Post New

G-score is a metric developed about 20 years ago to rank growth stocks, with one point given for 8 measures. I made a gtr1 screen based on Mohanram's G-score, and get similar results: stocks with higher G-scores had higher returns. Results from 19880302 to 20241107 for the screen {GscoreHigh_dv50p}:

  Gscore   CAGR  GSD  count
HIGH(6,7)   13   22    126
LOW(0,1)    -3   32     40
HIGH-LOW    16

G-score is only calculated for the lowest 20% of Book-to-Market stocks. I calculated a new field [GAscore] that has no Book value criteria. This increases the number of passing stocks, with similar results:

 GAscore   CAGR  GSD  count
HIGH(6,7)   14   19    427
LOW(0,1)     3   28    183
HIGH-LOW    11

Combining GAscore with 5-year sales growth has good CAGR, but also high GSD.
Define {GAscoreSG}
step0: [Security Type] == 11
step1: [Mkt Days Since Security Opened] >= 63
step2: [Average dollar-volume over 63 days] > 0
step3: [Average dollar-volume over 63 days] Top 50%
step4: [SalesY1] > 0
step5: [SalesY6] > 0
step6: [GAscore] == 7
step7: [[SalesY1]/[SalesY6]] Top 10; Cash When Fewer
Holding period = 21 mkt days, 0.4% friction

Results from 19880302 to 20241107:

      Screen       CAGR  GSD  MDD  Sharpe  Beta  AT
    GAscoreSG       17   33   -61   0.59   1.3   3
SP1500EqualWeight   12   21   -60   0.54   1.1   0

===================
Details:
Mohanram backtested G-score from 1979 to 1999. From the abstract: "This paper tests whether a strategy based on financial statement analysis of low book-to-market (growth) stocks is successful in differentiating between winners and losers in terms of future stock performance. I create an index (G_SCORE) based on a combination of traditional fundamentals such as earnings and cash flows and measures appropriate for growth firms such as the stability of earnings and growth and the intensity of R&D, capital expenditure and advertising. A strategy based on buying high G_SCORE firms and shorting low G_SCORE firms consistently earns significant excess returns... The stock market in general and analysts in particular are much more likely to be positively surprised by firms whose growth oriented fundamentals are strong, indicating that the stock market fails to grasp the future implications of current fundamentals." (See Mohanram's 2005 paper "Separating Winners from Losers among Low Book-to-Market Stocks using Financial Statement Analysis".)

G-score has 8 parts, with one point given for each:
G1: Return on Assets (ROA) exceeding the industry median.
G2: Cash flow ROA exceeding the industry median.
G3: Cash flow from operations exceeding net income.
G4: Earnings variability (measured by the variance of ROA over the past five years) being less than the industry median.
G5: Sales growth variability (variance over the past five years) being less than the industry median.
G6: R&D intensity (R&D expenditure as a percentage of total assets) exceeding the industry median.
G7: Capital expenditure intensity exceeding the industry median.
G8: Advertising intensity exceeding the industry median.

Mohanram used the 2-digit SIC code as the industry. Medians were calculated using the lowest 20% of Book-to-Market stocks (including stocks with negative equity). Mohanram did not deblank the fields, but assigned 0 to the variability measures if less than three years of information was available. All firms with price and book value information were included. Returns for 1 year and 2 year holds were calculated from 1979 to 1999, with a 21% CAGR difference between HIGH and LOW G-score stocks.

  G-score    count  pct   MeanReturnY1  firm-years
    ALL       994   100%      8.2%        20866
HIGH(6,7,8)   109   11%       17.4%        2290
  LOW(0,1)    157   16%       -4.0%        3295
  HIGH-LOW                    21.4%

G-score has many calculation steps, and so I separated the task into 2 parts: calculate G-scores, and then screen stocks. The gtr1 field [Gscore] is calculated uses SIP industry codes, and does not include G8.
step0 to step6: deblanking steps.
step7: Book-to-Market bottom 20%.
step8: calculate industry medians and [Gscore].

The gtr1 screen {GscoreHigh} imports the field [Gscore]
step0: Gscore = 6 or 7
(253 mkt days Holding period, 0.4% Friction)

Results from 19880302 to 19991231 (count is the average count of stocks passing the screen) are similar to Mohanram's results, with a 22% CAGR difference between HIGH and LOW Gscore stocks. (The time periods are different: Mohanram's backtest was 21 years ending in 1999, this {GscoreHigh} backtest was 12 years ending in 1999.)

  Gscore   count  pct   CAGR  GSD
   >=0     1215   100%  9.4   22
HIGH(6,7)   170   14%   20.0  19
LOW(0,1)    169   14%   -1.9  26
HIGH-LOW                21.9

Results, with 24 years post-discovery, from 19880302 to 20241107 continue to show higher CAGR for higher Gscore:

  Gscore   count  pct   CAGR  GSD
   >=0     1053   100%  6.0   27
HIGH(6,7)   154   15%   13.4  22
LOW(0,1)    134   13%   -4.1  32
HIGH-LOW                17.5

The screen {GscoreHigh_dv50p} uses the same [Gscore] calculation, but adds liquidity criteria to the screening stage.
import [Gscore] calculated using all stocks.
step0: no ETFs, CEFS, or LPs.
step 1 to step3: dollar volume top 50%
step 4: Gscore = 6 or 7
step 5: recent 10Q report.
step 6: no recent mergers.
(253 mkt days Holding period, 0.4% Friction)

Results from 19880302 to 20241107 continue to show higher CAGR for higher Gscore:

  Gscore   count  pct   CAGR  GSD
   >=0      579   100%  8.3   26
HIGH(6,7)   126   22%   12.8  22
LOW(0,1)     40    7%   -3.2  32
HIGH-LOW                16.0

G-score is only calculated for the lowest 20% of Book-to-Market stocks. I calculated a new field [GAscore] that has no Book value criteria. ([GAscore] calculates industry medians using all stocks. In contrast, [Gscore] calculates industry medians using only Book-to-Market bottom 20% stocks.)
Results from 19880302 to 20241107 show higher CAGR for higher GAscore:

 GAscore   count  pct   CAGR  GSD
   >=0     2387   100%  10.4  23
HIGH(6,7)   427   18%   13.6  19
LOW(0,1)    183    8%   2.6   28
HIGH-LOW                10.9

links to screens:

{GscoreHigh}
https://gtr1.net/2013/?!!QlpoMTFBWSZTWX0CUwoAB7Dfg...

{GscoreHigh_dv50p}
https://gtr1.net/2013/?!!QlpoMTFBWSZTWaewUKoACDXfg...

{GAscoreHigh_dv50p}
https://gtr1.net/2013/?!!QlpoMTFBWSZTWSl0!2FLUAB47...

{GAscoreSG}
https://gtr1.net/2013/?!!QlpoMTFBWSZTWciKjbcAB8jfg...

Lizgdal

Awesome coding of GTR1!

Could you please describe this field

Create [COAInd]: [0.5*[Top 50% Cut-Off by [COA], grouped by [IndGroup], at step8] + 0.5*[Bottom 50% Cut-Off by [COA], grouped by [IndGroup], at step8]]

I think it is meant to calculate the median, but I don't understand how. If I have 5 values (1,2,3,4,5), what is the value for Top 50% and Bottom 50%? If I have 6 values (1,2,3,4,5,6) what are the values?

Thanks

Aussi

Results, with 24 years post-discovery, from 19880302 to 20241107 continue to show higher CAGR for higher Gscore

Impressive. What, though, are the results strictly post-discovery: 20000101 to 20241107? That would eliminate the influence of the test period.

Eric Hines

Some industries only have a few stocks. The average of bpcut 50% and tpcut 50% is close to the median.

=======
Example that finds median of data: 1,2,3,4,5
number of data points = N = 5

bpcut 50%
0. sort data in ascending order: 1,2,3,4,5
1. bx = (N * Bottom%/100 rounded down) = 2
2. 2nd data point is 2

tpcut 50
0. sort data in descending order: 5,4,3,2,1
1. tx = (N * Top%/100 rounded down) = 2
2. 2nd data point is 4

=========
The average of tpcut and bpcut is equal to the median when the sample size is even. Some examples with 20241107 data:

Ticker  [IndGroup]  [ROAq1]  tpcut50  bpcut50  [ROAInd]
  NC      5010101     -7%      11%      -1%       5%
 HNRG     5010101     -1%      11%      -1%       5%
  BTU     5010101     9%       11%      -1%       5%
 ARCH     5010101     11%      11%      -1%       5%
 CEIX     5010101     15%      11%      -1%       5%

Ticker  [IndGroup]  [ROAq1]  tpcut50  bpcut50  [ROAInd]
  VET     5010201     -13%     7%       6%        6%
  MG      5010201     -1%      7%       6%        6%
  E       5010201     3%       7%       6%        6%
 SHEL     5010201     4.7%     7.2%     5.7%     6.5%
  TGS     5010201     5.7%     7.2%     5.7%     6.5%
 EQNR     5010201     7.2%     7.2%     5.7%     6.5%
  PBR     5010201     7.4%     7.2%     5.7%     6.5%
  TTE     5010201     8%       7%       6%        6%
  SU      5010201     8%       7%       6%        6%
  CVE     5010201     8%       7%       6%        6%

https://gtr1.net/2013/?~checkG1:lf-1lp-1h1::Assets...

Gscore results have been fairly consistent. Looking at a few 10-year periods, HIGH-LOW range was around 13 to 23.

                           CAGR
  From       To     yrs  HIGH-LOW
19880302  19991231  12      22       test data
19991231  20241107  25      16     post-discovery
19880302  20241107  37      18         overall
 
19880302  19980302  10      23     rolling 10-yr
19941104  20041105  10      20     rolling 10-yr
20041105  20141107  10      13     rolling 10-yr
20141107  20241107  10      18     rolling 10-yr

  Screen    CAGR  GSD  MDD    From       To     yrs  count
GscoreAll   6.0   27   -74  19880302  20241107  37   1053
GscoreHigh  13.4  22   -57  19880302  20241107  37    154
GscoreLow   -4.1  32   -89  19880302  20241107  37    134
 
GscoreAll   9.4   22   -41  19880302  19991231  12   1215
GscoreHigh  20.0  19   -31  19880302  19991231  12    170
GscoreLow   -1.9  26   -68  19880302  19991231  12    169
 
GscoreAll     4   28   -74  19991231  20241107  25    975
GscoreHigh   10   23   -57  19991231  20241107  25    146
GscoreLow    -5   34   -89  19991231  20241107  25    117
 
GscoreAll     7   26   -50  20141107  20241107  10    854
GscoreHigh   14   21   -35  20141107  20241107  10    146
GscoreLow    -3   33   -73  20141107  20241107  10    106
 
GscoreAll     7   24   -62  20041105  20141107  10    967
GscoreHigh   12   20   -51  20041105  20141107  10    131
GscoreLow    -1   29   -72  20041105  20141107  10    109
 
GscoreAll     5   33   -74  19941104  20041105  10   1333
GscoreHigh   15   27   -57  19941104  20041105  10    190
GscoreLow    -5   39   -84  19941104  20041105  10    170
 
GscoreAll     6   19   -33  19880302  19980302  10   1176
GscoreHigh   18   18   -26  19880302  19980302  10    164
GscoreLow    -5   22   -57  19880302  19980302  10    164

G-score can be summarized as:
Cash flow from operations exceeds net income.
High Return on Assets (ROA) and Cash flow ROA.
Low Earnings variability and Sales growth variability.
High R&D and Capital expenditures.

some links:

Man+Machine – The evolution of fundamental scoring models and ML implications, October 8, 2021

1968 Altman’s Z-Score: "praised at the time for a significant level of accuracy of 94% in predicting company bankruptcy one year before the event"

1996 Sloan ratio: "Sloan’s main conclusion is that earnings driven by accruals are less persistent than those low accruals."

1999 Beneish’s M-Score: "easy and quick way to spot the probability of earnings manipulation of a stock"

2014 SEC Account Quality Model: "AQM is designed to detect whether a company’s financial disclosures differ greatly from its industry peers."

2000 Piotroski’s F-Score: "main purpose is to deliver a single metric to allow investors to adopt a systematic approach to fundamental value investing... Piotroski’s model inherently favours an investment style with a clear value tilt"

2005 Mohanram’s G-Score: "with the objective of finding a systematic way of investing in growth stocks"

"A recently published study by Amor-Tapia and Tascon in 2016 tests multiple scoring models using four different European stock markets. Surprisingly, Mohanram’s G-Score and Piotroski’s F-Score were the only ones surviving an out-of-sample backtesting delivering statistically significant alpha despite the other rejected models being developed at a later date."
https://www.cfauk.org/pi-listing/man-machine-the-e...

In January 2024, yuvaltaylor posted: "I wrote about the F-Score at length here: Why Piotroski's F-Score No Longer Works. There are a number of other X-Score thingies out there. I use a modified version of Beneish’s M-Score, which I outlined here: Detecting Financial Fraud: A Close Look at the Beneish M-Score, as well as Ohlson’s O-Score and Dechow’s F-Score. There’s also Altman’s Z Score. They all have merits and problems. One problem with these scores is that they mainly rely on multilinear regression rather than weighted ranking, and that’s just not as good a method."
https://community.portfolio123.com/t/man-machine-t...

In Febuary 2020, yuvaltaylor posted: "here’s my formula for the g-score"
https://community.portfolio123.com/t/scoring-syste...

Eric: “What, though, are the results strictly post-discovery: 20000101 to 20241107?”
Just put a starting date 20000103 in lizgdal’s G-Score high GTR1 link and I got:

CAGR 11.6, GSD 20.6, UI 9.2, SAWR 9.9%

In the paper fundamental equity analysis “Man+Machine – The evolution of fundamental scoring models and ML implications”.
Evolution of Fundamental Scoring Models | CFA UK
They state:
“Despite their many weaknesses and newly developed scoring models challenging their dominance flourishing over the last two decades, both Piotroski’s F-Score and Mohanram’s G-Score have not been ditched by investors. A recently published study by Amor-Tapia and Tascon in 2016 tests multiple scoring models using four different European stock markets. Surprisingly, Mohanram’s G-Score and Piotroski’s F-Score were the only ones surviving an out-of-sample backtesting delivering statistically significant alpha despite the other rejected models being developed at a later date.”
You referenced P123’s Yuval Taylor post with his version of a G Score equation above which you can use in their ranking system or a simulation. But P123 Also has a Mohanram GScore Screen.
I just ran a Backtest using P123's GScore Screen for equities in the SP1500 from 1/3/2002.
25 stocks, 4 week rebalance, 0.25% slippage, CAGR 10.7% vs S&P 9.4% nice but not exceptional.
However almost all the outperformance came in the last 15 years.

Rerunning a rolling backtest(multiple start dates)
11/11/2014 – 11/11/2024
523 Samples, Start Frequency every Week, Holding Period 3 Months:
Scrn SP500
Annualized 20.06% 13.33%
One Month 4.46% 3.21%
Three Month 15.96% 12.58%
One Year 52.85% 39.75%
Three Year 23.57% 33.76%
Five Year 159.67% 109.17%
Total 522.85% 249.80%
Risk
Sharpe Ratio 0.91 0.74
Sortino Ratio 1.27 0.99
Max Drawdown -34.2% -33.7%
Standard Dev 19.6% 15.3%
Correlation 0.83 -
R-Squared 0.68 -
Beta 1.06 -
Alpha (annualized) 6.01%

Nice work, especially for an Annual screen (low trading for those that are taxed)!

Just adding the standard BCC (and a small change to 252 trading days for reasons - doesn't change the results really) to (mostly) reduce the moderately frequent negative years already improves the GscoreHigh screen to ...

Results				
Statistics are calculated from 19880302 to 20241108 over daily closing portfolio values for all 252 trading cycles.				

  			Avg 		Min 		Max 		SD 
 				
CAGR: 			15.028172	14.20667	15.926455	0.29794
TR: 			16984.583984	12976.708984	22527.361328	1629.75708
Log2TR: 		7.410031	7.030856	7.821925	0.137083
SAWR(20; 0.95): 	10.612482	9.939157	11.330691	0.300106
GSD(20): 		18.980368	18.355295	19.567358	0.24761
DIGSD(20; 0%): 		21.148109	20.386206	21.979813	0.377738
LDD(20; 0%): 		11.445286	11.032278	11.952506	0.22102
LDDD3: 			8.119306	7.436218	8.881767	0.268355
MDD: 			-38.182415	-46.089619	-33.897877	2.281894
UI(20): 		11.809819	10.573466	13.663349	0.571028
Sharpe(20): 		0.731354	0.683201	0.77604		0.015233
Beta(20): 		0.783587	0.7678		0.801364	0.007492
TI(20): 		16.158493	15.192879	17.07169	0.333716
AT: 			1.107518	1.064199	1.148228	

Slightly higher CAGR (+1.36pp), lower GSD (-2.93pp), circa +1.5 SAWR & lower drawdowns (-38.2 vs. -57.2) at the expense of greater AT (1.11 vs. 0.63)(but you'd mostly want to trade to cash at those times anyway). A fine starting point for a screen (perhaps needs a final sort).

https://gtr1.net/2013/?~GscoreHigh:h252f0.1::Gscor...

Thanks for posting the P123 results for comparison.

I just ran a Backtest using P123's GScore Screen for equities in the SP1500 from 1/3/2002.
25 stocks, 4 week rebalance, 0.25% slippage, CAGR 10.7% vs S&P 9.4%

The 10.7 CAGR is slightly lower than the gtr1 results:

           screen             CAGR  SDcg  depth    from       to     P123          notes
GAscore_top25_SP1500_20dhold  11.5  0.5     49   20020103  20241108  10.7      all GAscore=7
Gscore_top25_SP1500_20dhold   12.5  0.5     83   20020103  20241108        mix of Gscore=6 and 7
      SP500MktCapWeight       9.6   0.0    499   20020103  20241108  9.4

There is not much space between GAscore and Gscore results, and so it's difficult to say which one P123 is using. (There are more than 25 GAscore=7 stocks, with an average count of 49 passing stocks. There are sometimes fewer than 25 Gscore=7 stocks, and so the screen sometimes adds in Gscore=6 picks.)


Rerunning a rolling backtest(multiple start dates) 11/11/2014 – 11/11/2024
523 Samples, Start Frequency every Week, Holding Period 3 Months:
Scrn SP500
Annualized 20.06% 13.33%

The 20.1 CAGR is higher than the gtr1 results:

           screen             CAGR  SDcg  depth    from       to     P123
GAscore_top25_SP1500_20dhold  10.8  0.9     51   20141111  20241108
Gscore_top25_SP1500_20dhold   16.2  0.7     91   20141111  20241108  20.1
      SP500MktCapWeight       13.5  0.0    500   20141111  20241108  13.3

Growth stocks have higher returns recently, and so Gscore_top25 has higher returns.

Having an explicit tie break would make comparisons easier. For example, using price as a tie break makes the depth a consistent 25:
Price > 1
SortField = Gscore + 1/Price
SortField top 25

               Screen                 CAGR  SDcg  Depth    From       To
GAscoreLowPrice_top25_SP1500_20dhold  9.3   0.5     25   20141111  20241108
GscoreLowPrice_top25_SP1500_20dhold   16.7  0.9     25   20141111  20241108

Growth stocks (BP bottom 20%) have higher CAGR in the last 10 years:

      Screen       CAGR  GSD  MDD  Sharpe  Beta    From       To     Depth
    SP1500BPq1     7.6   34   -60   0.34   1.4   20141111  20241108   298   BP top 20%
    SP1500BPq2     8.9   24   -46   0.44   1.1   20141111  20241108   298
    SP1500BPq3     9.0   22   -41   0.47   1.1   20141111  20241108   295
    SP1500BPq4     10.4  20   -37   0.56   1.1   20141111  20241108   298
    SP1500BPq5     13.2  21   -37   0.68   1.1   20141111  20241108   298   BP btm 20%
SP1500EqualWeight  10.3  23   -42   0.51   1.1   20141111  20241108  1499

A SAWR of > 10% is perhaps the highest I have seen for an annual screen.

Aussi

the top 5 of the Nasdaq100 ranked by 200 day return generates a 18% CAGR with annual trading.

Post New

Berkshire Hathaway FAQ

---

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