Index | Recent Threads | Who's Online | Search

Posts: 373    Pages: 38    Prev 29 30 31 32 33 34 35 36 37 38 Next
Last Post
This topic has been viewed 17986 times and has 372 replies
yuvaltaylor
Re: FactSet beta site v1.0, NOW LIVE

@ Dan: For example -100/10 would be 100/10=10 --> that might be a great stock! It just might have just turned around and the price might not reflect that -->
I am with Jim here, it might not capture a growth, but it can be a good sentiment factor, especially on small caps.

I tried:

(CurFYEPSMean - CurFYEPS4WkAgo) /Max(0.02, Abs(CurFYEPS4WkAgo)) [formula #1]
(NextFYEPSMean - NextFYEPS4WkAgo) / Max(0.02, Abs(NextFYEPS4WkAgo)) [formula #1a]

and I tried
(CurFYEPSMean - CurFYEPS4WkAgo) / Abs(CurFYEPS4WkAgo) [formula #2]
(NextFYEPSMean - NextFYEPS4WkAgo) / Abs(NextFYEPS4WkAgo) [formula #2a]

they give (much!) less performance then

Abs(CurFYEPSMean) / Abs (CurFYEPS4WkAgo) [formula #3]
Abs (NextFYEPSMean) / Abs(CurFYEPS4WkAgo) [formula #3a]

Both above factors in one node and a rank weight of 12.5%.

at least with my ranking systems (which has some heavy quality and momentum components) and a filter on small caps on the buy rules


Let's look at what these factors are actually doing rather than judging them on "performance."

Abs(CurFYEPSMean) / Abs (CurFYEPS4WkAgo). This would rank stocks that had an EPS estimate of -0.13 four weeks ago and an EPS estimate of -0.01 right now very low. A stock that had an EPS estimate of -0.01 four weeks ago and an EPS estimate of -0.13 right now would be ranked extremely high. Is this what you want?

Abs (NextFYEPSMean) / Abs(CurFYEPS4WkAgo). If you want to compare next year's EPS estimate to this year's EPS estimate, why would you compare it to what this year's EPS estimate was four weeks ago rather than to this year's EPS estimate today? And once again, using Abs in the numerator makes no logical sense.

It's always important to judge factors on their logic BEFORE judging them on their performance.

As for their performance, if you run a 10-week ranking test on the PRussell 3000 excluding stocks with a price under $3, rebalancing every four weeks over the last 10 years, the top decile gets 11.1% with the formula #1, 11.0% with formula #2, and 10.2% with formula #3. It gets 12.7% with formulas #1a and #2a and only 8.3% with formula #3a. (I numbered and labeled the formulas above.)

Yuval Taylor
Product Manager, Portfolio123
invest(igations)
Any opinions or recommendations in this message are not opinions or recommendations of Portfolio123 Securities LLC.

Jul 16, 2020 10:19:11 AM       
Jrinne
Re: FactSet beta site v1.0, NOW LIVE

without judging any other formulas:

(CurFYEPSMean - CurFYEPS4WkAgo) / Abs(CurFYEPS4WkAgo)
(NextFYEPSMean - NextFYEPS4WkAgo) / Abs(NextFYEPS4WkAgo)

Using top decile (or quintile) minus bottom decile (or quintile) over max period rank performance both of these are highly statistically significant with t-scores greater than 4.

You might check the last 5 years, however. But then again, you might want to check all of your factors over the last 5 years.

BTW, this is what Zacks does in their screener. I do not use Zacks but any reputation they may have is based their handling of earnings estimates.

Edit: To expand slightly if you read Modern Portfolio Theory and Investment Analysis (9th Edition) this would fall under "event driven" market changes.

This equation just ensures that there is a relatively new (a month or less) event: a recent improvement in the consensus estimates. The equation also addresses the magnitude of this change. I do not think it is about growth. It is about how efficient the market is or how quickly it adapts to the new information—with earnings being a proved market driver.

The academics still have not decided how quickly the market adapts to this new information. Nor am I prepared to make any claims about the last 5 years.

Best,

Jim

From time to time you will encounter Luddites, who are beyond redemption.
--de Prado, Marcos López on the topic of machine learning for financial applications

Jul 16, 2020 10:41:27 AM       
Edit 6 times, last edit by Jrinne at Jul 16, 2020 11:01:45 AM
RTNL
Re: FactSet beta site v1.0, NOW LIVE

How long was the hold before re-balancing, Jim?

And since P123 does not do the significance tests, I assume you are doing it else where. If you don't mind, can you share where you do these?

Jul 16, 2020 11:07:40 AM       
Edit 1 times, last edit by RTNL at Jul 16, 2020 11:09:12 AM
Jrinne
Re: FactSet beta site v1.0, NOW LIVE

How long was the hold before re-balancing, Jim?

And since P123 does not do the significance tests, I assume you are doing it else where. If you don't mind, can you share where you do these?

RT,

Weekly. The image is weekly (FY equation above). NAs neutral so NAs do not affect the top or bottom quintile returns.

The Excel download from this page can easily be manipulated to get the t-score for top quintile minus bottom quintile.

Best,

Jim

Attachment Earnings Estimates.png (346316 bytes) (Download count: 59)


From time to time you will encounter Luddites, who are beyond redemption.
--de Prado, Marcos López on the topic of machine learning for financial applications

Jul 16, 2020 11:15:29 AM       
Edit 2 times, last edit by Jrinne at Jul 16, 2020 11:18:16 AM
RTNL
Re: FactSet beta site v1.0, NOW LIVE

Thanks, Jim.

Jul 16, 2020 11:20:03 AM       
InspectorSector
Re: FactSet beta site v1.0, NOW LIVE

Using top decile (or quintile) minus bottom decile (or quintile) over max period rank performance both of these are highly statistically significant with t-scores greater than 4.


So the first question is: How are you handling N/As Are they set to neutral in the ranking system? Otherwise a top decile versus bottom decile may be a meaningless analysis.

Steve

Jul 16, 2020 12:34:39 PM       
Edit 1 times, last edit by InspectorSector at Jul 16, 2020 12:35:23 PM
Jrinne
Re: FactSet beta site v1.0, NOW LIVE

From above:

NAs neutral so NAs do not affect the top or bottom quintile returns.

Steve,

This is an EXCELLENT QUESTION and your reminder of something that could be overlooked is much appreciated.

I did already think of this (this time).

Best,

Jim

From time to time you will encounter Luddites, who are beyond redemption.
--de Prado, Marcos López on the topic of machine learning for financial applications

Jul 16, 2020 12:40:36 PM       
InspectorSector
Re: FactSet beta site v1.0, NOW LIVE

I don't have any experience with FactSet's estimate data so I am going to base these comments on my experience with the previous data vendor. My comments are as follows:

- historical data was not Point In Time. i.e. the estimates were spilling over into Monday morning and the data was being overwritten. Therefore, the most recent week is untrustworthy.

- GIven that the most recent week is not PIT, you have to avoid the most recent week. The formulae (CurFYEPSMean - CurFYEPS4WkAgo) / Abs(CurFYEPS4WkAgo) and (NextFYEPSMean - NextFYEPS4WkAgo) / Abs(NextFYEPS4WkAgo) are not valid as PIT backtests.

- You have to modify the formulae to the following: (CurFYEPS1WkAgo - CurFYEPS4WkAgo) / Abs(CurFYEPS4WkAgo) and (NextFYEPS1WkAgo - NextFYEPS4WkAgo) / Abs(NextFYEPS4WkAgo)

Backtest results:
- using the original formulae on the PRussell3000, 5-Year RS performance with weekly rebalance, ignoring the fact that the results are probably not PIT... Only the bottom decile provides usable results. There is no monatonic increase in buckets. This likely means that an extreme negative change in earnings estimates may forecast lower stock prices ahead. However keep in mind that this may not be PIT and this may not be exploitable.
- using the modified formula (avoiding the most recent week), there is no useful advantage.

See attached.

Attachment Earnings Growth.gif (94340 bytes) (Download count: 47)


Jul 16, 2020 1:38:57 PM       
InspectorSector
Re: FactSet beta site v1.0, NOW LIVE

BTW, this is what Zacks does in their screener. I do not use Zacks but any reputation they may have is based their handling of earnings estimates.

Zacks determines which analysts give the best ratings and then they look for an increase in earnings estimates from those analysts specifically, immediately before the earnings report comes out. This info is used in the Zacks rating. We don't have this ability using P123.

Jul 16, 2020 1:43:38 PM       
Edit 1 times, last edit by InspectorSector at Jul 16, 2020 1:44:11 PM
Jrinne
Re: FactSet beta site v1.0, NOW LIVE

Steve,

The last 5 years are interesting but I would prefer not to live in interesting times. I cannot add much to what you illustrate here.

Not limited to earnings estimates however. Here is EBITDATTM/EV. This universe is the PRussell 3000 also.

Enough to make me not want to discount walk-forward methods that do not look back the entire 20 years. Something that adjusts to more recent events. Not that I have much out-of-sample data to show (either way).

Best,

Jim

Attachment EBITDATTM:EV.png (90311 bytes) (Download count: 45)


From time to time you will encounter Luddites, who are beyond redemption.
--de Prado, Marcos López on the topic of machine learning for financial applications

Jul 16, 2020 1:51:26 PM       
Edit 6 times, last edit by Jrinne at Jul 16, 2020 2:11:50 PM
Posts: 373    Pages: 38    Prev 29 30 31 32 33 34 35 36 37 38 Next
 Last Post