[R] do I need plyr, apply or something else?

[Russell Bowdrey]

Michael, Mikhail

Many thanks for your helpful comments. My faith in community support continues to grow.

Michael: I'm looking to use some sort of flexible spline-like fit (smooth.spline, lowess etc).


Many thanks for sharing your expertise. I actually cross posted this on to the "manipulatr" google group, here is the response from Peter Meilstrup:

" For (1) you might want to take a look at rollapply() and related functions in the zoo package.
for (2), don't put the different samples of your curve fit into different columns. Instead imagine generating a data frame with three columns:

bae.date (date each your fit is based around) 
prediction.date (date you are extrapolating to)
preciction (the fitted value)

so if you have 100 dates, and generate a 7 point curve from each date, you end up with 700 rows."

As ever time pressures kind of dictate that I start from what I know. I've only pretty basic database skills at the moment, so will try zoo/TTR first and try PostgreSQL if that isn't satisfactory.

-----Original Message-----
From: Mikhail Titov [mailto:mlt at gmx.us] 
Sent: 12 July 2012 00:22
To: R. Michael Weylandt
Cc: Russell Bowdrey; r-help at r-project.org
Subject: Re: do I need plyr, apply or something else?

"R. Michael Weylandt" <michael.weylandt at gmail.com> writes:

> On Wed, Jul 11, 2012 at 10:05 AM, Russell Bowdrey 
> <Russell.Bowdrey at justretirement.com> wrote:
>>
>> Dear all,
>>
>> This is what I'd like to do (I have an implementation using for 
>> loops, which I designed before I realised just how slow R is at 
>> executing them - this process currently takes days to run).
>>
>> I have a large dataframe containing corporate bond data, columns are:
>> BondID
>> Date (goes back 5years)
>> Var1
>> Var2
>> Term2Maturity
>>
>> What I want to do is this:
>>
>> 1)      For each bond, at each given date, look back over 1 year and append some statistics to each row ( sd(Var1), cor(Var1,Var2) over that year etc)
>>
>
> Look at the TTR package and the various run** functions. Much faster.
>
>> a.  It seems I might be able to use ddply for this, but I can't work 
>> out how to code the stats function to only look back over one year, 
>> rather than the full data range
>>
>> b.      For example: dfBondsWithCorr<-ddply(dfBonds, .(BondID), transform,corr=cor(Var1,Var2),.progress="text")
>> returns a dataframe where for each bond it has same corr for each 
>> date
>>
>> 2) On each date, subset dfBondsWithCorr by certain qualification 
>> criteria, then to the qualifiers fit a regression through a Var1 and 
>> Term2Maturity, output the regression as a df of curves (say for each 
>> date, a curve represented by points every 0.5 years)
>>
>> a.  I can do this pretty efficiently for a single date (and I suppose 
>> I could wrap that in a function) , but can't quite see how to do the 
>> filtering and spitting out of curves over multiple dates without 
>> using for loops
>>
>
> This ones harder. For simple linear regressions, you can solve the 
> regression analytically (e.g., slope = runCov / runVar and mean
> similarly) but doing it for more complicated regressions will pretty 
> much require a for loop of one sort or another. Can you say what sort 
> of model you are looking to use?
>
>> Would appreciate any thoughts, many thanks in advance

I feel like PostgreSQL will do the work better. It has support for basic statistics [1] and you can use window functions [2] to limit the scope for last year only. Then you get your data with RODBC or something.

I suspect you have you data in some sort of DB in the first place. Perhaps it has similar features.

[1] http://www.postgresql.org/docs/9.1/static/functions-aggregate.html#FUNCTIONS-AGGREGATE-STATISTICS-TABLE
[2] http://www.postgresql.org/docs/9.1/interactive/sql-expressions.html#SYNTAX-WINDOW-FUNCTIONS

--
Mikhail


This email and any attachments are confidential and inte...{{dropped:29}}