This vignette shows comparisons in terms of computation time with other packages. Specifically, we will make comparisons with the roll package and zoo package. It should be stressed though that the other solutions do additional things than this package does. E.g., there is not performed any rank test in the function in this package. We start by showing the comparisons of computation times and then we show different options. Some function definitions are shown at the end.
We start by simulating the data
set.seed(32981054)
n <- 10000
p <- 6
wdth = 50
X <- matrix(rnorm(p * n), n, p)
y <- drop(X %*% runif(p)) + rnorm(n)
df <- data.frame(y, X)
frm <- eval(parse(text = paste0(
"formula(y ~ -1 + ", paste0("X", 1:p, collapse = " + "), ")")))Then we check that the functions give the same (the function definitions are at the end of this document)
library(rollRegres)
base_res <- roll_regress_R_for_loop(X, y, wdth)
all.equal(base_res, roll_regres.fit(X, y, wdth)$coefs,
check.attributes = FALSE)
#R [1] TRUE
all.equal(base_res, roll_regres(frm, df, wdth)$coefs,
check.attributes = FALSE)
#R [1] TRUE
all.equal(base_res, roll_regress_zoo(X, y, wdth),
check.attributes = FALSE)
#R [1] TRUE
all.equal(base_res, roll_lm_func(X, y, wdth),
check.attributes = FALSE)
#R [1] TRUEand here we compare the computation time
microbenchmark::microbenchmark(
roll_regress = roll_regres.fit(X, y, wdth),
# this will be slower due to call to `model.matrix` and `model.frame`
roll_regress_df = roll_regres(frm, df, wdth),
roll_regress_zoo = roll_regress_zoo(X, y, wdth),
roll_regress_R_for_loop = roll_regress_R_for_loop(X, y, wdth),
roll_lm = roll_lm_func(X, y, wdth),
times = 5)
#R Unit: milliseconds
#R expr min lq mean median uq
#R roll_regress 4.949844 4.990765 5.116116 5.019353 5.034736
#R roll_regress_df 5.582242 5.638712 6.415709 5.681938 5.840059
#R roll_regress_zoo 513.045797 529.249343 527.956594 531.347242 533.030741
#R roll_regress_R_for_loop 289.162866 298.066299 308.334590 315.472252 316.488637
#R roll_lm 51.067710 51.812972 53.017997 51.846693 53.445712
#R max neval
#R 5.585880 5
#R 9.335593 5
#R 533.109848 5
#R 322.482896 5
#R 56.916899 5
# here is the formula used above
frm
#R y ~ -1 + X1 + X2 + X3 + X4 + X5 + X6This section will cover some additional features.
Here are expanding window regressions with additional output
#####
# simulate data
set.seed(65731482)
n <- 100
p <- 2
X <- matrix(rnorm(p * n), n, p)
y <- drop(X %*% runif(p)) + rnorm(n)
#####
# use package function
pck_out <- roll_regres.fit(
X, y, width = 50L, do_downdates = FALSE,
do_compute = c("sigmas", "r.squareds", "1_step_forecasts"))
#####
# assign R-version
R_func <- function(X, y, width){
n <- nrow(X)
p <- ncol(X)
out <- matrix(NA_real_, n, p)
sigmas <- rep(NA_real_, n)
r.squared <- rep(NA_real_, n)
one_step_forecasts <- rep(NA_real_, n)
for(i in width:n){
idx <- 1:i
fit <- lm(y[idx] ~ -1 + X[idx, , drop = FALSE])
out[i, ] <- fit$coefficients
su <- summary(fit)
sigmas[i] <- su$sigma
ss1 <- sum((y[idx] - mean(y[idx]))^2)
ss2 <- sum(fit$residuals^2)
r.squared[i] <- 1 - ss2 / ss1
if(i < n){
next_i <- i + 1L
one_step_forecasts[next_i] <- fit$coefficients %*% X[next_i, ]
}
}
list(coef = out, sigmas = sigmas, r.squared = r.squared,
one_step_forecasts = one_step_forecasts)
}
R_out <- R_func(X, y, 50L)
#####
# gives the same
stopifnot(
isTRUE(all.equal(R_out$coef , pck_out$coefs)),
isTRUE(all.equal(R_out$sigmas , pck_out$sigmas)),
isTRUE(all.equal(R_out$r.squared , pck_out$r.squared)),
isTRUE(all.equal(R_out$one_step_forecasts, pck_out$one_step_forecasts)))You can use the grp argument to make updates in blocks. E.g., here is an example with weekly data
#####
# simulate data
set.seed(68799379)
week <- as.integer(gl(25, 7))
head(week[1:10])
#R [1] 1 1 1 1 1 1
n <- length(week)
p <- 2
X <- matrix(rnorm(p * n), n, p)
y <- drop(X %*% runif(p)) + rnorm(n)
#####
# use package function
pck_out <- roll_regres.fit(
X, y, width = 10L, grp = week,
do_compute = c("sigmas", "r.squareds", "1_step_forecasts"))
#####
# assign R-version
R_func <- function(X, y, width, grp){
grp <- grp + 1L - min(grp)
u_grp = unique(grp)
n <- nrow(X)
p <- ncol(X)
out <- matrix(NA_real_, n, p)
sigmas <- rep(NA_real_, n)
r.squared <- rep(NA_real_, n)
one_step_forecasts <- rep(NA_real_, n)
start_val <- min(which(u_grp >= width))
for(g in u_grp[start_val:length(u_grp)]){
idx <- which(grp %in% (g - width + 1L):g)
i <- which(grp == g)
fit <- lm(y[idx] ~ -1 + X[idx, , drop = FALSE])
out[i, ] <- sapply(fit$coefficients, rep, times = length(i))
su <- summary(fit)
sigmas[i] <- su$sigma
ss1 <- sum((y[idx] - mean(y[idx]))^2)
ss2 <- sum(fit$residuals^2)
r.squared[i] <- 1 - ss2 / ss1
if(g != max(grp)){
next_g <- grp[min(which(grp > g))]
next_g <- which(grp == next_g)
one_step_forecasts[next_g] <- X[next_g, ] %*% fit$coefficients
}
}
list(coef = out, sigmas = sigmas, r.squared = r.squared,
one_step_forecasts = one_step_forecasts)
}
R_out <- R_func(X, y, 10L, grp = week)
#####
# gives the same
stopifnot(
isTRUE(all.equal(R_out$coef , pck_out$coefs)),
isTRUE(all.equal(R_out$sigmas , pck_out$sigmas)),
isTRUE(all.equal(R_out$r.squared , pck_out$r.squared)),
isTRUE(all.equal(R_out$one_step_forecasts, pck_out$one_step_forecasts)))Suppose that we are performing rolling regression on stock data over a yearly window. Further, suppose that there are some gaps in the data where we do not have data and we require at least 6 months of data. This can be done as follows
#####
# simulate data
set.seed(96235555)
n <- 10L * 12L * 21L # x years w/ 12 months of 21 trading days
month <- (seq_len(n) - 1L) %/% 21L + 1L # group by months
set.seed(29478439)
X <- matrix(rnorm(n * 4L), ncol = 4L)
y <- rnorm(n)
# randomly drop rows
keep <- seq_along(y) %in% sample.int(nrow(X), as.integer(n * .5))
X <- X [keep, ]
y <- y [keep]
month <- month[keep]
#####
# use package function
pck_out <- roll_regres.fit(
X, y, width = 12L, grp = month, min_obs = 21L * 6L, do_downdates = TRUE,
do_compute = c("sigmas", "r.squareds"))
#####
# assign R-version
R_func <- function(X, y, width, grp, downdate, min_obs){
grp <- grp + 1L - min(grp)
u_grp = unique(grp)
n <- nrow(X)
p <- ncol(X)
out <- matrix(NA_real_, n, p)
sigmas <- rep(NA_real_, n)
r.squared <- rep(NA_real_, n)
start_val <- min(which(u_grp >= width))
for(g in u_grp[start_val:length(u_grp)]){
idx <-
if(downdate)
which(grp %in% (g - width + 1L):g) else
which(grp %in% 1:g)
i <- which(grp == g)
if(length(idx) < min_obs)
next
fit <- lm(y[idx] ~ -1 + X[idx, , drop = FALSE])
out[i, ] <- sapply(fit$coefficients, rep, times = length(i))
su <- summary(fit)
sigmas[i] <- su$sigma
ss1 <- sum((y[idx] - mean(y[idx]))^2)
ss2 <- sum(fit$residuals^2)
r.squared[i] <- 1 - ss2 / ss1
}
list(coef = out, sigmas = sigmas, r.squared = r.squared)
}
R_out <- R_func(X, y, width = 12L, downdate = TRUE, grp = month,
min_obs = 21L * 6L)
#####
# gives the same
stopifnot(
isTRUE(all.equal(R_out$coef , pck_out$coefs)),
isTRUE(all.equal(R_out$sigmas , pck_out$sigmas)),
isTRUE(all.equal(R_out$r.squared, pck_out$r.squared)))sessionInfo()
#R R version 3.6.1 (2019-07-05)
#R Platform: x86_64-pc-linux-gnu (64-bit)
#R Running under: Ubuntu 18.04.3 LTS
#R
#R Matrix products: default
#R BLAS: /usr/lib/x86_64-linux-gnu/blas/libblas.so.3.7.1
#R LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.7.1
#R
#R locale:
#R [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C
#R [3] LC_TIME=en_US.UTF-8 LC_COLLATE=C
#R [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
#R [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
#R [9] LC_ADDRESS=C LC_TELEPHONE=C
#R [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
#R
#R attached base packages:
#R [1] compiler stats graphics grDevices utils datasets methods
#R [8] base
#R
#R other attached packages:
#R [1] rollRegres_0.1.3 roll_1.1.3 zoo_1.8-6
#R
#R loaded via a namespace (and not attached):
#R [1] Rcpp_1.0.3 lattice_0.20-38 digest_0.6.22
#R [4] grid_3.6.1 backports_1.1.5 magrittr_1.5
#R [7] evaluate_0.14 RcppParallel_4.4.4 rlang_0.4.1
#R [10] stringi_1.4.3 checkmate_1.9.4 rmarkdown_1.17
#R [13] tools_3.6.1 stringr_1.4.0 xfun_0.11
#R [16] yaml_2.2.0 microbenchmark_1.4-7 htmltools_0.4.0
#R [19] knitr_1.26Here are the function definitions
#####
# simple R version
roll_regress_R_for_loop <- function(X, y, width){
n <- nrow(X)
p <- ncol(X)
out <- matrix(NA_real_, n, p)
for(i in width:n){
idx <- (i - width + 1L):i
out[i, ] <- lm.fit(X[idx, , drop = FALSE], y[idx])$coefficients
}
out
}
#####
# zoo version
.zoo_inner <- function(Z) {
fit <- lm.fit(Z[, -1, drop = FALSE], Z[, 1])
fit$coefficients
}
library(zoo)
roll_regress_zoo <- function(x, y, width){
Z <- cbind(y, X)
rollapply(Z, width, FUN = .zoo_inner,
by.column = FALSE, align = "right", fill = NA_real_)
}
#####
# roll_lm
library(roll)
roll_lm_func <- function(x, y ,width)
roll_lm(X, matrix(y, ncol = 1), wdth, intercept = FALSE)$coefficients
# use one thread as other methods are easily to compute in parallel too
RcppParallel::setThreadOptions(numThreads = 1)
# compile functions
library(compiler)
roll_regress_R_for_loop <- cmpfun(roll_regress_R_for_loop)
.zoo_inner <- cmpfun(.zoo_inner)
roll_regress_zoo <- cmpfun(roll_regress_zoo)
roll_lm_func <- cmpfun(roll_lm_func)