roll_mean: Calculate a _matrix_ of mean (location) estimates over a...

In algoquant/HighFreq: High Frequency Time Series Management

Calculate a matrix of mean (location) estimates over a rolling look-back interval attached at the end points of a time series or a matrix.

Description

Calculate a matrix of mean (location) estimates over a rolling look-back interval attached at the end points of a time series or a matrix.

Usage

roll_mean(
  tseries,
  lookb = 1L,
  startp = 0L,
  endd = 0L,
  step = 1L,
  stub = 0L,
  method = "moment",
  confl = 0.75
)

Arguments

lookb	The number of end points in the look-back interval (the default is lookb = 1).
tseries	A time series or a matrix of data.
startp	An integer vector of start points (the default is startp = 0).
endd	An integer vector of end points (the default is endd = 0).
step	The number of time periods between the end points (the default is step = 1).
stub	An integer value equal to the first end point for calculating the end points (the default is stub = 0).
method	A character string representing the type of mean measure of (the default is method = "moment").

Details

The function roll_mean() calculates a matrix of mean (location) estimates over rolling look-back intervals attached at the end points of the time series tseries.

The function roll_mean() performs a loop over the end points, and at each end point it subsets the time series tseries over a look-back interval equal to lookb number of end points.

It passes the subset time series to the function calc_mean(), which calculates the mean (location). See the function calc_mean() for a description of the mean methods.

If the arguments endd and startp are not given then it first calculates a vector of end points separated by step time periods. It calculates the end points along the rows of tseries using the function calc_endpoints(), with the number of time periods between the end points equal to step time periods.

For example, the rolling mean at 25 day end points, with a 75 day look-back, can be calculated using the parameters step = 25 and lookb = 3.

The function roll_mean() with the parameter step = 1 performs the same calculation as the function roll_mean() from package RcppRoll, but it's several times faster because it uses C++ RcppArmadillo code.

The function roll_mean() is implemented in RcppArmadillo RcppArmadillo C++ code, which makes it several times faster than R code.

If only a simple rolling mean is required (not the median) then other functions like roll_sum() or roll_vec() may be even faster.

Value

A matrix of mean (location) estimates with the same number of columns as the input time series tseries, and the number of rows equal to the number of end points.

Examples

## Not run: 
# Define time series of returns using package rutils
retp <- na.omit(rutils::etfenv$returns$VTI)
# Calculate the rolling means at 25 day end points, with a 75 day look-back
meanv <- HighFreq::roll_mean(retp, lookb=3, step=25)
# Compare the mean estimates over 11-period look-back intervals
all.equal(HighFreq::roll_mean(retp, lookb=11)[-(1:10), ], 
  drop(RcppRoll::roll_mean(retp, n=11)), check.attributes=FALSE)
# Define end points and start points
endd <- HighFreq::calc_endpoints(NROW(retp), step=25)
startp <- HighFreq::calc_startpoints(endd, lookb=3)
# Calculate the rolling means using RcppArmadillo
meanv <- HighFreq::roll_mean(retp, startp=startp, endd=endd)
# Calculate the rolling medians using RcppArmadillo
medianscpp <- HighFreq::roll_mean(retp, startp=startp, endd=endd, method="nonparametric")
# Calculate the rolling medians using R
medians = sapply(1:NROW(endd), function(i) {
  median(retp[startp[i]:endd[i] + 1])
})  # end sapply
all.equal(medians, drop(medianscpp))
# Compare the speed of RcppArmadillo with R code
library(microbenchmark)
summary(microbenchmark(
  Rcpp=HighFreq::roll_mean(retp, startp=startp, endd=endd, method="nonparametric"),
  Rcode=sapply(1:NROW(endd), function(i) {median(retp[startp[i]:endd[i] + 1])}),
  times=10))[, c(1, 4, 5)]

## End(Not run)

algoquant/HighFreq documentation built on Oct. 26, 2024, 9:20 p.m.