Using R in equity research

Andrew,

I had been using Minitab in my MBA program at NYU, but
my professor of regression and data analysis Jeff
Simonoff suggested that ambitious students might try
R. I didn't use it for his course but gradually taught
myself to perform regression analysis and diagnostics
with R. 

I'm definitely not from a mathematical/statistical
background. I've been driven to statistics by a desire
to test assumptions I form about fundamental company
research and equity analysis.

Most equity research people tend to not check
diagnostics very well, and thus violate basic

I think that is generally true 'on the street'.

assumptions I think, because they'd rather be
overconfident about their assumptions. That's an
over-generalization though.

I've been doing valuation studies within industries,
looking at how some valuation measures relate to
company characteristics and external factors over
time. My professor suggested using mixed effects
models for such longitudinal data studies, and I don't
have a budget for this sort of thing, so using R and
the NLME package was a natural choice. I think I'm
doing some things with it I haven't seen other
analysts do.  

Sounds interesting. Do you have any write-ups?

I notice that most financial users of R and S-Plus
tend to be focused on derivatives and time-series
applications. I haven't seen many approaching time
series data from a multi-factor relationship
perspective. Anyone else seen good applications,
publications on that kind of analysis?

For what I know, it is used fairly extensive in portfolio management (for
equity as well as other portfolio) but I don't have a quick reference I
could point you to for more.  

Suggestions, anyone?

Best regards, Dirk

FEATURE:  VW Beetle license plate in California

I've been doing valuation studies within industries, looking at how
some valuation measures relate to company characteristics and
external factors over time. My professor suggested using mixed
effects models for such longitudinal data studies, and I don't have
a budget for this sort of thing, so using R and the NLME package was
a natural choice. I think I'm doing some things with it I haven't
seen other analysts do.

I'm not sure I understand what you are doing, but I've often thought
about doing the following: Suppose you estimate regression models
_within_ a homogeneous industry, where you put P/E or P/B on the
l.h.s. and you use a bunch of firm characteristics as explanatory
variables. Would the outliers be places to take a good look for a
profit opportunity? (Is this what you have in mind?)

The problem with this (AFAICT) is that the cross section of accounting
ratios / data tends to be pretty nasty in terms of
distributions. You'll always have a few weird observations which drive
the result. R might be particularly good at this, by virtue of
bringing a variety of statistical and graphical tools to bear on weird
observations, non-normal distributions, etc.

All this is just guesswork, I haven't actually done it. If you have,
do show us examples?

Ajay Shah                                                   Consultant
ajayshah@mayin.org                      Department of Economic Affairs
http://www.mayin.org/ajayshah           Ministry of Finance, New Delhi

Ajay,
Yes, that's part of what I've been working on lately. 

I cover companies in the industrial and materials
sector. I can't do research just trying to apply the
major academic studies done on large universes, by
increasing my loadings on SMB, HML, earnings
surprises, and putting negative loadings on accruals
(though it's good to keep in mind). That kind of
strategy doesn't work so well within a universe of 30
companies and a 1 year measurement horizon. 

I mostly work on fundamental analysis, and DCF models,
but wanted to supplement that with value-added
relative value analysis. The typical sell-side "based
on average p/e blah blah" is quite weak, so I looked
at what Aswath Damodaran suggested in his valuation
book. He had some basic ideas of performing a
regression on one industry at one point in time,
regressing something like p/e on expected growth and
beta, for example, or ev/ebitda on some other factors.
But then he showed how the results vary widely year to
year, performing separate regressions each year, and
kind of just threw up his hands. 

Fortunately, I've got the Compustat database at work,
and can pull industry data like that for many points
in time, so I can create a longitudinal set of data
for an industry. I first tried piling multiple years
and companies into one big pile and performing a
regression on it, but I knew that would be very bad. I
asked my former NYU professor of regression for advice
on how to tackle such an analysis. He suggested using
mixed-effects models, such as nlme in R, and after
researching it and buying the Bates/Pinheiro book, and
performing some analyses, it definitely makes more
sense using this approach. Then I tie the
relationships that I find into my relative value
models using my fundemental forecasts (stuff like
growth, margins, expected leverage, etc.) as
predicting factors.

I don't expect a tight fit, but just to help make
better informed, more objective valuation estimates,
which tend to be required of a sell-side analyst. 

I definitely have to transform variables, do variable
weightings, look for outliers, check for
autocorrelations, etc. The RCMDR and NLME packages
make these things reasonably easy to do, and the best
thing about R is that I have the code for my study
after its done, so if I have second thoughts, I can go
back in fairly easily.

I've been doing valuation studies within

industries, looking at how

some valuation measures relate to company

characteristics and

external factors over time. My professor suggested

using mixed

effects models for such longitudinal data studies,

and I don't have

a budget for this sort of thing, so using R and

the NLME package was

a natural choice. I think I'm doing some things

with it I haven't

seen other analysts do.

I'm not sure I understand what you are doing, but
I've often thought
about doing the following: Suppose you estimate
regression models
_within_ a homogeneous industry, where you put P/E
or P/B on the
l.h.s. and you use a bunch of firm characteristics
as explanatory
variables. Would the outliers be places to take a
good look for a
profit opportunity? (Is this what you have in mind?)

The problem with this (AFAICT) is that the cross
section of accounting
ratios / data tends to be pretty nasty in terms of
distributions. You'll always have a few weird
observations which drive
the result. R might be particularly good at this, by
virtue of
bringing a variety of statistical and graphical
tools to bear on weird
observations, non-normal distributions, etc.

All this is just guesswork, I haven't actually done
it. If you have,
do show us examples?

-- 
Ajay Shah                                           
       Consultant
ajayshah@mayin.org                      Department
of Economic Affairs
http://www.mayin.org/ajayshah           Ministry of
Finance, New Delhi

_______________________________________________
R-sig-finance@stat.math.ethz.ch mailing list

https://www.stat.math.ethz.ch/mailman/listinfo/r-sig-finance