Post New

ROE_Cash and YEY seem to complement each other when blended 50/50

Blend 9/97 to 20240103
CAGR 19.3
SAWR 12.34
GSD 20.0
MDD -47.3
UI 8.2
Sharpe 0.96
AT 4.1

https://gtr1.net/2013/blend.cgi?!!QlpoMTFBWSZTWSD9...

Could you live with a perpetual 12.3% withdrawal rate? If only backtests were future realities.

RAM

What are the rules for the blend; they're unclear to me from the gtr1 setup?

First two or three from one and then from ranks three or two on down from the other to a total of five? Top five from each and buy ten? Alternating from each? Something else?

Thanks

Eric Hines

What are the rules for the blend;

I set up both screens for a hold 10 each. So unless I have misunderstood the blender
I believe it is just running the 2 screens in parallel each month with 20 total equities.

I assume you could run each screen hold 10 individually, download the daily portfolio values
and add them together to verify if that is what the blender is actually doing.

RAM 

I set up both screens for a hold 10 each. So unless I have misunderstood the blender
I believe it is just running the 2 screens in parallel each month with 20 total equities.

I assume you could run each screen hold 10 individually, download the daily portfolio values
and add them together to verify if that is what the blender is actually doing.

The blender link has a rebalance every 1 period. As such, I don't think it is running the two screens independently. Each stock is rebalanced to 5% weight each period.

Craig

Aussie "The blender link has a rebalance every 1 period. As such, I don't think it is running the two screens independently."

Why not both screens have the same hold start date, trading started and stopped on the same dates and both screens had the same period.
Seems to me they would be in sync.

RAM

Why not both screens have the same hold start date, trading started and stopped on the same dates and both screens had the same period.
Seems to me they would be in sync.

Screen 0 has a rebalance every 3 periods (60 days). Screen 1 has a rebalance every 1 period (20 days) The blend rebalances at the longest period (I think), which means every 60 days, all stocks will be rebalanced to equal weight. i.e. money will be moved from the screen that has performed the best over the last 60 days to the screen than has not performed as well.

There is a wrinkle that I do not know how it is handled. Screen 0 has a minimum hold of 2 periods. So if the stock has only been held for 1 period at rebalance time, does that mean then some quantity of the stock could be sold? Not sure.

If the screens are run independently, no money moves between the screens, therefore you will get a different result than the blend.

Craig

Aussie; As you pointed out the rebalance period and minimum hold period is different for
Screen 0 and screen 1 but I still believe the blender runs the two screens independently and
Then merges the resulting daily returns together just like an individual would.
However just to verify this doesn’t significantly change the results I have rerun the blend
Setting the rebalance to 1 for both screens and the minimum hold to 1 for both screens.
The modified blend:
https://gtr1.net/2013/blend.cgi?!!QlpoMTFBWSZTWcbE...
CAGR 19.76 from 19.3
SAWR 12.39 from 12.34
GSD 21.6 from 20.0
MDD -48.5 from -47.3
UI 8.8 from 8.2
Sharpe 0.93 from 0.96
AT 3.6 from 4.1
RAM

CAGR 19.76 gets my attention. Someone shoot holes in this screen before I start doing it.

Then merges the resulting daily returns together just like an individual would.

Perhaps I am not the typical individual. I use five different screens and I don't merge the money between screens, even though some are in the same account. My thought process for not merging the money between screens is that if one screen is really not adding value (such as a data mined screen), I don't want to take money from a successful screen (one can only hope) and move the money to an unsuccessful screen. I do realize that if all the screens do add value, and are uncorrelated, then moving money between screens would be good. However, I have little confidence that all my screens add value, or are uncorrelated.

But, back to your blend. I think the backtester is doing as you say, which is merging the results together, by reducing money from the screen that has gone up and adds to the screen that has gone down.

Craig

Philly Tide: “Someone shoot holes in this screen before I start doing it.”

You certainly have reason to be skeptical , of the many thousands of screens proposed on many sites
over the last 20 years very few have performed for even a short period post discovery and fewer that
continued to perform over a several year period.

Yeild Earn Year has been one of the most tried and true screens originally a Value Line Screen modified
to a Stock Investor Pro version when VL became unavailable. It was perhaps the best overall performer
until most abandoned it when it performed poorly from 2014 thru 2018 but did very well in 2022 &23.

ROE_Cash is a much newer screen proposed by mungofitch as a SPY alternative on 5/24/2020 with a hold 40 till drop 45. I can verify personally that it has continued to work for me from 7/2020 with some minor modifications and a hold 10.

I don’t think you can shoot down the logic of either screen but I also wouldn’t put a large portion of my
Investments in either screen. That said I have consistently beat the market by a few percent over 25 plus
Years with less volatility using multiple screens.

Aussi: “I think the backtester is doing as you say, which is merging the
results together, by reducing money from the screen that has gone up
and adds to the screen that has gone down.”

Very good point, I hadn’t thought about the blender taking money from
the screen that was doing well. That’s not the way I’ve been allocating
funds, in fact I usually move more into the screens that are working (A WWL approach)

RAM

CAGR 19.76 gets my attention. Someone shoot holes in this screen before I start doing it.

Yield-earn-year backtested at a 15%+ CAGR. Until it crashed and turned to crap.

To Ray's point... pre-2014 (the halcyon period for YEY - yes, cherrypicked end date)

	     Avg CAGR	Min	Max	SD
19971231	5.6	0.7	9.5	2.2
19981231	34.3	21.0	45.0	7.9
19991231	0.6	-7.1	8.7	4.4
20001229	34.4	19.9	48.5	8.0
20011231	29.0	23.3	36.2	3.4
20021231	12.7	7.3	20.1	3.7
20031231	61.3	54.6	71.1	4.5
20041231	28.2	21.5	38.5	4.7
20051230	13.6	7.5	18.0	2.7
20061229	26.7	22.4	30.4	2.4
20071231	56.1	46.7	68.6	5.0
20081231	-24.5	-32.2	-13.2	4.7
20091231	35.5	22.8	46.4	5.3
20101231	28.6	23.6	32.4	2.2
20111230	11.4	6.1	17.9	3.3
20121231	29.0	22.8	35.2	3.5
20131231	48.3	39.9	55.2	4.4

And since:

20141231	8.2	4.6	11.9	2.2
20151231	2.4	-1.2	7.1	2.1
20161230	8.5	1.3	17.3	4.3
20171229	16.5	11.0	21.4	3.2
20181231	-3.5	-10.1	1.4	2.7
20191231	23.0	19.9	29.9	2.5
20201231	14.0	5.2	21.4	4.6
20211231	37.2	27.8	48.7	5.3
20221230	-3.1	-12.7	4.2	4.4
20231229	15.9	9.6	21.1	3.8

There was much speculation about "what happened to YEY". One of the impacts was when it started to include MLPs.

Yield-earn-year backtested at a 15%+ CAGR. Until it crashed and turned to crap.

Obviously the point that the blend supports is that while the YEY screen is sucky, the ROE_Cash screen is kicking butt, and "vicey versy".

It would be interesting to see what the ROE_Cash screen backtest look like, and add those columns to FlyingCircus's YEY post.

Tails

(BTW, h/t to FlyingCircus for his post about GTR1...)

Is ROE_Cash the same as what Jim calls LargeCapCash?

I always felt that YEY was rather hinky. Not a lot of holdings, and dividends are pretty much a nothingburger if you believe that dividends are not free money. But it was good while it lasted. ::sigh::

LargeCapCash or the LargeCapCashDividends version intuitively feel more robust. And 40 stocks has a better diversification. I'm not very confident that it will beat the S&P500 these days, since it doesn't include many of the top 7 companies that have been driving the S&P500 lately.

Decisions, decisions, decisions. I suspect that certain people would just say, "To heck with it, just buy BRK."

RAMc,
Do you have a P123 script version of the screen?

thanks,
Taz

Taz2: "Do you have a P123 script version of the screen?"

Sorry but no, for screens I have more confidence in GTR1's daily start dates and the flexibility.
P123 has limited rolling tests but they are not up to GTR1's standards.
For P123 I've been concentrating on understanding their new historical factor downloads for machine learning.
Their documentation for their API and ML factor data downloads isn't very straight forward.
Machine learning isn't the bottleneck for me it's the P123 interface and limitations.

RAM

That said I have consistently beat the market by a few percent over 25 plus
Years with less volatility using multiple screens.

That's interesting to hear. I thought that most people on here had not beaten the market.

Mark19: “That's interesting to hear. I thought that most people on here had
not beaten the market.”

Judging by the number of people that left the board over the years and the way so
many screens that underperformed post discovery it seems obvious that most
people using MI did not do well. Somehow perhaps by a combination of using MI in
the earlier years and just by pure random luck I’ve managed to come out significantly
ahead over the long term. But my outperformance was almost all in earlier years,
my MI has lagged the market since 2008 to the point that I’ve been almost entirely
in Tactical Asset Allocations for the last few years. But interestingly the smaller portion
I’ve kept in screens have been doing well again recently.

RAM

Judging by the number of people that left the board over the years and the way so
many screens that underperformed post discovery it seems obvious that most
people using MI did not do well. Somehow perhaps by a combination of using MI in
the earlier years and just by pure random luck I’ve managed to come out significantly
ahead over the long term. But my outperformance was almost all in earlier years,
my MI has lagged the market since 2008 to the point that I’ve been almost entirely
in Tactical Asset Allocations for the last few years. But interestingly the smaller portion
I’ve kept in screens have been doing well again recently.

One of the more interesting ideas I heard was by Andrew Lo. Physics is not a good analogy for the stock market. The laws of physics never change. Biology is better, since the market always evolves. I think fundamental indexing becoming popular may have hurt the screens. Or maybe just screening becoming so mainstream.

One of the more interesting ideas I heard was by Andrew Lo. Physics is not a good analogy for the stock market. The laws of physics never change. Biology is better, since the market always evolves. I think fundamental indexing becoming popular may have hurt the screens. Or maybe just screening becoming so mainstream.

Another possibility is that our timescales have been too short to see that screens don't work all the time.
We intuitively know that. But, we don't know how long periods of underperformance might last.
If a screen "works" for 10 years, then doesn't work for four years, then works again for three years, is it broke?
I suppose it is, because few will use a screen that doesn't work for four years.

I recall one of the most famous example of a screen that stopped working is RS26, due to overtuning or herd effect.
It would be interesting to see if it still has not worked. I don't have VL, so am not familiar with it.

There are all these potential reasons why screens might not work ... fooled by randomness, herd effect, cycles, quants, machine learning and now AI.
It can also be combinations of these things.


Mark

I thought that most people on here had not beaten the market.

It looks like there are 7 MI screens that have beaten the market over the past 10 years, with annual returns ranging from 13.3% to 17.5%. These are being tracked at https://discussion.fool.com/t/curated-screens-2024...

It looks like there are 7 MI screens that have beaten the market over the past 10 years, with annual returns ranging from 13.3% to 17.5%. These are being tracked at https://discussion.fool.com/t/curated-screens-2024...

That list may be factually true, but the conclusion that many people would draw from it could be a classic Fooled-by-Randomness trap.

If you start with a list of 100 screens, let's say, and rank them by 10 year performance, and the average screen lags the market by 5%, then you will randomly get a few screens that beat the market by a few points and a lot more screens that lag the market by more than 5%. You are virtually guaranteed that, all else being equal, a different set of screens will beat the market over the next ten years.

Elan

Does anyone have the backtest results for ROEPLOW & ROIC?

I am interested in how those screens have held up over the long term.

Does anyone have the backtest results for ROEPLOW & ROIC?

Backtesting of VL screens has not been possible on gtr1 for some years now, after VL decided not not to permit it.

I primarily use portfolio123 for backtesting. It has a user-friendly UI where you can easily construct screens that use factors like ROE (like ROEPLOW does) and backtest them there.

There was a post in 2016 listing returns for {ROEPLOW} and {ROIC}. I made SIP versions. {ROEPLOW} has inconsistent results, maybe because of the zero dividend criteria. {ROIC} might outperform the market by about 2. A ranked version {rankROTC} does better. Results (5 deep, 0.4% friction, 21 day hold):

Screen             CAGR  GSD  MDD  UI  Sharpe    From       To
ROEPLOW             12   28   -48  16   0.53   20060526  20160527
ROIC                 8   25   -59  26   0.43   20060526  20160527
ticker_SPY           7   23   -55  17   0.40   20060526  20160527
ROEPLOW_SI          14   24   -48  14   0.71   20060526  20160527
ROTC_SI             10   25   -57  22   0.50   20060526  20160527
SP1500EqualWeight    9   25   -60  14   0.45   20060526  20160527
rankROTC            12   26   -56  22   0.56   20060526  20160527
 
ROEPLOW_SI           6   22   -36  12   0.31   20131231  20231229
ROTC_SI             11   22   -42  14   0.55   20131231  20231229
SP1500EqualWeight    9   23   -42  8    0.49   20131231  20231229
rankROTC            12   24   -46  16   0.57   20131231  20231229
 
ROEPLOW_SI          13   30   -78  33   0.47   19870302  20240116
ROTC_SI             13   26   -60  22   0.52   19870302  20240116
SP1500EqualWeight   11   22   -60  11   0.50   19870302  20240116
rankROTC            16   28   -62  23   0.61   19870302  20240116

2016 results are from: Comparing GTR1 results and Weekly Trade Ledgers, 06/01/2016
http://www.datahelper.com/mi/search.phtml?nofool=y...

https://gtr1.net/2013/?~ROEPLOW_SI:h21f0.4::styp.a...
https://gtr1.net/2013/?~ROTC_SI:h21f0.4::styp.a:et...
https://gtr1.net/2013/?~rankROTC_20240120_lizgdal:...

Define {ROEPLOW}
Uses [Domicile Code] [Div'd Yield] [% Retained to Common Equity] [Market Cap $ (Mil)] [Total Return 26-Week]
Deblank [Domicile Code] [Div'd Yield] [% Retained to Common Equity] [Market Cap $ (Mil)]
Keep :[Domicile Code]="us"
Keep :[Div'd Yield]=0
Keep :[% Retained to Common Equity]>20
Sort Descending [Market Cap $ (Mil)]
Top :15 PlusTies
Deblank [Total Return 26-Week]
Sort Descending [Total Return 26-Week]
; Top :5
End

Define {ROEPLOW_SI}
step0: [Security Type; lag=1 days] == 0,10,11,12,18,30,31,48
step1: [Mkt Days Since Security Opened; lag=1 days] >= 1265
step2: [MktCap] Top 1700
step3: country is U.S.
step4: [Div] == 0
step5: [Equity] > 0
step6: [ROE] > 0.2
step7: [MktCap] Top 15
step8: [Total Return % over 126 days; lag=1 days] Top 5

Define {ROIC}
Uses [Return on Total Capital] [Market Cap $ (Mil)] [Total Return 13-Week] [Total Return 26-Week] [Total Return 1-Year]
Deblank [Return on Total Capital] [Market Cap $ (Mil)]
Keep :[Return on Total Capital]>=25
Sort Descending [Market Cap $ (Mil)]
Top :20 PlusTies
Deblank [Total Return 13-Week] [Total Return 26-Week] [Total Return 1-Year]
Create [RSW] :.4*[Total Return 13-Week]+.3*[Total Return 26-Week]+.3*[Total Return 1-Year]
Sort Descending [RSW]
; Top :5
End

Define {ROTC_SI}
step0: [Security Type; lag=1 days] == 0,10,11,12,18,30,31,48
step1: [Mkt Days Since Security Opened; lag=1 days] >= 1265
step2: [MktCap] Top 1700
step3: [ROTC] > 0.25
step4: [Equity] > 0
step5: [MktCap] Top 20
step6: [RSW] Top 5

Define {rankROTC}
Create [sumRanks]: [[rMC] + [rROTC] + [rMo]]
step0: [Security Type; lag=1 days] == 0,10,11,12,18,30,31,48
step1: [Mkt Days Since Security Opened; lag=1 days] >= 1265
step2: [MktCap] Top 1700
step3: [Equity] > 0
step4: [sumRanks] Top 5

There was a post in 2016 listing returns for {ROEPLOW} and {ROIC}...

FWIW, for the original VL versions in the 8 years since that post, ROEPLOW has been largely useless (market tracking or a small penalty), but ROIC has done very nicely.
ROIC has beat the S&P by around 3%/year, monthly at any depth from 5 to 25.

ROIC at depth 25 quarterly after trading costs, advantage 3.26%/year, beating the S&P in 85% of the rolling years.
GSD barely above the S&P if you care about such things (16.6 vs 16.0)


As an aside, never underestimate how much randomness there is in any backtest of only 5 or 10 stocks. This is the biggest issue by far on why so many screens have disappointed after they were created: we found the luckiest screens and variants and posted them. The roll-off in rolling 10 year performance is not uncorrelated with the fraction of each 10-year rolling period that was within the original lucky backtest window.

Jim

Another possibility is that our timescales have been too short to see that screens don't work all the time.
We intuitively know that. But, we don't know how long periods of underperformance might last.
If a screen "works" for 10 years, then doesn't work for four years, then works again for three years, is it broke?...

True enough.

e.g., YldEarnYear got absolutely pasted in the pandemic crash.
But someone with perfect foresight would have decided to pile in just then.
Starting April 2020 to end 2023 (3.75 years), the standard screen 10 deep monthly beat the S&P by 22.8%/year. (CAGR 41.4% vs 18.6%)
Without holding any of the magnificent seven, that's quite an accomplishment.

Even at 25 stocks deep it beat SPY by 19%/year.

Jim

As an aside, never underestimate how much randomness there is in any backtest of only 5 or 10 stocks. This is the biggest issue by far on why so many screens have disappointed after they were created: we found the luckiest screens and variants and posted them.

Interesting theory. If this is true, it seems we should see evidence that screens that performed well in backtest at deeper/larger numbers of holds (say holding 30-40 stocks) have continued to perform well in forward test.

Has anyone done an analysis like that? To see if the top ranked screens by that metric have performed better in forward test than those that are bottom ranked by that metric?

Rather than looking at total return for the full hold, another metric could be developed based on breaking up the screen ranking positions of stocks into "mini" portfolios and looking at how consistent they are out to deeper levels, comparing 1-10 to 11-20 to 21-30 to 31-40. Those that remain strong out to deeper ranks might be expected to hold up better in forward test, based on this theory.

Interesting theory. If this is true, it seems we should see evidence that screens that performed well in backtest at deeper/larger numbers of holds (say holding 30-40 stocks) have continued to perform well in forward test.

Many if not most of the previously tested screens would not qualify for such a test because they have a final sorting step for top 10.

Elan

True enough.

Very nice to have GTR1 working. It did not take too much to throw this quick analysis of YEY vs SP500 together, FWIW.

This is a comparison of rolling (1,2,3,4,5) year returns of YEY_SI vs SP500 from 19870302 through 20240122.
On all the graphs, the blue line is YEY_SI and the orange line is the SP500.
Periods where the blue line is below the orange line are those dreaded underperformance periods.

Rolling 1-year returns https://1drv.ms/i/s!AmCzJdeikH-cipEnRd-0-2gmvcrBMw...
Rolling 2-year returns https://1drv.ms/i/s!AmCzJdeikH-cipEor8fLrC4PgzzkMQ...
Rolling 3-year returns https://1drv.ms/i/s!AmCzJdeikH-cipEpZT0vJ0bup7wmfQ...
Rolling 4-year returns https://1drv.ms/i/s!AmCzJdeikH-cipEq9LFdcThtwVosLw...
Rolling 5-year returns https://1drv.ms/i/s!AmCzJdeikH-cipErzrpeUTrs2AcWeQ...

Not sure if this is a good way to look at it. On the 5 year graph, it is notable that from 2017-2020 there is a clear period of underperformance.
It is almost unique in the history, although there was a shorter period of underperformance in the late '90's.

For reference, for YEY_SI I used the SI version that Robbie posted in post #256567 (note that it did not do as good as 41.4% after the pandemic bottom):
https://gtr1.net/2013/?~YLDEARNYEAR_SI:h21::styp.a...
For the SP500 return, the total return index ^S5T is used.
Daily portfolio values from 19870302 through 20240122 for both were exported into a spreadsheet.
From there, rolling (1,2,3,4,5) year returns were calculated for each.
For example, starting on 19880301, you can calculate the 1 year return from 19870302. The next day, you calculate another rolling 1 year return.


Mark

Not sure if this is a good way to look at it. On the 5 year graph, it is notable that from 2017-2020 there is a clear period of underperformance.

I think it's very informative.
I like the 3 year a bit better for information content, but whatever.

What would be interesting as an overlay:
An equally weighted portfolio of the universe of qualifying dividend payers that the YEY portfolio was using.
I mention this because the vast underperformance of YEY for a while was in large part because of the vast underperofrmance of all decent dividend payers in general during that stretch.
From (say) Dec 2010 to April 2020, the S&P was up 164% and an equally weighted portfolio of VL stocks that YEY might pick (top half of dividend payers by yield) returned 77%.
The steps in the screen other than "look for a decent dividend" really had their work cut out for them, starting from a 4.7%/year handicap.
As a generalization, all dividend screens did very poorly for a few years, simply because all dividend paying stocks did pretty poorly. The screens were picking from a bag of laggards.


Ramblings on the same general subject---

Sometimes I look at stock performance broken into three categories:
* Non dividend payers
* Dividend payers, the lower half by yield
* Dividend payers, the upper half by yield

Among other things, the very recent relative performance of the portfolios (rank order) can be used as an input to a not-bad market timing signal.
Ties back to Zeelotes' aggressive/defensive signal family.

Jim

Mungofitch posted:
What would be interesting as an overlay:
An equally weighted portfolio of the universe of qualifying dividend payers that the YEY portfolio was using.

Here is the rolling 3-year graph with the addition of the qualifying dividend payers.
So the graph has the SP500 total return, YEY_SI top 10, and YEY-SI according to this screen:
https://gtr1.net/2013/?~YLDEARNYEAR_SI:s19870302h2...
Comparing to the standard YEY-SI screen https://gtr1.net/2013/?~YLDEARNYEAR_SI:h21::styp.a...
Steps 4,5,6 are omitted, meaning without the ratio of dividend yield to PE and without the final sort.
GTR1 reports that an average of 1144 stocks had a dividend yield > 0.

https://1drv.ms/i/s!AmCzJdeikH-cips9YcAol_W97JOq5g...

It is clear that the ratio(cdy,cpe) top 10% and the final sort have a very large impact on the result.
Without those, the screen is a tad bit better than the S&P, but not a lot.
Other than that, I'm not sure what the conclusions are.

There was a period 1990-1997 where YEY_SI outperformed greatly, but the dividend yielders were just about the same as the index.
So in that period, the final steps added much value.
However, in 2016-2022 that is not the case, FWIW.


Mark

Post New

US Policy FAQ

---

Contact Shrewd'm
Contact the developer of these message boards.