Please be respectful of others' privacy, and avoid sharing personal information or sensitive content without their permission. If you are unsure if something is appropriate to share, ask for permission (use the 'Privately email' option when replying to their post) or avoid sharing it altogether.
- Manlobbi
Stocks A to Z / Stocks B / Berkshire Hathaway (BRK.A)
No. of Recommendations: 7
Profitability subsumes all of “quality” investing, explaining both the performance of the strategies that industry markets and the factors that academics employ. It also has striking power pricing “defensive equity” strategies that overweight low-beta or low-volatility stocks. Profitability tilts explain all the abnormal performance of popular “alternative value” strategies, including those adjusted for “intangibles,” and half of value's post-2007 underperformance. Profitability is crucial for pricing a wide array of seemingly unrelated anomalies, yielding a more parsimonious understanding of the cross section of expected returns.
https://papers.ssrn.com/sol3/papers.cfm?abstract_i...
No. of Recommendations: 8
There are 2 versions of the paper, one free and one costing $5. Profitability is:
(REVT−COGS−(XSGA−XRD)−XINT) / (BE+MIB)
REVT is Sales
COGS is Cost-of-Goods-Sold
XSGA is selling, general, and administrative expenses
XRD is R&D expenditure
XINT is interest expense
BE is book-equity
MIB is minority interest
So, roughly: (net income + R&D) / equity
or, ROE + R&D/equity
No. of Recommendations: 9
(REVT−COGS−(XSGA−XRD)−XINT) / (BE+MIB)
...
So, roughly: (net income + R&D) / equity
or, ROE + R&D/equity
In the right direction, but not that great an approximation - ROE would usually be after-tax after-depreciation after-everything income. The paper makes the case for using the most "gross" income metric available.
Minority interests are rarely a big factor and could probably be skipped.
So a simple version would be (Gross margin - [non-R&D SG&A expenses] - interest) / book
(meaning gross margin as a dollar amount, not percentage: total revenues minus cost of goods sold)
Anybody up to an SIpro implementation?
Personally, I'd want to include profit averaged over more than a year. A lot of very good firms have somewhat variable gross earnings.
If I used this as a screen to pick companies to invest in, I'd add a sanity test that their debt not be too high. A firm that paid out all its book value as unsupported dividends would have a small denominator and a big result.
I usually compare debt to [cyclically adjusted] net income...if it would take more than 5-10 years of profit to wipe out the debt, it's leveraged. Under 5 years is pretty much always safe, up to 10 for the very most reliable earners, and over that is too junky for me.
Jim
No. of Recommendations: 4
The Profitability factor seems completely rational to me but unless I’ve bungled this somehow, I show a historical inverse relationship to profitability and return.
I used the Portfolio123’s FactSet equivalents for the Compustat factors to give me a quarterly based profitable formula without the minority factor in the denominator:
Notes on converting Compustat to P123
REVT:compustat total revenue I used Sales for REVT
COGS:Cost of Goods Sold
compustat SALE-COGS -> P123 GrossProfit
XSGA:Administrative expenses
XRD:Research and Development Expense
compustat XSGA-XRD -> P123 SGandA
XINT: Interest and Related Expense Total
-> P123 IntExp
BE: Book Equity
-> P123 BookValQ
MIB: Minority Interest, portion of net income (not parent company)-- I left this out
Final P123 Formula: (SalesQ -CostGQ -SGandAQ -IntExpQ)/BookValQ
Running a 20 year 20 bin backtest on the SP1500 results in a set of nicely declining total return bins with the lowest bin having a 13.4% annual return and the highest bin having a 5.8% return. The universe had a 10.1% return.
For such a small number of factors this does very well but in the inverse of what it supposed to be.
So did I misunderstand or misinterpret the profitability formula?
Excess vs. Benchmark Entire Period
Benchmark 10.24744
Bucket 1 3.113532 3.1% above universe
Bucket 2 3.521425
Bucket 3 2.903438
Bucket 4 1.987549
Bucket 5 -0.67059
Bucket 6 0.609326
Bucket 7 -0.10803
Bucket 8 0.740753
Bucket 9 2.175545
Bucket 10 -10.2474
Bucket 11 -10.2474
Bucket 12 -1.83211
Bucket 13 -0.27276
Bucket 14 -0.0286
Bucket 15 -0.73058
Bucket 16 -0.99633
Bucket 17 -1.86139
Bucket 18 -2.70347
Bucket 19 -1.5514
Bucket 20 -4.47536 4.5% less than universe
No. of Recommendations: 0
Should note that the bins 10 and 11 had no returns which results in -10% vs universe because I eliminated
equities which didn't report one of the terms for the period.
No. of Recommendations: 0
mungo: I'd want to include profit averaged over more than a year
I ran the screen using 13 week hold and all factors previous 12 month. The results were virtually unchanged.
No. of Recommendations: 5
OOps! Found there was a little bubble on the menu that I missed inversing the ranking. So indeed, the profitability formula gives positive results.
No. of Recommendations: 10
Profitability subsumes all of “quality” investing, explaining both the performance of the strategies that industry markets and the factors that academics employ. It also has striking power pricing “defensive equity” strategies that overweight low-beta or low-volatility stocks. Profitability tilts explain all the abnormal performance of popular “alternative value” strategies, including those adjusted for “intangibles,” and half of value's post-2007 underperformance. Profitability is crucial for pricing a wide array of seemingly unrelated anomalies, yielding a more parsimonious understanding of the cross section of expected returns.
Wow, talk about indecipherable jargon. The average word length is in inverse proportion to intelligibility.
Elan
No. of Recommendations: 4
I know that most here have doubts that selecting equities by Machine Learning are in the same category as equating the S&P500 prediction to butter in Bangladesh production. But as a comparison to using the profitability equation which has a 10.2% advantage over the universe in the upper bin. My novice ML system still in the early development stage currently has backtests with a 17% CAGR advantage (20% over the last 5 years). It works with gradually declining returns by selecting 10, 25, 50 or even a 150 stocks with lower turnover than my old screens. The problem is that ML has turned out to be an order of magnitude more difficult than the academic papers or textbooks make it appear.
No. of Recommendations: 2
Sorry for the dumb question, but precisely what generated those results ?
For example, Jim produced similar looking results of a "value bin" versus a "forward return", for both BRK and S&P500, based on the P/B and P/E 'value measures' (respectively, iirc).
IIRC, his results weren't the results of investing using a screen, he simply looked at the forward CAGR of each stock in the universe for e.g. 3 years forward, 5 years forward, etc and associated these with bins of the value measure. He also averaged CAGRs a bit over endpoint dates but that's not conceptually critical.
In your table you used the 'value meaure'
Final P123 Formula: (SalesQ -CostGQ -SGandAQ -IntExpQ)/BookValQ
Does SGandAQ include R&D expense or not?
I ran the screen using 13 week hold and all factors previous 12 month. The results were virtually unchanged.
So you created a screen/backtest based on this value measure, and your table is the result of this?
Could you elaborate on the screen, e.g. is there some threshold for which you drop stocks, etc? Or perhaps I misunderstood?
Lastly, most here have doubts that selecting equities by Machine Learning
I don't, if it was done correctly. Have you posted about your work on it before?
No. of Recommendations: 4
Tetherdog: but precisely what generated those results ?
First the formula:
The SSRN paper used the formula (REVT−COGS−(XSGA−XRD)−XINT) / (BE+MIB)
The terms are from compustat, I wanted to test this on Portfolio123 which now uses factset data but in the
Past used compustat and I have a table cross referencing P123 and compustat. I tried to show the relationship
Of compustat and P123 above but I’ll try again:
Compustat REVT revenue total
Compustat COGS cost of goods sold
Compustat SALE-COGS is equivalent to P123 GrossProfit ??don't understand compustat REVT vs SALE
Compustat XSGA Administrative expenses
Compustat XRD Research and Development Expense
And Compustat XSGA-XRD is equivalent to P123 SGandA
Compustat XINT Interest and Related Expense Total is equivalent to P123 IntExp
Compustat BE Book Equity is equivalent to BookVal
Compustat MIB is minority interest which is almost always negligible and I could not find a P123 equv so I left it out
Result a P123 version of the formula base on the last quarterly values is:
(SalesQ -CostGQ -SGandAQ -IntExpQ)/BookValQ
Tetherdog: So you created a screen/backtest based on this value measure, and your table is the result of this?
Could you elaborate on the screen, e.g. is there some threshold for which you drop stocks, etc? Or perhaps I misunderstood?
Not exactly, I created a sort of histogram of the performance equations actual return over history:
1. For the 1st historical date rank all 1500 of the SP1500 stocks using the performance equation
2. Sort the resulting 1500 by the descending value of the performance equation
3. Group the equities into 20 equal “predicted” performance bins from highest to lowest
4. Average the gain for each bin over the future sample period in this case 4 weeks. If the prediction has any value the first bins will have a higher average gain than the last bins.
5. Repeat the above for the next period and accumulate the gains from each bins.
The resulting 1st bin gives the actual performance of the highest predicted by the performance equation and the 20 bin the lowest predicted. There was no calculation of any trading friction nor any measurement of how consistent the ranking was between sample date.
Last Yes I have posted about Machine Learning here but the feedback wasn’t exactly positive, the topic is different enough that ML posts are better posted on a ML blog than here. But thanks for the bringing up the performance equation, I’ll obviously drop it into my ML factors and see if it helps.
No. of Recommendations: 2
Thank you! I guess the comment about 'screen' threw me off
I ran the screen using 13 week hold and all factors previous 12 month. The results were virtually unchanged.
The reported gain in your table is the short-term gain over the subsequent 4 weeks to each date you calculate that metric?
I'm probably misunderstanding something again, it seems surprising that a relatively simple metric based on quarterly reported data would correlate well with short-term, i.e. four week, performance.
Jim had a somewhat similar table where he looked at bins of P/B for BRK and showed a relation to long-term CAGRs.
No. of Recommendations: 4
Tedthedog: it seems surprising that a relatively simple metric based on quarterly reported data would correlate well with short-term, i.e. four week, performance.
The factors change only during the first week of a quarter and remain the same for the entire quarter. So a stocks rank will stay the same for the entire quarter and will very likely remain in the same bin and be reselected for each of the following 4 week samples remaining in the quarter. I looked at the average hold time for the top bin and it was averaged over a year if I remember correctly. But as there was no penalty for trade friction it would not make a difference.
I just ran a simulation with trading friction from 1/1/2002 till present. Selecting the top ranked 45 stocks. The results are not particularly good. Only 12.45 CAGR (poor) with a 43.5% annual turnover (excellent low taxes) a max drawdown of 45.5% (9% less than the universe) 59% winners, 0.7 Sharp. Although this isn’t spectacular it is outstanding for one simple formula. Add a few simple additions like mungo’s “sanity test that their debt not be too high” and this could be the foundation for a nice screen. But I’ll stick with MI.
No. of Recommendations: 0
Interesting! Thanks for sharing your work.
