No. of Recommendations: 11
(For reference see
https://web.archive.org/web/20170315000958/http://... )
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: of 284296
Subject: Post-Discovery Performance of the SI Pro screens
Date: 1/20/2007 2:25 AM
No. of Recommendations: 60
Recently there has been much interest on this board on the post discovery returns of the VL screens. Some of these now have reasonably long post-discovery records, and it seemed timely to examine their post-discovery returns. In this work, I compare the post-discovery returns of the 42 standard Si Pro screens, reported each week by Keelix, to the Russell 2000 index (RUT).
The discovery dates, post #s and authors for these 42 screens are (in alphabetic order):
Screen Post# Author Date
3pt_Relative_Value 189123 ?? 5/29/06
78RPM 157494 DrBob2 2/12/04
Advanced 186991 Retieff 4/11/06
ARS 152264 DrBob2 11/16/03
BI 189123 ?? 5/29/06
Bob 186688 Retieff 4/11/06
CANSLIM-26 140325 DrBob2/AAII 2/02/03 AAII screen 1997
Commodity 153726 DrBob2 12/17/03
Eastwood 159437 DrBob2 3/22/04
FCF-26 159215 DrBob2/AAII 3/16/04
Fried 500 178201 DrBob2 11/19/05
GSX 183654 Astrophool 2/12/06
GSX2 187289 Astrophool 4/24/06
Gentle Screamers 147947 DrBob2 7/28/03
GS Mungo 178610 mungofitch/DB2 11/29/05
GS Mungo Voom 178610 mungofitch/DB2 11/29/05
High Relative Value 162979 DrBob2 6/24/04
Incoming_Cash 194485 winker 12/20/06
Low_Mult 183654 ?? 2/12/06
LowPS 189599 ?? 6/12/06
Melange 157347 DrBob2 2/10/04
Microcap Momentum 151991 DrBob2 11/10/03
Net-nets 112249/58 DB2/Grahamified 12/06/01
Optiman 95695 DrBob2 3/07/01
PIH Naked 158442 DrBob2 2/26/04
POG 160600 DrBob2 4/15/04
POI 168792 DrBob2 1/22/05
P/S I Love You 167780 winker 12/27/04 SI version H52EarnPS
RS-100 189123 ?? 5/20/06
Quality Earnings 169778 DrBob2 2/17/05
S&P; Peg 147882 DrBob2 7/26/03
Shrinkage 141725 DrBob2 2/22/03
Silver Parachute 155908 DrBob2 1/17/04
Small Value 111479 DrBob2 11/20/01
Steady Growth 133476 DrBob2 9/19/02
Turnarounds 156628 DrBob2 1/27/04
Up5X3 187511 Astrophool 4/28/06
Up 5% 159290 DrBob2 3/18/04
Value at the Top 134696 DrBob2 10/18/02
WK_Voom 183654 ?? 2/12/06
Z26saTA 186773 Retieff 4/12/06
Zweig-26 129802 DrBob2/AAII 7/20/02 AAII Jan/02
To investigate the returns of these screens, I combined two jobs run
on Keelix's backtester:
Job #138869: all 42 screens, plus RUT, SP500, held 5 deep,
traded monthly from the start of monthly data (1997-08-31)
through 2006-12-29.
Job #138870: all 42 screens, plus RUT, SP500, held 5 deep,
traded weekly from the start of weekly data (2003-01-03)
through 2006-12-29.
The data from these runs were combined to use the monthly data from 1997-08-31 to 2003-01-03 and weekly data thereafter. This database was then interpolated to weekly dates (Fridays) beginning 1997-09-05 (the first Friday in the dataset). Of course the interpolated weekly data between 1997 through 2003 is not meaningful, but the analysis of CAGR will be unaffected and it makes the analysis easier having a regular time series. For each screen, the CAGR was then calculated from the start of data on 1997-09-05 through the Friday following the screen discovery date (PreCAGR in the following table). A few (5) screens are missing SI data fields in the early years; for these screens the CAGR is computed from the start of data through the discovery date. Also, for each screen, the CAGR is computed from the Friday following the discovery date (PosCAGR) through year end 2006 (2006-12-29). The GSD is computed only from the start of true weekly data (2003-01-03) through 2006-12-29. Much work on this board has shown that screen GSD is a good predictor of out-of-sample GSD. Therefore, we assume the 2003-2006 GSD is typical of the entire dataset. It's the best we can do with the limited weekly data available.
Next the Russell 2000 (RUT) CAGR is calculated for the same post discovery period for each screen (PosRUT). The RUT was chosen as the market index most typical of the stocks selected for the SI Pro screens. The excess performance of the screen over the RUT is dCAGR= PosCAGR - PosRUT. The number of weeks pre-discovery (#Pre) and post-discovery (#Pos) are also shown in the table.
Finally I computed the cumulative excess screen return (not annualized) over the Russell 2000 for the entire post discovery period, divided by the GSD scaled to a period of this length. This result (called dCAGR/GSD in the table heading) gives the excess post-discovery return of each screen over the Russell 2000 in units of screen standard deviations. A screen that did as well as the overall market would have a zero entry in this column. Note that this number is NOT just the ratio of the (annualized) dCAGR and GSD columns (I agree the nomenclature is not the best but it's late at night and I'm not going to fix it now). Assuming these statistics are normally distributed (they may well not be), then 2 standard deviations (a 2-sigma deviation) in screen return over the Russell 2000 has only a 5% probability of occurring by chance.
It is seen that some screens with several years of post-discovery data do not significantly outperform the Russell 2000 index. But many SI Pro screens do significantly outperform the Russell 2000, while not a single SI Pro screen significantly underperformed the Russell 2000. To further emphasize the significance of this result, outperforming screens (dCAGR/GSD > 2 sigma) with long post-discovery records (#Pos > 104 weeks) were selected. There are 12 screens that meet these criteria. An equiweighted blend of these 12 outperforming screens had a CAGR= 45 and GSD= 18 over the 104 week post-discovery period, compared to the Russell 2000's CAGR= 13. This is an outperformace (dCAGR/GSD) of 3.05 sigma, corresponding to a chance probability (assuming normal statistics) of only 0.2%. The total number of stocks held in this blend is 5*12= 60, over 104 weeks, so the total number of stock picks in the post-discovery result is 6240, which exceeds mungofitch's value of 4000 for believable backtests.
Furthermore, although I have not tried to quantify this further, there does not appear to be a decline in outperformance with increasing time from discovery. Two of the earliest SI Pro screens, Small Value and the Net-Nets, with post-discovery backtests exceeding 5 years, have demonstrated very substantial (6-7 sigma!) outperformance over the Russell 2000. This result is very different from recent results reported for the VL screens.
Here are the results (the 12 outperforming screens are in bold):
Screen Discovery_Date PreCAGR PosCAGR PosRUT dCAGR #Pre #Pos GSD dCAGR/GSD
3pt_Relative_Val 29-May-2006 52.92 49.15 22.29 26.87 457 30 24.70 0.75
78RPM 12-Feb-2004 64.34 73.30 11.20 62.10 337 150 44.66 4.04
Advanced 11-Apr-2006 97.11 24.15 2.84 21.31 306 37 36.90 0.48
ARS 16-Nov-2003 14.58 40.77 12.45 28.32 325 162 24.05 3.16
BI 29-May-2006 67.02 56.69 22.29 34.40 457 30 27.48 0.85
Bob 11-Apr-2006 56.43 25.90 2.84 23.05 450 37 25.59 0.75
CANSLIM-26 2-Feb-2003 14.78 28.05 22.21 5.84 284 203 29.10 0.67
Commodity 17-Dec-2003 71.83 29.91 12.22 17.69 329 158 20.67 2.05
Eastwood 22-Mar-2004 36.64 29.89 10.10 19.79 343 144 23.89 1.77
FCF-26 16-Mar-2004 31.04 10.29 12.10 -1.81 342 145 22.47 -0.15
Fried_500 19-Nov-2005 31.06 12.61 12.76 -0.15 430 57 21.56 -0.01
GSX 12-Feb-2006 54.41 68.22 8.06 60.16 442 45 26.55 2.04
GSX2 24-Apr-2006 62.54 43.55 1.11 42.44 452 35 32.31 1.04
Gentle_Screamers 28-Jul-2003 27.34 72.91 17.46 55.45 309 178 36.10 6.22
GS_Mungo 29-Nov-2005 51.36 13.45 13.28 0.17 431 56 22.61 0.01
GS_Mungo_Voom 29-Nov-2005 50.20 24.92 13.28 11.64 431 56 22.32 0.55
High_Relative_Va 24-Jun-2004 42.67 53.31 12.71 40.60 356 131 20.21 4.66
Incoming_Cash 20-Dec-2006 52.40 0.00 0.00 0.00 486 1 69.01 0.00
Low_Mult 12-Feb-2006 98.77 58.32 8.06 50.25 163 45 32.30 1.41
LowPS 12-Jun-2006 88.66 33.41 27.84 5.57 459 28 42.45 0.09
Melange 10-Feb-2004 47.41 56.47 11.20 45.27 337 150 37.12 3.21
Microcap_Momentu 10-Nov-2003 43.66 19.68 13.75 5.93 324 163 44.10 0.28
Net-Nets_Grahami 6-Dec-2001 24.60 69.90 10.39 59.51 223 264 65.72 6.17
OptiMan 7-Mar-2001 18.76 10.77 9.53 1.24 184 303 12.87 0.34
PIH_Naked 26-Feb-2004 54.27 8.52 10.06 -1.54 339 148 33.17 -0.08
POG 15-Apr-2004 61.30 25.53 11.20 14.33 346 141 38.91 0.72
POI 22-Jan-2005 111.06 22.25 11.62 10.63 387 100 35.77 0.45
P/S
Love_You 27-Dec-2004 41.93 54.44 13.33 41.11 383 104 29.18 2.52
RS-100 20-May-2006 85.04 27.75 11.70 16.06 456 31 36.64 0.33
Quality_Earnings 17-Feb-2005 31.57 21.13 12.00 9.13 390 97 32.12 0.42
S&P;_Peg 26-Jul-2003 30.81 31.72 17.46 14.26 309 178 23.20 1.77
Shrinkage 22-Feb-2003 12.95 34.52 22.88 11.64 287 200 19.28 2.23
Silver_Parachute 17-Jan-2004 27.42 28.98 10.91 18.07 334 153 26.77 1.51
Small_Value 20-Nov-2001 32.25 46.53 11.05 35.48 221 266 28.96 6.88
Steady_Growth 19-Sep-2002 19.56 21.60 20.32 1.28 264 223 17.57 0.26
Turnarounds 27-Jan-2004 11.84 11.25 10.77 0.47 335 152 35.66 0.02
Up5X3 28-Apr-2006 79.00 36.64 1.11 35.53 452 35 24.03 1.17
Up_5% 18-Mar-2004 47.69 22.69 12.10 10.60 342 145 31.26 0.69
Value_at_the_Top 18-Oct-2002 11.72 32.51 19.32 13.19 268 219 22.89 2.22
WK_Voom 12-Feb-2006 121.92 81.83 8.06 73.77 163 45 34.65 1.91
Z26saTA 12-Apr-2006 86.05 40.95 2.84 38.10 224 37 29.09 1.07
Zweig-26 20-Jul-2002 -8.50 90.64 18.07 72.56 30 231 37.45 16.21
Outperforming 12-screen blend 44.31 45.42 13.33 32.09 383 104 18.54 3.05
Screens in blend:
78RPM ARS Commodity Gentle_Screamers
High_Relative_Va Melange Net-Nets_Grahami P/SLove_You
Shrinkage Small_Value Value_at_the_Top Zweig-26
=================
Author: FlyingCircus Big red star, 1000 posts Old School Fool
Number: 195804 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/20/2007 8:04 PM
No. of Recommendations: 3
Great work!
One comment is that some of those GSDs put the screens in the Forbidden Zone, even with more data points used & created. For instance, 78RPM with 44.66, NetNets with 65.72. The high GSDs put those screens purely across that desert wasteland with a big flashing sign of "You Take The Risk!"
From that list I'd start with Shrinkage and HRV as candidates - GSD of 20 is sustainable risk in a blend.
Matt
=================
Author: TGMark Big red star, 1000 posts Old School Fool
Number: 195805 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/20/2007 11:24 PM
No. of Recommendations: 1
I agree, nice summary.
I'm not sure I understand the scaling the GSD to the post-discovery length thing. How exactly did you do this?
An observation is that the high performing screens are all older ones, as they've had a longer post-discovery period and this scaling affects the dCAGR/GSD favorably.
Newer screens, such as Up5X3, GSX2, BI, Bob, Advanced, etc score poorly in comparison.
That seems a little curious, but the newer screens have very short post discovery periods, so concluding anything about them is probably difficult.
Mark
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195806 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/20/2007 11:39 PM
No. of Recommendations: 5
One comment is that some of those GSDs put the screens in the Forbidden Zone, even with more data points used & created. For instance, 78RPM with 44.66, NetNets with 65.72. The high GSDs put those screens purely across that desert wasteland with a big flashing sign of "You Take The Risk!"
From that list I'd start with Shrinkage and HRV as candidates - GSD of 20 is sustainable risk in a blend.
I totally agree with you. My list of outperforming screens wasn't meant as a recommendation to invest in these -- it was simply the list of screens that outperformed the Russell 2000 for a minimum 2-yr post-discovery period. Personally, I avoid screens with GSD > 30, although the risk of including a high GSD screens in a large blend (such as this 12-screen blend) is still fairly small. In fact, removing the two high GSD screens you mention (78RPM, Net-Nets) only drops the blend GSD from 18.5 to 17.7. Here is the result for the 10-screen blend (dCAGR/GSD > 2.0, #Pos > 104 weeks, GSD < 40):
PreCAGR PosCAGR PosRUT dCAGR #Pre #Pos GSD dCAGR/GSD
Best screen blend 38.71 49.20 13.33 35.87 383 104 17.69 3.63
Screens in blend:
ARS Commodity Gentle_Screamers High_Relative_Va
Melange P/SLove_You Shrinkage Small_Value
Value_at_the_Top Zweig-26
Being even more restrictive and keeping only screens with GSD < 30 (dCAGR/GSD > 2.0, #Pos > 104 weeks, GSD < 30) brings the resulting 7-screen blend GSD down to 16.0:
PreCAGR PosCAGR PosRUT dCAGR #Pre #Pos GSD dCAGR/GSD
Best screen blend 38.04 41.97 13.33 28.64 383 104 15.97 3.14
Screens in blend:
ARS Commodity High_Relative_Va P/SLove_You
Shrinkage Small_Value Value_at_the_Top
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195807 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 12:12 AM
No. of Recommendations: 5
TGMark wrote:
I'm not sure I understand the scaling the GSD to the post-discovery length thing. How exactly did you do this?
The rightmost column labeled dCAGR/GSD is the ratio of the cumulative return over the post-discovery period to the GSD over this period (not the annualized value shown in the GSD column). The easiest way to show exactly what I did is to write this in pseudocode. CumRet refers to the weekly screen cumulative return (from an initial value of 1.0 at the start of data), RUTCumRet is the same quantity for the Russell 2000. IncrRet is the incremental change from week to week expressed as a fraction (not percentage). They are related by: IncrRet[i] = CumRet[i]/CumRet[i-1] where the index refers to the i-th week of the dataset.
ScreenRet= ( CumRet[2006-12-29]/CumRet[Discovery_Date] - 1. )*100.
RUTRet= ( RUTCumRet[2006-12-29]/RUTCumRet{Discovery_Date] - 1.)*100.
Numerator= ScreenRet - RUTRet
SD= StdDev( Log10( IncrRet[2003-01-03]: IncrRet[2006-12-29] )*sqrt( Post_discovery weeks )
Denominator= ( 10.0^SD - 1. )*100.
dCAGR/GSD= Numerator/Denominator
An observation is that the high performing screens are all older ones, as they've had a longer post-discovery period and this scaling affects the dCAGR/GSD favorably.
Newer screens, such as Up5X3, GSX2, BI, Bob, Advanced, etc score poorly in comparison.
That seems a little curious, but the newer screens have very short post discovery periods, so concluding anything about them is probably difficult.
Some of the new screens appear to be good performers. Compare, e.g., the very good post-discovery CAGR for GSX and UP5X3 to the comparable Russell 2000 CAGR. It's just that the post-discovery period for these screens is so short that this difference isn't (yet) terribly statistically significant (although dCAGR/GSD for GSX > 2 after only 45 weeks post-discovery). Longer post-discovery periods allows the stochastic component of the screen returns (which determine the GSD) to be separated from the mean return (the CAGR). This is easily seen for a hypothetical well-behaved screen which always returns R each cycle, with a consistent standard deviation of S. Then after N cycles, CumRet= R^N, and SD(N-cycles)= S^sqrt(N). Thus the ratio CumRet/SD increases with the length of the backtest.
My guess is that some of these new screens will also be outperformers. Time will tell. Stayed tuned for updates (I will probably post these each quarter).
=================
Author: DrBob2 Big funky green star, 20000 posts Top Recommended Fools Old School Fool
Number: 195811 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 8:09 AM
No. of Recommendations: 4
One comment is that some of those GSDs put the screens in the Forbidden Zone, even with more data points used & created. For instance, 78RPM with 44.66, NetNets with 65.72.
Now there's an image for you: Ben Graham as a wild and crazy, high volatility trader.
For the last several years there have been very few net-nets available; perhaps that effected the GSD (I would use a basket of 10 or 12 with an annual hold.)
DB2
=================
Author: behzat One star, 50 posts Old School Fool
Number: 195823 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 3:11 PM
No. of Recommendations: 0
Thank you for this very helpful summary.
I noticed a problem.Could you verify one point.
In job #138869 ( Monhly hold) CAGR of the blend is 55 but looking to returns of 1997/11/28 monthly return is 39.8 % which I am not sure this is correct.All the individual screens of the blend average is -1.36% for that month which is very different of blend's return for that month.
If the blend consists of all the screens of this job then the problem is screen POI's 1701 % return for this month!!! which I doubt is correct
Behzat
=================
Author: comewhatmay One star, 50 posts Old School Fool
Number: 195826 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 6:59 PM
No. of Recommendations: 0
Another anomaly in that job #138869 is the Commodity screen for 2002-03-28, showing a return of 2179.63% for the month. Any blend I run having that data point naturally has a GSD which doesn't look too good.
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195828 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 8:04 PM
No. of Recommendations: 9
Thank you for this very helpful summary.
I noticed a problem.Could you verify one point.
In job #138869 ( Monhly hold) CAGR of the blend is 55 but looking to returns of 1997/11/28 monthly return is 39.8 % which I am not sure this is correct.All the individual screens of the blend average is -1.36% for that month which is very different of blend's return for that month.
If the blend consists of all the screens of this job then the problem is screen POI's 1701 % return for this month!!! which I doubt is correct
Well, this problem was *supposed* to be taken care of. Some of the early data fields in the SI Pro database contain bad data, and properly, the way to fix this is to run each screen separately, and ignore the offending ticker. There are several others of these that I noticed, e.g.: the Commodity screen has an monthly change of 2179.63% for 2002-03-28, Fried-500 has 83.44% on 2002-08-30, etc. For this calculation, I was a bit lazy and just capped the extreme incremental changes to either +40% (for a large positive change) or -40% (for a large negative change). Unfortunately, due to a bug in my program, these changes did not get propagated to my routine that interpolated these changes to weekly changes, which I then combined with job #138870 to produce this report.
So, indeed, some of the pre-discovery screens' CAGR and GSD have been skewed by bad data points. After redoing the calculation with the extreme data points capped to +/- 40%, the 1997-2006 42-screen blend CAGR is slightly reduced (by -3%) to CAGR= 41.35, GSD= 17.30, CAGR/GSD= 2.39. None of this affects the main results and conclusions of this post. The post-discovery screen CAGR and GSD results are unchanged, as is the conclusion that several of the SI Pro screens have significantly outperformed the general market (represented by the Russell 2000 index). For completeness, here are the corrected pre/post-discovery results for the 42 standard SI Pro screens:
Screen Discovery_Date PreCAGR PosCAGR PosRUT dCAGR #Pre #Pos GSD dCAGR/GSD
3pt_Relative_Val 29-May-2006 52.92 49.15 22.29 26.87 457 30 24.70 0.75
78RPM 12-Feb-2004 64.34 73.30 11.20 62.10 337 150 44.66 4.04
Advanced 11-Apr-2006 97.11 24.15 2.84 21.31 306 37 36.90 0.48
ARS 16-Nov-2003 14.58 40.77 12.45 28.32 325 162 24.05 3.16
BI 29-May-2006 67.02 56.69 22.29 34.40 457 30 27.48 0.85
Bob 11-Apr-2006 56.43 25.90 2.84 23.05 450 37 25.59 0.75
CANSLIM-26 2-Feb-2003 14.78 28.05 22.21 5.84 284 203 29.10 0.67
Commodity 17-Dec-2003 10.56 29.91 12.22 17.69 329 158 20.67 2.05
Eastwood 22-Mar-2004 36.64 29.89 10.10 19.79 343 144 23.89 1.77
FCF-26 16-Mar-2004 28.95 10.29 12.10 -1.81 342 145 22.47 -0.15
Fried_500 19-Nov-2005 26.85 12.61 12.76 -0.15 430 57 21.56 -0.01
GSX 12-Feb-2006 54.41 68.22 8.06 60.16 442 45 26.55 2.04
GSX2 24-Apr-2006 62.54 43.55 1.11 42.44 452 35 32.31 1.04
Gentle_Screamers 28-Jul-2003 27.34 72.91 17.46 55.45 309 178 36.10 6.22
GS_Mungo 29-Nov-2005 42.72 13.45 13.28 0.17 431 56 22.61 0.01
GS_Mungo_Voom 29-Nov-2005 41.92 24.92 13.28 11.64 431 56 22.32 0.55
High_Relative_Va 24-Jun-2004 42.67 53.31 12.71 40.60 356 131 20.21 4.66
Incoming_Cash 20-Dec-2006 42.65 0.00 0.00 0.00 486 1 32.32 0.00
Low_Mult 12-Feb-2006 98.77 58.32 8.06 50.25 163 45 32.30 1.41
LowPS 12-Jun-2006 86.39 33.41 27.84 5.57 459 28 42.45 0.09
Melange 10-Feb-2004 47.41 56.47 11.20 45.27 337 150 37.12 3.21
Microcap_Momentu 10-Nov-2003 43.66 19.68 13.75 5.93 324 163 44.10 0.28
Net-Nets_Grahami 6-Dec-2001 19.31 62.10 10.39 51.72 223 264 57.03 5.65
OptiMan 7-Mar-2001 18.76 10.77 9.53 1.24 184 303 12.87 0.34
PIH_Naked 26-Feb-2004 41.77 8.52 10.06 -1.54 339 148 33.17 -0.08
POG 15-Apr-2004 61.30 25.53 11.20 14.33 346 141 38.91 0.72
POI 22-Jan-2005 49.73 22.25 11.62 10.63 387 100 35.77 0.45
P/SLove_You 27-Dec-2004 41.93 54.44 13.33 41.11 383 104 29.18 2.52
RS-100 20-May-2006 84.58 27.75 11.70 16.06 456 31 36.64 0.33
Quality_Earnings 17-Feb-2005 31.57 21.13 12.00 9.13 390 97 32.12 0.42
S&P;_Peg 26-Jul-2003 30.81 31.72 17.46 14.26 309 178 23.20 1.77
Shrinkage 22-Feb-2003 12.95 34.52 22.88 11.64 287 200 19.28 2.23
Silver_Parachute 17-Jan-2004 27.42 28.98 10.91 18.07 334 153 26.77 1.51
Small_Value 20-Nov-2001 32.25 46.53 11.05 35.48 221 266 28.96 6.88
Steady_Growth 19-Sep-2002 19.56 21.60 20.32 1.28 264 223 17.57 0.26
Turnarounds 27-Jan-2004 11.84 11.25 10.77 0.47 335 152 35.66 0.02
Up5X3 28-Apr-2006 79.00 36.64 1.11 35.53 452 35 24.03 1.17
Up_5% 18-Mar-2004 47.69 22.69 12.10 10.60 342 145 31.26 0.69
Value_at_the_Top 18-Oct-2002 10.82 32.51 19.32 13.19 268 219 22.89 2.22
WK_Voom 12-Feb-2006 121.92 81.83 8.06 73.77 163 45 34.65 1.91
Z26saTA 12-Apr-2006 86.05 40.95 2.84 38.10 224 37 29.09 1.07
Zweig-26 20-Jul-2002 -8.50 90.64 18.07 72.56 30 231 37.45 16.21
Outperforming screen blend 36.81 45.42 13.33 32.09 383 104 18.18 3.12
Screens in blend:
78RPM ARS Commodity Gentle_Screamers
High_Relative_Va Melange Net-Nets_Grahami P/SLove_You
Shrinkage Small_Value Value_at_the_Top Zweig-26
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195829 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 8:16 PM
No. of Recommendations: 0
comewhatmay wrote:
Another anomaly in that job #138869 is the Commodity screen for 2002-03-28, showing a return of 2179.63% for the month. Any blend I run having that data point naturally has a GSD which doesn't look too good.
Note that the GSD reported (2nd column from the right) is calculated only using the weekly data from Jan 2003 on. Most of the data anomalies predate this period and do not affect this GSD calculation (which is why the POI and Commodity GSDs remain unchanged in my revised report). A minor exception occurs for the Net-Nets, for which the revised GSD is slightly reduced, due to the capping (at +40%) of the 76.56% weekly change for this screen on 2004-11-12.
Phil
=================
Author: elann Big funky green star, 20000 posts Top Favorite Fools Top Recommended Fools Old School Fool
Number: 195837 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 11:28 PM
No. of Recommendations: 2
A minor exception occurs for the Net-Nets, for which the revised GSD is slightly reduced, due to the capping (at +40%) of the 76.56% weekly change for this screen on 2004-11-12.
I think that instead of capping the return at 40% it's better to correct anything above 40% down to zero. A one month return of 40%, if it's false, is enough to elevate the CAGR of a multi-year test by several percentage points.
Elan
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195839 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/21/2007 11:42 PM
No. of Recommendations: 2
I think that instead of capping the return at 40% it's better to correct anything above 40% down to zero. A one month return of 40%, if it's false, is enough to elevate the CAGR of a multi-year test by several percentage points.
I agree that just setting the anomalous weekly/monthly to zero would be a fairer way of rejecting outliers. The problem is that there are quite a few entries around 30% and most of these are probably valid. I selected a cutoff by plotting all the incremental weekly/monthly changes for all the screens against the Russell 2000 changes for the same period (273 cycles * 42 screens = 11,466 points), and looking where the distribution cut off. There was a fairly sharp edge above +35% and below about -30%, so I used +/- 40% to reject obvious outliers (the points with a 2000% increase in a single month,etc.). I don't want to reject the tails of the real return distribution, so in the end I decided to stay with my +/- 40% cutoff. At same point I will go through and clean up the bad stocks in these screens, which is the proper way of doing this.
Phil
=================
Author: Keelix Big red star, 1000 posts Old School Fool
Number: 195844 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/22/2007 4:20 AM
No. of Recommendations: 2
At same point I will go through and clean up the bad stocks in these screens, which is the proper way of doing this.
If you do and make it available in a list or something I can promise I will put it into the backtester so the rest of us can reap the benefits from your good work.
Keelix
=================
Author: AstroPhool Three stars, 500 posts Old School Fool
Number: 195868 of 284296
Subject: Re: Post-Discovery Performance of the SI Pro scr
Date: 1/22/2007 5:49 PM
No. of Recommendations: 0
If you do and make it available in a list or something I can promise I will put it into the backtester so the rest of us can reap the benefits from your good work.
Thanks, Keelix. I will post the list of what I find.
Phil
