Beta Desc 47.0 47.5

What screens would this strategy blend for right now?

I know this is the question many are asking. Well, the top ranking position for bearish periods is GSD Ratio. This is one of my home-grown measures.

   GSD Ratio
1 YLDEARNYEAR
2 YIELD4
3 ROIC
4 Overlap_4/13
5 RSWEPS

The second best measure is the Ulcer Index with the Sleep Ratio right behind it. These two are basically one and the same. Here is what you'd get with these.

   Ulcer Index
1 SomeMoSafe
2 PIH_MCP
3 PIH_CSO_safe
4 LLTD
5 HIGHCASHFLOW


The next one down in the list is the one that came out on top for bearish and bullish periods -- that is, for not choosing between periods but invested the same all the time. As you can see in this list, it pays to apply some overlap removal method to blends since many of these will choose the same stocks.

   Treynor/GSD
1 PIH_MCP
2 PIH_CSO_simple
3 SomeMoSafe
4 PIH_CSO_safe
5 LPCF

The other thing to keep in mind is that much of my later testing was done on the Olympic standard screens of Gold, Silver, and Bronze. Those tests would result in some of these screens being removed.

Anyway, I consider this to have arrived at a culmination to this research project. If you have any further questions don't hesitate to ask.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211501 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/25/2008 10:56 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
For the curious, here are the best blending measures for 2008, from 1/1/2008 to 7/21/2008 to be exact. Remember that this is building a blend of five screens holding ranks 1-4 from each screen for a total hold of twenty stock positions. So if you had chosen to build your screen by one of these measures you'd be doing pretty good so far this year.

You'll note from above that if you sorted so that the best was at the top of the two periods Ulcer Index or Sleep Ratio would be vying for the top position, and they are the ones that are best for 2008.

        Screen              Sort       ROI    CAGR    GSD   Sharpe  Ulcer Index  Blend Win %  Screen Win %
Sleep Ratio Ascending 19.18% 37.99% 25.63 1.36 2.57% 100% 100%
Ulcer Index Ascending 19.18% 37.99% 25.63 1.36 2.57% 100% 100%
CAGR/(UPI^x) Ascending 19.18% 37.99% 25.63 1.36 2.57% 100% 100%
CAGR/(UI^x) Descending 18.01% 35.51% 26.34 1.29 2.80% 100% 100%
Ulcer Performance Index Descending 18.01% 35.51% 26.34 1.29 2.80% 100% 100%
Sharpe/(GSD^x) Descending 15.99% 31.28% 25.57 1.17 3.77% 100% 80%
Treynor/(Beta^x) Descending 13.73% 26.64% 21.25 1.17 2.76% 100% 80%
Treynor/(GSD^x) Descending 13.73% 26.64% 21.25 1.17 2.76% 100% 80%
CAGR/(SF^x) Descending 13.64% 26.45% 23.98 1.06 3.71% 100% 80%
Sortino Descending 13.64% 26.45% 23.98 1.06 3.71% 100% 80%
Calmar Descending 13.64% 26.45% 23.98 1.06 3.71% 100% 80%
Upside Potential Ratio Descending 13.64% 26.45% 23.98 1.06 3.71% 100% 80%
CAGR/(GSD^x) Descending 13.14% 25.43% 24.23 1.03 3.28% 100% 80%
Sharpe Descending 13.14% 25.43% 24.23 1.03 3.28% 100% 80%


Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211502 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/25/2008 1:59 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
I thought I'd post what I would consider the best that I have to offer on the subject of what is the optimal way to sort screens for bearish and bullish periods.

I'm thoroughly confused now. I thought your earlier table on the subject showed Sharpe at the top for the bullish period and Sharpe/GSD for the bearish period.

Elan

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211516 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/25/2008 8:20 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Elan wrote:
I'm thoroughly confused now. I thought your earlier table on the subject showed Sharpe at the top for the bullish period and Sharpe/GSD for the bearish period.

That is what was found as the absolute best measures to sort with when only VL data from 1989 to 2008 was used. My conclusions at the bottom of this thread are based on that test combined with a test of the SIPRO universe from 1999-2008 and another VL test from 1978 to 1988. I simply used the rank function and found the best for each universe and then averaged those ranks to arrive at the final list of what was optimal.

You'll note in my post just above your last one I showed the returns for 2008 using these various measures. The Sharpe/GSD would have resulted in a gain to date of 15.99%, while the absolute best measure of Ulcer Index returned 19.18%. Even Sharpe as the sort, which it wouldn't be due to the bearish enviroment since last October per my NH-HL signal, returned 13.14%. I'd venture a guess that nearly all of the results in that table exceed nearly everyone's experience this year.

My point is this: adopting this strategy will beat any randomly chosen method, or employing something based on anecdotal evidence. Which one you choose from the top is not as important as choosing one and mechanically using it.

Still confused?

Print the post Back To Top
Author: SteveAl Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211525 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/26/2008 5:48 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
What screens would this strategy blend for right now?


Assuming I chose, say, UI for bearish periods as defined by your NH-NL indicator. Where would I find the data to calculate the current top screens at the time of a switch signal?

Steve

Print the post Back To Top
Author: stimpson Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211554 of 251809
Subject: Re: Blending at a Whole New Level Date: 7/27/2008 6:22 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
I was wondering about that also. Are the screen criteria/statistics published anywhere?

Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211656 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/1/2008 6:08 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
Zee,

I've arrived a bit late to this thread. I've been on vacation the last two weeks with limited internet access and just look at the feast I've returned to. I'm still quite a bit behind in my reading but I just had to read all 46 posts in this thread.

I saw your comments about people focusing on the less important part of the research. I'll try to comment on the important part.

Questions

1. Lets say one decided to go with this concept and use Sharpe to select the screens for the bull period and Sharpe/GSD for the bear period. When selecting screens for say the bull period, I assume you look at the Sharpe Ratio of each screen using all of it's backtest history and not just the "bull" part of the backtest history? Can you confirm that my understanding is correct.

2. I may not have fully absorbed all 46 posts so please excuse me if the answer is clearly stated in this thread, but when you were ranking each measure for selecting screens for bull and bear periods, did you alway rank them by resulting CAGR or by resulting Sharpe?

Comments

Elan raises the question of over curve fitting. I would to some extent share his concerns, but I think that this may well fall into the category of what Jim (mungofitch) calls harmless over tuning. In your first post in this thread, Sharpe came out as the best measure for bull periods and Sharpe/GSD for bear periods. In subsequent posts, I realise that some other measure may have come out on top, but I think these ones remained somewhere in the front runners. Let's assume we decide to go with these one. Now lets assume that the timing signal has absolutely no predictive power whatsoever, what then?

Well lets see now. When this useless signal tells us its a bull market, we would select our screens using the Sharpe Ratio. Well that's what most of us considered the best measure for selecting our screens for all markets up until last November so that can't be too bad.

Again, what would we do when this useless signal tells us it is a bear market? We would select our screens using Sharpe/GSD. Well the testing last November showed that this was better than Sharpe for selecting screens for all markets.

My conclusion is that this fits Jim's ideal of timing. Great if it works and harmless if it doesn't.

My main concern is that I don't like trading much (time, hassle, tax time, etc) and I use HTD extensively to minimise trading. Any timing methods go against the philosophy of HTD so I would probably change my implementation along the lines of.

* Select my bull and bear screens.
* On my scheduled buy day, check the indicator and only buy new stocks from the applicable screens, but don't use the indicator for selling stocks - i.e. just sell based on my normal HTD rules.

These tweaks would mean that I would probably not get much of the benefit of the timing except on prolonged runs, but it should also mean that I should not see an increase in my trading.

StevnFool

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211672 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/2/2008 5:25 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 5
Stevn wrote:
I've arrived a bit late to this thread. I've been on vacation the last two weeks with limited internet access and just look at the feast I've returned to. I'm still quite a bit behind in my reading but I just had to read all 46 posts in this thread.

Welcome back. Frankly, I've missed your participation in this research. I believe your contributions last November helped greatly.

1. Lets say one decided to go with this concept and use Sharpe to select the screens for the bull period and Sharpe/GSD for the bear period. When selecting screens for say the bull period, I assume you look at the Sharpe Ratio of each screen using all of it's backtest history and not just the "bull" part of the backtest history? Can you confirm that my understanding is correct.

Yes, I do not divide up the history. I wish that would be possible, but it is not. At least, not right now. Interesting idea though.

The bull signals have a lookback of six years of data. The bear signals have a lookback of three years of data. I found these to be more or less optimal. Using all the history, however, does not result in something significantly different than what I've posted. The lookback length, IOW, is not that important.

2. I may not have fully absorbed all 46 posts so please excuse me if the answer is clearly stated in this thread, but when you were ranking each measure for selecting screens for bull and bear periods, did you alway rank them by resulting CAGR or by resulting Sharpe?

I'm not sure what you mean here. The measure stated was the measure used for ranking the screens and determining which screens to select. So they would never use CAGR since that was not an option. Sharpe was used for when it was the measure selected. If you are referring to my various tables, most of the time I determined which measures were best by their Sharpe, not their CAGR.

Again, what would we do when this useless signal tells us it is a bear market? We would select our screens using Sharpe/GSD. Well the testing last November showed that this was better than Sharpe for selecting screens for all markets.

My conclusion is that this fits Jim's ideal of timing. Great if it works and harmless if it doesn't.


I'd agree with all your comments on this. But let me add a bit of data for your consideration. Here is a table of the blend's returns for each bull and bear signal in order to compare holding Sharpe/GSD for both, or Sharpe for both. I also include the Russell 2000 return for the sake of comparison. But first, let me share the summary tables of the full table.

The average return for Sharpe/GSD during the buy signals is 10.42% compared to 16.28% for Sharpe alone. During the bullish periods the Russell 2000 gained just over 6%. In contrast, when a bearish period comes the Sharpe/GSD has an average gain of 6.69% compared to the Sharpe gain of 4.95%. The index loses -1.11%.

Signal  Sharpe/GSD  Sharpe  Rus2k
BUY 10.42% 16.28% 6.04%
SELL 6.69% 4.95% -1.11%

Based on this you'd almost think the best option is to just use Sharpe all the time, or even to just use Sharpe/GSD all the time. Here is a table showing the various options with the far right column showing a blend where Sharpe is used to sort in the bullish periods and Sharpe/GSD in the bearish periods.

        Sharpe/GSD  Sharpe    Blend
CAGR % 34.60% 39.75% 45.41%
GSD 15.41 26.72 19.43
Sharpe 1.83 1.34 1.93
UI 4.97% 14.98% 5.25%
Max DD -28.66% -54.51% -26.63%

Here is the count of the ranks for each period -- full period, bull period and bear period:

                       Sharpe/GSD  Sharpe  Rus2k
All Signals Rank 1 33 36 7
Rank 2 33 30 13
Rank 3 10 10 56
Buy Signals Rank 1 15 19 4
Rank 2 17 16 5
Rank 3 6 3 29
Sell Signals Rank 1 18 17 3
Rank 2 16 14 8
Rank 3 4 7 27

And finally, the table of all returns for each period.

   Date     Signal  Sharpe/GSD  Sharpe    Rus2k   Sharpe/GSD  Sharpe  Rus2k
1/3/1989 BUY 24.93% 50.76% 13.63% 2 1 3
6/30/1989 SELL 12.37% 18.60% 1.12% 2 1 3
1/2/1990 BUY -4.64% -3.57% -8.28% 2 1 3
2/1/1990 SELL 7.08% 9.74% 4.97% 2 1 3
3/7/1990 BUY -0.73% 3.78% 1.25% 3 1 2
4/6/1990 SELL 1.58% 5.80% -1.25% 2 1 3
5/7/1990 BUY -0.08% 4.40% 4.26% 3 1 2
6/27/1990 SELL -6.21% -2.92% -19.07% 2 1 3
1/24/1991 BUY 64.52% 72.63% 32.39% 2 1 3
11/26/1991 SELL 35.23% 25.55% 13.07% 1 2 3
5/6/1992 BUY -0.10% -1.66% -2.23% 1 2 3
6/11/1992 SELL 4.09% 3.37% -2.83% 1 2 3
7/13/1992 BUY -7.31% -8.64% -0.73% 2 3 1
8/24/1992 SELL 28.79% 27.81% 0.38% 1 2 3
10/15/1992 BUY 16.01% 15.49% 14.28% 1 2 3
2/23/1993 SELL 25.10% 20.78% 15.82% 1 2 3
12/1/1993 BUY 13.73% 20.72% 2.65% 2 1 3
3/29/1994 SELL 1.32% 4.39% -4.01% 2 1 3
6/2/1994 BUY 5.76% 4.94% -3.74% 1 2 3
7/5/1994 SELL 2.74% 7.38% 2.21% 2 1 3
8/10/1994 BUY 2.85% 4.89% 1.80% 2 1 3
10/5/1994 SELL 6.50% 9.06% -0.94% 2 1 3
1/10/1995 BUY 35.20% 53.97% 18.62% 2 1 3
10/10/1995 SELL 27.36% 30.18% 16.24% 2 1 3
7/3/1996 BUY 1.50% 6.91% -2.56% 2 1 3
10/29/1996 SELL 7.36% 10.03% 4.12% 2 1 3
11/29/1996 BUY 8.09% 10.96% 1.33% 2 1 3
3/17/1997 SELL 0.92% 5.62% -0.23% 2 1 3
5/7/1997 BUY 26.94% 20.99% 18.84% 1 2 3
10/28/1997 SELL 9.52% 6.00% 4.02% 1 2 3
12/31/1997 BUY 14.89% 21.08% 7.96% 2 1 3
5/20/1998 SELL 1.29% -1.57% -2.34% 1 2 3
7/1/1998 BUY -6.73% -10.12% -8.23% 1 3 2
7/31/1998 SELL 13.81% 7.36% -9.94% 1 2 3
11/2/1998 BUY 5.21% 3.73% 2.59% 1 2 3
12/14/1998 SELL -0.83% 4.05% 7.46% 3 2 1
1/13/1999 BUY 2.76% 0.22% -6.88% 1 2 3
2/10/1999 SELL 2.17% -1.68% -0.51% 1 3 2
3/11/1999 BUY 13.12% 37.84% 10.09% 2 1 3
7/29/1999 SELL -1.04% 1.53% -4.31% 2 1 3
8/30/1999 BUY -1.10% 1.00% -2.45% 2 1 3
9/27/1999 SELL 0.10% 5.36% 2.77% 3 1 2
10/29/1999 BUY 6.49% 110.62% 35.95% 3 1 2
3/17/2000 SELL 3.05% -14.68% -20.02% 1 2 3
4/17/2000 BUY 2.96% 4.98% 11.53% 3 2 1
5/16/2000 SELL -0.18% 6.64% 4.11% 3 1 2
7/5/2000 BUY 3.60% 9.61% -3.89% 2 1 3
8/4/2000 SELL 3.50% 17.85% 7.92% 3 1 2
9/5/2000 BUY 0.29% -8.11% -6.87% 1 3 2
10/3/2000 SELL 26.01% -43.94% -5.07% 1 3 2
1/12/2001 BUY 13.59% 5.85% -0.07% 1 2 3
2/21/2001 SELL 5.09% 1.45% -3.82% 1 2 3
4/19/2001 BUY 7.97% 10.63% 4.76% 2 1 3
6/19/2001 SELL 3.89% -11.29% -13.46% 1 2 3
10/17/2001 BUY 13.61% 11.57% 7.03% 1 2 3
2/22/2002 SELL 19.46% 35.99% -18.53% 2 1 3
10/31/2002 BUY 8.73% 8.10% 2.47% 1 2 3
12/19/2002 SELL -2.12% -6.26% -3.52% 1 3 2
3/20/2003 BUY 30.17% 34.66% 52.80% 3 2 1
5/6/2004 SELL 4.90% 6.48% 1.55% 2 1 3
6/7/2004 BUY 4.87% 4.49% -1.24% 1 2 3
7/8/2004 SELL 3.85% 3.23% -3.83% 1 2 3
8/25/2004 BUY 40.10% 48.37% 14.30% 2 1 3
3/16/2005 SELL -0.83% -3.10% -2.61% 1 3 2
5/19/2005 BUY 13.64% 26.03% 4.95% 2 1 3
10/10/2005 SELL -0.52% -3.52% 2.40% 2 3 1
11/9/2005 BUY 5.55% 10.29% 10.61% 3 2 1
5/17/2006 SELL 0.28% -0.60% -2.39% 1 2 3
7/5/2006 BUY 0.90% -2.13% -4.04% 1 2 3
8/4/2006 SELL 2.75% 2.26% 3.28% 2 3 1
9/5/2006 BUY 10.86% 9.22% 5.93% 1 2 3
3/5/2007 SELL 5.89% 6.21% 5.05% 2 1 3
4/4/2007 BUY 7.17% 16.77% 0.45% 2 1 3
7/24/2007 SELL -3.46% -5.53% -4.11% 1 3 2
9/4/2007 BUY 10.57% 7.24% 0.29% 1 2 3
10/19/2007 SELL 3.35% 0.40% -15.80% 1 2 3
7/19/2008

My main concern is that I don't like trading much (time, hassle, tax time, etc) and I use HTD extensively to minimise trading. Any timing methods go against the philosophy of HTD so I would probably change my implementation along the lines of.

* Select my bull and bear screens.
* On my scheduled buy day, check the indicator and only buy new stocks from the applicable screens, but don't use the indicator for selling stocks - i.e. just sell based on my normal HTD rules.

These tweaks would mean that I would probably not get much of the benefit of the timing except on prolonged runs, but it should also mean that I should not see an increase in my trading.


I certainly appreciate your thoughts in this regard, but I guess I see the results above showing the worth in adopting this approach. If I could test what you suggest it would be great confirmation of whether your method would really impact the results or not. I have no way of knowing. Very hard to say.

Any other ideas on how to improve what has been done to date?

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211732 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/4/2008 6:41 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Welcome back. Frankly, I've missed your participation in this 
research.

With other things going on, I have only got up to date on this board 
now.  I was also very interested in your "Blends ride waves too!" 
thread which is quite convincing in the validity of the use of Sharpe 
for bull periods and Sharpe/GSD for bear periods.

I'm not sure what you mean here. The measure stated was the measure 
used for ranking the screens and determining which screens to select. 
So they would never use CAGR since that was not an option. Sharpe was 
used for when it was the measure selected. If you are referring to 
my various tables, most of the time I determined which measures were 
best by their Sharpe, not their CAGR.

The bit in bold answers my question.

Any other ideas on how to improve what has been done to date?

Nothing really at the moment, but I have a few more questions.

I want to get clarification on some of the numbers.  In post 211390, 
you posted results for the VL universe from 1989 - 9th Nov 2007 and you 
show the following results using Sharpe for Bullish and Sharpe/GSD for 
bearish for selecting a 20 stock 5 screen blend.

CAGR:  48.58%
GSD:   19.17
Sharpe:  2.07

I'm not quite sure what I am looking at below:

        Sharpe/GSD  Sharpe    Blend
CAGR %    34.60%    39.75%    45.41%  
GSD       15.41     26.72     19.43
Sharpe     1.83      1.34      1.93  
UI         4.97%    14.98%     5.25%
Max DD   -28.66%   -54.51%   -26.63%

What is listed here under Blend I presume is something close to what 
you reported in post 211390, but why are the results not identical.  Is it:

1.  Different time frame?
2.  Different universe?
3.  Something else?

Also, are the other columns only for the bear and bull periods 
respectively or are they for the full period?  if the latter, the 
Sharpe for Sharpe/GSD at 1.83 seems very high when compared to the 
resulting Sharpe of 1.57 for Treynor/GSD as reported in post 211385 
considering that Treynor/GSD was considered better than Sharpe/GSD
for selecting a blend for holding in both bull and bear.

I guess what I am really asking here is that 211390 seemed to be telling
me that by using this method rather than Treynor/GSD all the time, I could
increase the resulting Sharpe from 1.57 to 2.07 which is truly amazing.
The results above seem to be saying that using this method rather than
Sharpe/GSD will only increase the resulting Sharpe from 1.83 to 1.93 if
the 1.83 refers to what you would get using Sharpe/GSD all of the time.

Am I misreading something?

I need to read the "Blends ride waves too!" thread again as it may 
answer some of my questions, but it is too late tonight.

StevnFool


Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211737 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/4/2008 9:40 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 7
Stevn asked:
In post 211390, you posted results for the VL universe from 1989 - 9th Nov 2007 and you show the following results using Sharpe for Bullish and Sharpe/GSD for bearish for selecting a 20 stock 5 screen blend.

FYI: All the posted results up to 11/9/2007 were done this way to make them comparable to last November's research. I did this simply because I did not want to rerun all those tests again, but I did want an apples to apples comparison.

I'm not quite sure what I am looking at below:

Unfortunately, the confusion comes from the fact that this is up to the present, not up to 11/9/2007.

Also, are the other columns only for the bear and bull periods
respectively or are they for the full period? if the latter, the
Sharpe for Sharpe/GSD at 1.83 seems very high when compared to the
resulting Sharpe of 1.57 for Treynor/GSD as reported in post 211385
considering that Treynor/GSD was considered better than Sharpe/GSD
for selecting a blend for holding in both bull and bear.


This is partly due to the advantage gained from re-selecting screens at the indicators transition points. I have not yet been able to arrive at a good explanation of why this improves on returns so much, but it most definitely does. IOW it would be better to just use the indicator and stay with the same measure than it would be to simply rebalance the screens you use at the end of each year. All that is changed is the date when you rebalance. Of course, you rebalance a whole lot more, but I've proven already that rebalancing at fixed time periods actually hurts results, not helps them, so the issue is not adding more rebalance points, but rather, adding them at the points in time decided on by the NH-NL indicator.

The other factor is that the 2.07 Sharpe resulted from the first run where I allowed the indicator to produce a switch in less than thirty days.

I guess what I am really asking here is that 211390 seemed to be telling
me that by using this method rather than Treynor/GSD all the time, I could
increase the resulting Sharpe from 1.57 to 2.07 which is truly amazing.
The results above seem to be saying that using this method rather than
Sharpe/GSD will only increase the resulting Sharpe from 1.83 to 1.93 if
the 1.83 refers to what you would get using Sharpe/GSD all of the time.

Am I misreading something?


The post 211390 was the result of the fix where I had not included the screen history from 1986 to 1989 which resulted in poor results for 1989 and 1990 since they insufficient data. But it was actually after that that I modified the indicator so that no changes were done within one month. Here is a comparison between the two. My own opinion is that I'd still choose the one on the right despite the lower Sharpe. Partly I would do this due to the pratical implications of the reduced rebalance points, but also because I don't see the UI getting hurt by the reduced trading. Plus, when I compare the yearly returns I don't see something massively different. Your thoughts? Can Sharpe alone decide it for you?

                          No Sig in     
One Mth
Begin Date: 1/2/1989 1/2/1989
End Date: 11/9/2007 11/9/2007

Bullish: Sharpe Sharpe
Desc Desc
Bearish: Sharpe/GSD Sharpe/GSD
Desc Desc
CAGR 48.58% 47.48%
GSD 19.17 19.61
Sharpe 2.07 1.99
Ulcer Index 5.48% 5.33%

Bullish:
Reb Win % 78% 84%
Screen Win % 70% 76%
Bearish:
Reb Win % 83% 79%
Screen Win % 72% 71%

Yrs >= Index 89.47% 89.47%
Yrs >= 0 100.00% 100.00%
Drawdown -27.57% -27.57%
Bullish 59.79% 59.23%
Bearish 30.43% 29.89%

Years ROI ROI
1989 67.99% 67.99%
1990 0.84% 4.45% 1 Mth
1991 93.46% 92.93%
1992 69.87% 67.97%
1993 42.28% 44.85% 1 Mth
1994 30.27% 25.97%
1995 91.62% 72.75%
1996 39.90% 41.37% 1 Mth
1997 41.36% 43.55% 1 Mth
1998 36.26% 30.93%
1999 144.16% 127.36%
2000 77.02% 81.13% 1 Mth
2001 37.17% 37.41% 1 Mth
2002 35.77% 35.77%
2003 36.15% 36.15%
2004 43.54% 41.12%
2005 56.08% 57.70% 1 Mth
2006 15.58% 12.73%
2007 13.03% 19.01% 1 Mth

Let me share a table that compares the two peroids -- the ones on the left stop on July 19, 2008 while the others end last November. The result for 1999 for the Sharpe/GSD makes me very wary. Would you keep using the method after experiencing such a massive year of under-performance? Personally, I'm inclined to think not.

Start Date:   1/3/1989                      1/3/1989                   
End Date: 7/19/2008 11/9/2007
Sharpe/GSD Sharpe Blend Sharpe/GSD Sharpe Blend
CAGR % 34.60% 39.75% 45.41% 35.68% 41.37% 47.48%
GSD 15.41 26.72 19.43 14.95 26.65 19.61
Sharpe 1.83 1.34 1.93 1.92 1.38 1.99
UI 4.97% 14.98% 5.25% 4.81% 15.16% 5.33%
Max DD -28.66% -54.51% -26.63% -28.66% -54.51% -27.57%
1989 39.21% 75.51% 67.99%
1990 -5.44% 15.31% 4.45%
1991 83.87% 91.52% 92.93%
1992 70.88% 56.49% 67.97%
1993 43.15% 39.85% 44.85%
1994 20.78% 39.70% 25.97%
1995 51.69% 74.88% 72.75%
1996 34.55% 45.46% 41.37%
1997 46.34% 45.77% 43.55%
1998 30.78% 22.73% 30.93%
1999 13.27% 141.82% 127.36%
2000 53.22% -11.00% 81.13%
2001 48.69% 3.29% 37.41%
2002 34.59% 54.36% 35.77%
2003 31.63% 30.55% 36.15%
2004 36.24% 42.39% 41.12%
2005 35.95% 49.44% 57.70%
2006 13.48% 11.21% 12.73%
2007 15.04% 19.25% 19.01%
2008 8.24% 3.02% 8.24%

On a related note: If there was sufficient interest, I would think it feasible for a project to be done as a joint effort where a tool could be made available to facilitate finding the new blend components at each transition point. I'm not offering to do it on my own, but I am willing to assist if there are some ready to participate. In fact, I'm already working toward this goal... we'll see where it leads.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211756 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/5/2008 1:17 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3
In an earlier post, you said:

The bull signals have a lookback of six years of data. The bear signals have a lookback of three years of data. I found these to be more or less optimal. Using all the history, however, does not result in something significantly different than what I've posted. The lookback length, IOW, is not that important.

And now:

This is partly due to the advantage gained from re-selecting screens at the indicators transition points. I have not yet been able to arrive at a good explanation of why this improves on returns so much, but it most definitely does.

If using all of the data for selecting screens for you bull and bear blends, one would expect that as time goes on, the results would converge towards "the best blends" and that the timing of the re-selecting process should matter less and less. You have however found that the timing matters a great deal.

Can you confirm that the improved results of re-selecting screens at transition points were all achieved using the shorter lookbacks of 6 and 3 years respectively?

Would you expect that in this case using all history would actually harm the backtested result?

In general, I am more comfortable with using more history for screen selection (i.e. longer lookback), and when you said, "The lookback length, IOW, is not that important.", I was planning to use the maximum lookback, but what you are showing now may suggest that a shorter lookback is better. I would like your comments on this.

My own opinion is that I'd still choose the one on the right despite the lower Sharpe. Partly I would do this due to the pratical implications of the reduced rebalance points, but also because I don't see the UI getting hurt by the reduced trading. Plus, when I compare the yearly returns I don't see something massively different. Your thoughts? Can Sharpe alone decide it for you?

In general for evaluating backtests, I have tended to stick with Sharpe, but I think once you introduce timing elements or have times when you are in cash, the Sharpe Ratio may no longer be best. Something like CAGR/UI or (CAGR - rfr)/UI is probably better. I'm not sure to what extent we would call this whole blending process "timing", but I would tend to agree with you that UI probably tells you more than GSD and CAGR and UI probably tell you more than Sharpe.

Another thought I have had on this whole process. When I first suggested Sharpe/GSD last november, the idea was to still find screens or blends with good Sharpe Ratios, but to have GSD's at the lower end of the range. I think this is very prudent in bear times when a high GSD really bites. In bull times, GSD can in fact be our friend which would make it logical that Sharpe rather than Sharpe/GSD may be a better screen selection filter - i.e. the bull / bear indicator causes us to switch between less / more conservative screens.

I wonder for assessing the best strategy (I'm not talking about a screen selection filter here) would somthing like the Sortino Ratio be better than Sharpe as I believe it is not symmetrical in how looks at volatility.

Let me share a table that compares the two periods -- the ones on the left stop on July 19, 2008 while the others end last November. The result for 1999 for the Sharpe/GSD makes me very wary. Would you keep using the method after experiencing such a massive year of under-performance? Personally, I'm inclined to think not.

This has always been a bit of a worry for me. I believe Treynor/GSD is similar Sharpe/GSD in relation to 1999 and I often wondered if I would stick it if we have a 1999 repeat. The place to be then was momentum. I would think Sharpe with a shorter lookback would have found plenty of momentum to be in for 1999.

On a related note: If there was sufficient interest, I would think it feasible for a project to be done as a joint effort where a tool could be made available to facilitate finding the new blend components at each transition point. I'm not offering to do it on my own, but I am willing to assist if there are some ready to participate. In fact, I'm already working toward this goal... we'll see where it leads.

I love the idea, but I would be afraid to commit time to this as a find it hard to do all the things I want to do each day as it is. One challenge I see is that for most of us, daily backtest results are only available through the GTR1 backtester which typically has a backtest end date somewhere between 4 months and 16 months in the past. this obviously would be no use to select screens in real time at a transition point. Somewhat better is Jamie's monthly backtester, but it can be several months out of data too. It looks like there are some considerable challenges here to do this in rela time.

Thanks again for all your work in this area. I certainly think this is a Whole New Level for Blend selection.

StevnFool

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211762 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/5/2008 6:16 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
I wrote:

One challenge I see is that for most of us, daily backtest results are only available through the GTR1 backtester which typically has a backtest end date somewhere between 4 months and 16 months in the past. this obviously would be no use to select screens in real time at a transition point. Somewhat better is Jamie's monthly backtester, but it can be several months out of data too. It looks like there are some considerable challenges here to do this in rela time.

Zee,

If you have the interest in testing it, I have an idea which would make it a lot easier for anyone to select the screens for the bull and bear blend, but I don't know how much it would hurt returns.

Your idea is based on selecting screens at transition points. Lets say the indicator turned bullish today, I understand that your idea would be to select the 5 top screens sorted by Sharpe Ratio using six years of backtest data up to today for ranks 1 - 4.

I don't have backtest data up to today, but I can run any screen through the GTR1 backtester and get data up to the end of 2007. Lets assume that the last bullish indicator prior to the end of 2007 was 1st August 2007 (I just pulled this date out of the air). If I selected my 5 screens to run now by filter by Sharpe Ratio using 6 years of backtest data ending 1st Aug 2007 (the last bullish transition point for which I have data available to me), would it be reasonable to expect similar returns going forward.

If you don't have the interest or inclination to test this, I would be interested to see how much the blend changes from one bullish transition point to the next bullish transition point and similarly from one bearish transition point to the next bearish transition point.

If this method works, it may be optimal to use somewhat longer lookbacks. By this I mean that lets say in a worst case, the most recent transition point I have data for is 2 years old. A 3 year lookback ending 2 years ago may be less relevant today than a 5 year lookback ending 2 years ago.

Just a few thoughts which may make this more accessible to anyone without the need for a massive joint project, but some testing might discredit the idea.

StevnFool

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211763 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/5/2008 9:18 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Stevn asked:
Can you confirm that the improved results of re-selecting screens at transition points were all achieved using the shorter lookbacks of 6 and 3 years respectively?

Would you expect that in this case using all history would actually harm the backtested result?


Yes, I definitely used the 6/3 combination.

Here is a backtest over the full period of time using Sharpe in the bullish and Sharpe/GSD in the bearish periods:

Look-back Value     6        6       20       20
Look-back Value 3 20 3 20
CAGR 45.88% 46.22% 43.56% 43.89%
GSD 19.84 20.29 19.06 19.53
Sharpe 1.92 1.9 1.89 1.87
Ulcer Index 5.47% 6.33% 5.34% 6.53%
Average 11.48% 11.62% 10.72% 10.86%
Signal 84% 84% 82% 82%
Stock 76% 76% 74% 74%
Signal 79% 71% 79% 71%
Stock 71% 71% 71% 71%
Yrs >= Index 90.00% 90.00% 95.00% 95.00%
Yrs >= 0 100.00% 100.00% 100.00% 100.00%
Drawdown -27.57% -27.57% -27.57% -27.82%
Bullish 59.23% 59.23% 54.71% 54.71%
Bearish 29.89% 30.64% 29.89% 30.64%

I see this clearly pointing to a better result for 6/3. I have a backtest I ran on all measures that consistently put these two lookbacks to the top and that is the primary basis of my selecting these two lookbacks.

In general for evaluating backtests, I have tended to stick with Sharpe, but I think once you introduce timing elements or have times when you are in cash, the Sharpe Ratio may no longer be best. Something like CAGR/UI or (CAGR - rfr)/UI is probably better. I'm not sure to what extent we would call this whole blending process "timing", but I would tend to agree with you that UI probably tells you more than GSD and CAGR and UI probably tell you more than Sharpe.

I am definitely of this mindset.

CAGR/UI produces a value that has a much greater variance that enables one to see much more easily the mound of toast. The variance in Sharpe is just way too close to be anything more than noise.

In bull times, GSD can in fact be our friend which would make it logical that Sharpe rather than Sharpe/GSD may be a better screen selection filter - i.e. the bull / bear indicator causes us to switch between less / more conservative screens.

GSD can indeed be our friend. The fact that many of the top bullish measures are GSD descending based proves that out.

I wonder for assessing the best strategy (I'm not talking about a screen selection filter here) would somthing like the Sortino Ratio be better than Sharpe as I believe it is not symmetrical in how looks at volatility.

May be the case. What is the advantage of this over just using CAGR/UI?

This has always been a bit of a worry for me. I believe Treynor/GSD is similar Sharpe/GSD in relation to 1999 and I often wondered if I would stick it if we have a 1999 repeat. The place to be then was momentum. I would think Sharpe with a shorter lookback would have found plenty of momentum to be in for 1999.

This is most definitely a very strong reason for me to say a method such as I'm proposing is important to an MI investor. You have to be realistic about your own investor psychology.

It looks like there are some considerable challenges here to do this in real time.

I'd agree and is why I'm proposing a project be done to resolve the issue. Frankly, though, I've done the research. Others who see the value in the idea need to carry the ball to make it workable on a daily basis.

I don't have backtest data up to today, but I can run any screen through the GTR1 backtester and get data up to the end of 2007. Lets assume that the last bullish indicator prior to the end of 2007 was 1st August 2007 (I just pulled this date out of the air). If I selected my 5 screens to run now by filter by Sharpe Ratio using 6 years of backtest data ending 1st Aug 2007 (the last bullish transition point for which I have data available to me), would it be reasonable to expect similar returns going forward.

If you don't have the interest or inclination to test this, I would be interested to see how much the blend changes from one bullish transition point to the next bullish transition point and similarly from one bearish transition point to the next bearish transition point.


I ran a test with a 12 month LAG to explore your idea. Of course, a thorough check this does not make, but it provides an evaluation based off of the optimal blend combination I've found of Sharpe for bullish and Sharpe/GSD for bearish:

                                      12 Mth LAG
No Sig in No Sig in
One Mth One Mth
Begin Date: 1/2/1989 1/2/1989 1/2/1989
End Date: 11/9/2007 11/9/2007 11/9/2007

Bullish: Sharpe Sharpe Sharpe
Desc Desc Desc
Bearish: Sharpe/GSD Sharpe/GSD Sharpe/GSD
Desc Desc Desc
CAGR 48.58% 47.48% 41.43%
GSD 19.17 19.61 18.19
Sharpe 2.07 1.99 1.87
Ulcer Index 5.48% 5.33% 6.28%

Bullish:
Reb Win % 78% 84% 76%
Screen Win % 70% 76% 71%
Bearish:
Reb Win % 83% 79% 68%
Screen Win % 72% 71% 69%

Yrs >= Index 89.47% 89.47% 94.74%
Yrs >= 0 100.00% 100.00% 100.00%
Drawdown -27.57% -27.57% -26.99%
Bullish 59.79% 59.23% 52.90%
Bearish 30.43% 29.89% 24.88%

Years ROI ROI ROI
1989
67.99% 67.99% 64.39%
1990 0.84% 4.45% 3.18%
1991 93.46% 92.93% 108.81%
1992 69.87% 67.97% 68.93%
1993 42.28% 44.85% 52.17%
1994 30.27% 25.97% 21.96%
1995 91.62% 72.75% 84.24%
1996 39.90% 41.37% 40.71%
1997 41.36% 43.55% 46.24%
1998 36.26% 30.93% 6.23%
1999 144.16% 127.36% 32.48%
2000 77.02% 81.13% 76.88%
2001 37.17% 37.41% 13.41%
2002 35.77% 35.77% 12.85%
2003 36.15% 36.15% 35.43%
2004 43.54% 41.12% 46.19%
2005 56.08% 57.70% 53.07%
2006 15.58% 12.73% 27.24%
2007 13.03% 19.01% 30.52%

I'd say the damage is too severe to make it workable for me, but that doesn't mean it is not usable for others. What do you think?

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211767 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 12:58 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 3

Look-back Value 6 6 20 20
Look-back Value 3 20 3 20
CAGR 45.88% 46.22% 43.56% 43.89%
GSD 19.84 20.29 19.06 19.53
Sharpe 1.92 1.9 1.89 1.87
Ulcer Index 5.47% 6.33% 5.34% 6.53%
Average 11.48% 11.62% 10.72% 10.86%
Signal 84% 84% 82% 82%
Stock 76% 76% 74% 74%
Signal 79% 71% 79% 71%
Stock 71% 71% 71% 71%
Yrs >= Index 90.00% 90.00% 95.00% 95.00%
Yrs >= 0 100.00% 100.00% 100.00% 100.00%
Drawdown -27.57% -27.57% -27.57% -27.82%
Bullish 59.23% 59.23% 54.71% 54.71%
Bearish 29.89% 30.64% 29.89% 30.64%


I see this clearly pointing to a better result for 6/3. I have a backtest I ran on all measures that consistently put these two lookbacks to the top and that is the primary basis of my selecting these two lookbacks.

I see no statistically meaningful difference.

Elan

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211769 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 1:26 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Elan wrote:
I see no statistically meaningful difference.

Yes, that is why I wrote:
I see this clearly pointing to a better result for 6/3. I have a backtest I ran on all measures that consistently put these two lookbacks to the top and that is the primary basis of my selecting these two lookbacks.

AND

The bull signals have a lookback of six years of data. The bear signals have a lookback of three years of data. I found these to be more or less optimal. Using all the history, however, does not result in something significantly different than what I've posted. The lookback length, IOW, is not that important.

So yes, there is not much statistically meaningful... though you may change your mind slightly if you saw all the data I've seen on this theme. Still, the point remains, which lenght of lookback you use is not all that important.

Print the post Back To Top
Author: elann Big gold star, 5000 posts Top Favorite Fools Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211770 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 1:44 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Elan wrote:
I see no statistically meaningful difference.

Yes, that is why I wrote:
I see this clearly pointing to a better result for 6/3.


Not to quibble, but I read the sentence above as indicating that you thought there is a meaningful difference. If you say otherwise then no problem.

Elan

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211771 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 2:13 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Elan wrote:
Not to quibble, but I read the sentence above as indicating that you thought there is a meaningful difference. If you say otherwise then no problem.

Quibble away... but I think if you read all I quoted you can see that based on the table alone that is not what I was saying. But what seems crystal clear to me, may be the opposite to others.

Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211773 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 9:06 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Yes, I definitely used the 6/3 combination.

Here is a backtest over the full period of time using Sharpe in the bullish and Sharpe/GSD in the bearish periods:


Given what I had said before, was expecting greater drop for the 20 year lookback relative to 6/3.

May be the case. What is the advantage of this over just using CAGR/UI?

This was in respone to my suggestion of using Sortino. The short answer is that I don't really know. The reason I suggested Sortino is that it is a recognized way of combining returns and volatility to evaluate returns in a risk adjusted way. Some function of CAGR and UI will also do this, I'm just not certain if dividing CAGR by UI is the right way to do it.

I ran a test with a 12 month LAG to explore your idea. Of course, a thorough check this does not make, but it provides an evaluation based off of the optimal blend combination I've found of Sharpe for bullish and Sharpe/GSD for bearish:

...
...

I'd say the damage is too severe to make it workable for me, but that doesn't mean it is not usable for others. What do you think?


If you have the means to do it in real time, I would agree with you. If you don't, then even with a 12 month lag, it is a lot better than just using Sharpe/GSD or Treynor/GSD and evaluating each yearend so in the absence of the ability to do it in real time, I would say that yes it is workable.

Could I ask you to run the 12 month lag test with 12/6 lookback instead of 6/3 to confirm if my theory is correct that a longer lookback can in some way make up for the lag?

Also, are you calculating Sharpe and GSD using an all day start like the GTR1 backtester or a single start like Jamies backtester. If a single start is sufficient, we could probably use a shorter lag.

StevnFool

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211778 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 11:27 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Stevn asked:
Could I ask you to run the 12 month lag test with 12/6 lookback instead of 6/3 to confirm if my theory is correct that a longer lookback can in some way make up for the lag?

                                        6 / 3       12 / 6
12 Mth LAG 12 Mth LAG
No Sig in No Sig in No Sig in
One Mth One Mth One Mth
Begin Date: 1/2/1989 1/2/1989 1/2/1989 1/2/1989
End Date: 11/9/2007 11/9/2007 11/9/2007 11/9/2007

Bullish: Sharpe Sharpe Sharpe Sharpe
Desc Desc Desc Desc
Bearish: Sharpe/GSD Sharpe/GSD Sharpe/GSD Sharpe/GSD
Desc Desc Desc Desc
CAGR 48.58% 47.48% 41.43% 41.24%
GSD 19.17 19.61 18.19 18.49
Sharpe 2.07 1.99 1.87 1.84
Ulcer Index 5.48% 5.33% 6.28% 6.90%
CAGR/UI 8.86 8.91 6.60 5.98

Bullish:
Reb Win % 78% 84% 76% 82%
Screen Win % 70% 76% 71% 73%
Bearish:
Reb Win % 83% 79% 68% 71%
Screen Win % 72% 71% 69% 71%

Yrs >= Index 89.47% 89.47% 94.74% 94.74%
Yrs >= 0 100.00% 100.00% 100.00% 100.00%
Drawdown -27.57% -27.57% -26.99% -30.13%
Bullish 59.79% 59.23% 52.90% 53.60%
Bearish 30.43% 29.89% 24.88% 24.77%

Years ROI ROI ROI ROI
1989
67.99% 67.99% 64.39% 64.39%
1990 0.84% 4.45% 3.18% 1.47%
1991 93.46% 92.93% 108.81% 105.67%
1992 69.87% 67.97% 68.93% 66.07%
1993 42.28% 44.85% 52.17% 43.88%
1994 30.27% 25.97% 21.96% 23.04%
1995 91.62% 72.75% 84.24% 82.98%
1996 39.90% 41.37% 40.71% 39.56%
1997 41.36% 43.55% 46.24% 46.49%
1998 36.26% 30.93% 6.23% 2.77%
1999 144.16% 127.36% 32.48% 23.14%
2000 77.02% 81.13% 76.88% 76.35%
2001 37.17% 37.41% 13.41% 36.98%
2002 35.77% 35.77% 12.85% 11.79%
2003 36.15% 36.15% 35.43% 32.98%
2004 43.54% 41.12% 46.19% 49.40%
2005 56.08% 57.70% 53.07% 43.81%
2006 15.58% 12.73% 27.24% 36.68%
2007 13.03% 19.01% 30.52% 30.40%

Also, are you calculating Sharpe and GSD using an all day start like the GTR1 backtester or a single start like Jamies backtester. If a single start is sufficient, we could probably use a shorter lag.

Single start. How much shorter a lag do you suggest?

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211779 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/6/2008 1:13 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Thanks a lot Zee,

It looks like 6/3 is probably the one to stick with even with the lag as most measures are better with 6/3 rather than 12/6.

Single start. How much shorter a lag do you suggest?

I just looked at RS26 in Jamie's Standard backtester.

It shows picks and returns for the month of July, but does not show a return for the S&P500. I think I recall Jamie saying before that the most recent months return may not be valid. If we assume that during the month of August, we have valid returns for up to the end of June, that would correspond to a lag of between 1 & 2 months.

Lets go with 2 months lag.

Can you also clarify exactly what you mean by lag in this backtest. Take for example the test with 12 months lag. If there was a bullish transition on 1 Mar 2005. Did you:

A. select the screens using data up to 1 Mar 2004

Or

B. select them using data up to the most recent bullish transition prior to 1 Mar 2004.

My intention was B, but I'm not sure if your backtester can handle it. If it is A. then you would expect the results to be worse as you are no longer selecting screens at the transition points.

StevnFool

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211794 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/7/2008 7:51 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Stevn asked:
Lets go with 2 months lag.

Can you also clarify exactly what you mean by lag in this backtest. Take for example the test with 12 months lag. If there was a bullish transition on 1 Mar 2005. Did you:

A. select the screens using data up to 1 Mar 2004

Or

B. select them using data up to the most recent bullish transition prior to 1 Mar 2004.


I'm using A -- as you supposed, I cannot test B.

I ending up testing from 1 to 4 month lag and found that it's not a very smooth progression by any means.

Lookbacks      6 / 3    6 / 3    6 / 3    6 / 3
Lag Months 1 2 3 4
CAGR 43.53% 42.08% 40.27% 43.78%
GSD 18.85 19.01 18.7 18.26
Sharpe 1.9 1.83 1.78 1.96
Ulcer Index 5.29% 5.42% 5.70% 5.30%
Average 10.46% 10.20% 9.76% 10.48%
Signal 82% 76% 76% 76%
Stock 74% 73% 71% 73%
Signal 79% 76% 74% 76%
Stock 72% 69% 70% 74%
Yrs >= Index 89.47% 89.47% 89.47% 94.74%
Yrs >= 0 100.00% 100.00% 100.00% 100.00%
Drawdown -26.42% -24.70% -27.57% -27.57%
Bullish 53.47% 52.80% 48.46% 52.43%
Bearish 28.28% 26.13% 26.95% 29.75%

Let me know if you have any further ideas for improving on this.

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211797 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/7/2008 9:18 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
It is interesting that the single start doesn't seem to hurt all that much.

12 month Lag All start gave a Sharpe of 1.87
Average Sharpe of 1 - 4 month lag with single start is 1.86
Average Sharpe of 1 - 2 month lag with single start is 1.865

This would perhaps suggest we should use single starts if the data is fresher.

I ending up testing from 1 to 4 month lag and found that it's not a very smooth progression by any means.

What follows is pure speculation. You mentioned earlier in this thread of getting an average of 5.5 signals per year. Presumably this equates to 2.75 bull signals and 2.75 bear signals per year. Expressed in a different manner this equates to 4.36 months on average between similar signals. Note that your best return was for 4 months lag where you may have on average been aligning better with a previous similar signal. If this theory is correct, a 5 month lag should not be much worse than a 4 month lag and a 6 month lag should drop off further.

If this were the case, it would support the idea of using single start data from Jamie's backtester and selecting screens using data up to the most recent previous similar signal.

StevnFool

Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211801 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/7/2008 11:13 AM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Here are a few more data points to consider:

Lookbacks      6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3
Lag Months 1 2 3 4 5 6 7 8
CAGR 43.53% 42.08% 40.27% 43.78% 42.64% 42.68% 44.24% 44.13%
GSD 18.85 19.01 18.7 18.26 18.14 18.78 18.81 18.8
Sharpe 1.9 1.83 1.78 1.96 1.94 1.88 1.95 1.94
Ulcer Index 5.29% 5.42% 5.70% 5.30% 5.45% 5.65% 5.88% 6.02%
Average 10.46% 10.20% 9.76% 10.48% 10.34% 10.40% 10.89% 10.81%
Signal 82% 76% 76% 76% 82% 82% 82% 76%
Stock 74% 73% 71% 73% 73% 72% 73% 72%
Signal 79% 76% 74% 76% 74% 74% 74% 68%
Stock 72% 69% 70% 74% 73% 75% 72% 69%
Yrs >= Index 89.47% 89.47% 89.47% 94.74% 94.74% 94.74% 94.74% 94.74%
Yrs >= 0 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00% 100.00%
Drawdown -26.42% -24.70% -27.57% -27.57% -24.70% -24.70% -27.57% -27.57%
Bullish 53.47% 52.80% 48.46% 52.43% 52.09% 52.28% 56.32% 58.48%
Bearish 28.28% 26.13% 26.95% 29.75% 29.01% 28.89% 27.27% 24.95%


Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: blm Two stars, 250 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211803 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/7/2008 12:17 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
I ending up testing from 1 to 4 month lag and found that it's not a very smooth progression by any means.

Lookbacks      6 / 3    6 / 3    6 / 3    6 / 3
Lag Months 1 2 3 4
CAGR 43.53% 42.08% 40.27% 43.78%
GSD 18.85 19.01 18.7 18.26
Sharpe 1.9 1.83 1.78 1.96
Ulcer Index 5.29% 5.42% 5.70% 5.30%
Average 10.46% 10.20% 9.76% 10.48%
Signal 82% 76% 76% 76%
Stock 74% 73% 71% 73%
Signal 79% 76% 74% 76%
Stock 72% 69% 70% 74%
Yrs >= Index 89.47% 89.47% 89.47% 94.74%
Yrs >= 0 100.00% 100.00% 100.00% 100.00%
Drawdown -26.42% -24.70% -27.57% -27.57%
Bullish 53.47% 52.80% 48.46% 52.43%
Bearish 28.28% 26.13% 26.95% 29.75%

I admit I haven't read the whole thread yet so maybe I don't understand what you mean by "it's not a very smooth progression", but to me the columns look pretty close, probably statistically indistinguishable.

Brian

Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211805 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/7/2008 12:58 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
Lookbacks      6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3    6 / 3
 Lag Months      1        2        3        4        5        6        7        8
    CAGR      43.53%   42.08%   40.27%   43.78%   42.64%   42.68%   44.24%   44.13%
    GSD        18.85    19.01     18.7    18.26    18.14    18.78    18.81     18.8
   Sharpe       1.9      1.83     1.78     1.96     1.94     1.88     1.95     1.94

Well it looks like if you measure the return by Sharpe, I was on the 
right track when I said 

If this theory is correct, a 5 month lag should not be much worse 
than a 4 month lag and a 6 month lag should drop off further.

After that, I think it is getting probably just noise.  Even if it is 
all noise.  The average Sharpe for all of the above is 1.90 which is 
not bad.

StevnFool


Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211823 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/8/2008 1:20 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 4
Zee,

I'm getting wary of single start per month sorting of screens for a 3 
year lookback.

I calculated Sharpe/GSD for 148 "Standard VL" screens using data from 
Jamie's backtester from October 2004 through September 2007 - i.e. the 
three years prior to the last bear signal.

For a few of these screens I have also calculated Sharpe/GSD using the 
GTR1 backtester with a start date of 20041018 and an end date of 
20071017 with a 21 day hold.  I know these dates do not line up exactly 
but they are not all that far apart.

In all cases it was for holding ranks 1-5 HTD 10 and a friction of 1%.  
I used the HTD because I typically use HTD and I went with 1% friction 
as it favours less trading.  The results are dramatically different.

My conclusion is that there is just too much noise in a 3 year lookback 
with 1 start per month.

These are sorted in a descending sort by Sharpe/GSD calculated from the 
all day start backtester.  Compare the ranking of say NoMo and NoMoSafe 
by Single day versus All day Start.

I know it didn't come out top in your tests, but I'm leaning towards just
selecting a bull and bear blend now using all start data up to end of 2007
and using these blends for any switches this year until the 2008 data
becomes available in the GTR1 backtester.  I havent decided yet if I 
would use all history or the 6/3 lookback for this approach.

StevnFool

	         Single	AllDay
SomeMoSafe	 0.17	0.15
NoMoSafe	 0.13	0.14
YIELD4	         0.17	0.14
SomeMoC	         0.15	0.13
HIYIELD	         0.17	0.13
SomeMo	         0.17	0.13
NoMo	         0.19	0.09
CDPD	         0.12	0.09
PIH_CSO_simple	 0.06	0.08
3PT_SCV_pst	 0.15	0.08
YLDEARNYEAR	 0.08	0.08
YldDiv	         0.16	0.07
LPE_YLD	         0.09	0.06
SomeMoJoeSafe	 0.14	0.06
YLDEARNYEAR2	 0.06	0.05
YLDYEAR	         0.08	0.03


Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: lsmr409 Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211825 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/8/2008 1:44 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 0
StevnFool,

I know you haven't finalized your approach yet, but I just had a few questions for you...

How many screens and stocks would you plan to hold at one time? Would you potentially be holding (for example) SomeMoSafe, NoMoSafe, and SomeMoC all at the same time, or would you eliminate screens with heavy overlap, and move further down the list? If you decide to use an all-history lookback, would you pick a fixed starting point for all screens (say 12/1988)?

Is there a place to find the GTR1 'recipes' for the NoMo and SomeMo screens? I'd love to take a look at those.....

Thanks very much and regards,

Todd

Print the post Back To Top
Author: StevnFool Big red star, 1000 posts Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211832 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/8/2008 4:34 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 2
How many screens and stocks would you plan to hold at one time?

Zee in his testing has tended to always hold 4 stock per screen and mostly he tested holding a blend of 5 screens. That would be 20 stocks in total. In some of his recent testing, the best results came from holding between 4 and 6 screens. He made some comment about trying to decide between 5 or 6 screens.

4 screens = 16 stocks
5 screens = 20 stocks
6 screens = 24 stocks

I looked at how many stocks to hold in a blend some time last year and concluded also that 20 - 25 was about optimum. I can't recall exactly how I came to that conclusion. The Blend I am currently running which was developed late last year has 3 stocks in each of 8 screens = 24 stocks. I'll probably stick with this number.

Would you potentially be holding (for example) SomeMoSafe, NoMoSafe, and SomeMoC all at the same time, or would you eliminate screens with heavy overlap, and move further down the list?

I would look at the overlap using Jamie's Blender and see how many months have excessive overlap. See here for example.

http://www.backtest.org/8907BL(SomeMo)15p1(SomeMoC)15p1

This is a Blend of:

SomeMo 1-5
SomeMoC 1-5

1 period with 1 stock
187 periods with 5 stocks
39 periods with 6 stocks
1 period with 7 stocks

To me this is way to much overlap - i.e. in most (187 out of 228) periods the picks are identical.

However, I currently have, NoMoSafe, SomeMoC, YLDEARNYEAR & YLDEARNYEAR2 in my blend. The overlap is not always as bad as you might initially think. I use HTD extensively which also helps with this problem.

I think I would probably only use one SomeMo variant and one NoMo variant.

If you decide to use an all-history lookback, would you pick a fixed starting point for all screens (say 12/1988)

I have used from 1989 onwards for the last couple of years, but with the GTR1 backtester, I may use less history if I want to include extra screens that do have as much history.

Is there a place to find the GTR1 'recipes' for the NoMo and SomeMo screens? I'd love to take a look at those.....

NoMOSafe 1-5 HTD 10:
http://www.backtest.org/gtr1/
Dread:lt7:cpe:gt0:cdy:gt0:ratio(cdy,cpe)tp12:elqw.v:tp60:pih.v:bn5:htd4:pih.v:bn10:pri:vprc(0,2):cpe:ratio(pri,ces.v):cdy:ratio(cdv.v,pri):Dread:linear(1,tim.v,1,sft.v)

SomeMoC 1-5 HTD 10:
http://www.backtest.org/gtr1/
tim.v:lt6:cpe:gt0:cdy:gt0:product(cpe,cdy)lt1.35:ratio(ph253.g:3,cpe)tp33.3:cdy:tp33.3:elqw.v:tp33.3:pih.v:bn5:htd6:pih.v:bn10:pri:vprc(0,2):cpe:ratio(pri,ces.v):cdy:ratio(cdv.v,pri)

By the way, the screens in my previous post are not necessarily the best of the 148 screens. They were just the ones where I had run them in the GTR1 backtester.

StevnFool

Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Print the post Back To Top
Author: Zeelotes Big red star, 1000 posts Feste Award Nominee! Old School Fool Add to my Favorite Fools Ignore this person (you won't see their posts anymore) Number: 211838 of 251809
Subject: Re: Blending at a Whole New Level Date: 8/8/2008 9:48 PM
Post New | Post Reply | Reply Later | Create Poll . Report this Post | Recommend it!
Recommendations: 1
Stevn wrote:
I'm getting wary of single start per month sorting of screens for a 3
year lookback. ....

My conclusion is that there is just too much noise in a 3 year lookback
with 1 start per month.


I can appreciate that. Remember that although I'm using a single start per month to select the stocks held, the data I'm calculating off from is always daily, not monthly. I hadn't thought about this as it relates to your using it with just the monthly granularity, but I would be inclined to agree that it could be problematic.

I would be a whole lot more comfortable with your suggested solution of using daily data from the GTR1 for the year after it is available compared to using monthly data.

I havent decided yet if I would use all history or the 6/3 lookback for this approach.

It seems to me that whichever you choose is not that crucial. The main advantage of the system is not found in this question, but rather, in using a different blend for one period compared to another.

Print the post Back To Top
UnThreaded | Threaded | Whole Thread (69) | Ignore Thread Prev Thread | Next Thread
Advertisement