[R] dataframe to a timeseries object

[Daniele Amberti]

I found that plyr:::daply is more efficient than base:::by (am I doing something wrong?), below updated code for comparison (I also fixed a couple things).
Function daply from plyr package has also a .parallel argument and I wonder if creating timeseries objects in parallel and then combining them would be faster (Windows XP platform); does someone has experience with this topic? I found only very simple examples about plyr and parallel computations and I do not have a working example for such kind of implementation (daply that return a list of timeseries objects).

Thanks in advance,
Daniele Amberti


set.seed(123)

N <- 10000
X <- data.frame(
  ID = c(rep(1,N), rep(2,N,), rep(3,N), rep(4,N)),
  DATE = as.character(rep(as.POSIXct("2000-01-01", tz = "GMT")+ 0:(N-1), 4)),
  VALUE = runif(N*4), stringsAsFactors = FALSE)
X <- X[sample(1:(N*4), N*4),]
str(X)

library(timeSeries)
buildTimeSeriesFromDataFrame <- function(x, env)
{
  {
    if(exists("xx", envir = env))
      assign("xx",
        cbind(get("xx", env), timeSeries(x$VALUE, x$DATE,
          format = '%Y-%m-%d %H:%M:%S',
          zone = 'GMT', units = as.character(x$ID[1]))),
        envir = env)
    else
      assign("xx",
        timeSeries(x$VALUE, x$DATE, format = '%Y-%m-%d %H:%M:%S',
          zone = 'GMT', units = as.character(x$ID[1])),
        envir = env)

    return(TRUE)
  }
}

tsBy <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- by(X, X$ID, buildTimeSeriesFromDataFrame,
      env = e1, simplify = TRUE)
  return(get("xx", e1))
}

Time01 <- replicate(100,
  system.time(tsBy(X, X$ID, simplify = TRUE))[[1]])
median(Time01)
hist(Time01)
ATS <- tsBy(X, X$ID, simplify = TRUE)


library(xts)
buildXtsFromDataFrame <- function(x, env)
{
  {
    if(exists("xx", envir = env))
      assign("xx",
        cbind(get("xx", env), xts(x$VALUE,
          as.POSIXct(x$DATE, tz = "GMT",
            format = '%Y-%m-%d %H:%M:%S'),
          tzone = 'GMT')),
        envir = env)
    else
      assign("xx",
        xts(x$VALUE, as.POSIXct(x$DATE, tz = "GMT",
            format = '%Y-%m-%d %H:%M:%S'),
          tzone = 'GMT'),
        envir = env)

    return(TRUE)
  }
}

xtsBy <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- by(X, X$ID, buildXtsFromDataFrame,
      env = e1, simplify = TRUE)
  return(get("xx", e1))
}

Time02 <- replicate(100,
  system.time(xtsBy(X, X$ID,simplify = TRUE))[[1]])
median(Time02)
hist(Time02)
AXTS <- xtsBy(X, X$ID, simplify = TRUE)

plot(density(Time02), col = "red",
  xlim = c(min(c(Time02, Time01)), max(c(Time02, Time01))))
lines(density(Time01), col = "blue")
#check equal, a still a problem with names
AXTS2 <- as.timeSeries(AXTS)
names(AXTS2) <- names(ATS)
identical(getDataPart(ATS), getDataPart(AXTS2))
identical(time(ATS), time(AXTS2))

# with plyr library and daply instead of by:
library(plyr)

tsDaply <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- daply(X, "ID", buildTimeSeriesFromDataFrame,
      env = e1)
  return(get("xx", e1))
}

Time03 <- replicate(100,
  system.time(tsDaply(X, X$ID))[[1]])
median(Time03)
hist(Time03)

xtsDaply <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- daply(X, "ID", buildXtsFromDataFrame,
      env = e1)
  return(get("xx", e1))
}

Time04 <- replicate(100,
  system.time(xtsDaply(X, X$ID))[[1]])

median(Time04)
hist(Time04)

plot(density(Time04), col = "red",
  xlim = c(
    min(c(Time02, Time01, Time03, Time04)),
    max(c(Time02, Time01, Time03, Time04))),
  ylim = c(0,100))
lines(density(Time03), col = "blue")
lines(density(Time02))
lines(density(Time01))





-----Original Message-----
From: Daniele Amberti
Sent: 11 March 2011 14:44
To: r-help at r-project.org
Subject: dataframe to a timeseries object

I’m wondering which is the most efficient (time, than memory usage) way to obtain a multivariate time series object from a data frame (the easiest data structure to get data from a database trough RODBC).
I have a starting point using timeSeries or xts library (these libraries can handle time zones), below you can find code to test.
Merging parallelization (cbind) is something I’m thinking at (suggestions from users with experience on this topic is highly appreciated), any suggestion is welcome.
My platform is Windows XP, R 2.12.1, latest available packages on CRAN for timeSeries and xts.


set.seed(123)

N <- 9000
X <- data.frame(
  ID = c(rep(1,N), rep(2,N,), rep(3,N), rep(4,N)),
  DATE = rep(as.POSIXct("2000-01-01", tz = "GMT")+ 0:(N-1), 4),
  VALUE = runif(N*4))

library(timeSeries)
buildTimeSeriesFromDataFrame <- function(x, env)
{
  {
    if(exists("xx", envir = env))
      assign("xx",
        cbind(get("xx", env), timeSeries(x$VALUE, x$DATE, format = '%Y-%m-%d %H:%M:%S',
          zone = 'GMT', units = as.character(x$ID[1]))),
        envir = env)
    else
      assign("xx",
        timeSeries(x$VALUE, x$DATE, format = '%Y-%m-%d %H:%M:%S',
          zone = 'GMT', units = as.character(x$ID[1])),
        envir = env)

    return(TRUE)
  }
}


fooBy <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- by(X, X$ID, buildTimeSeriesFromDataFrame,
      env = e1, simplify = TRUE)
  return(get("xx", e1))
}

Time01 <- replicate(100,
  system.time(fooBy(X,
    X$ID, buildTimeSeriesFromDataFrame,
    simplify = TRUE))[[1]])

median(Time01)
hist(Time01)

library(xts)

buildXtsFromDataFrame <- function(x, env)
{
  {
    if(exists("xx", envir = env))
      assign("xx",
        cbind(get("xx", env), xts(x$VALUE,
          as.POSIXct(x$DATE, format = '%Y-%m-%d %H:%M:%S'),
          tzone = 'GMT')),
        envir = env)
    else
      assign("xx",
        xts(x$VALUE, as.POSIXct(x$DATE, format = '%Y-%m-%d %H:%M:%S'),
          tzone = 'GMT'),
        envir = env)

    return(TRUE)
  }
}

fooBy <- function(...)
{
  e1 <- new.env(parent = baseenv())
  res <- by(X, X$ID, buildXtsFromDataFrame,
      env = e1, simplify = TRUE)
  return(get("xx", e1))
}

Time02 <- replicate(100,
  system.time(fooBy(X,
    X$ID, buildTimeSeriesFromDataFrame,
    simplify = TRUE))[[1]])

median(Time02)
hist(Time02)

plot(density(Time02), xlim = c(min(c(Time02, Time01)), max(c(Time02, Time01))))
lines(density(Time01))


Best regards,
Daniele Amberti

ORS Srl

Via Agostino Morando 1/3 12060 Roddi (Cn) - Italy
Tel. +39 0173 620211
Fax. +39 0173 620299 / +39 0173 433111
Web Site www.ors.it

------------------------------------------------------------------------------------------------------------------------
Qualsiasi utilizzo non autorizzato del presente messaggio e dei suoi allegati è vietato e potrebbe costituire reato.
Se lei avesse ricevuto erroneamente questo messaggio, Le saremmo grati se provvedesse alla distruzione dello stesso
e degli eventuali allegati.
Opinioni, conclusioni o altre informazioni riportate nella e-mail, che non siano relative alle attività e/o
alla missione aziendale di O.R.S. Srl si intendono non  attribuibili alla società stessa, né la impegnano in alcun modo.