R 需要具有从开始到停止索引的快速滚动应用功能
下面是一段代码。它给出了滚动15分钟(历史)窗口的贸易价格水平百分比。如果长度为500或1000,它运行得很快,但正如您所看到的,有45K个观测值,而对于整个数据,它的速度非常慢。我可以应用任何plyr功能吗?欢迎提出任何其他建议 贸易数据是这样的:R 需要具有从开始到停止索引的快速滚动应用功能,r,performance,plyr,R,Performance,Plyr,下面是一段代码。它给出了滚动15分钟(历史)窗口的贸易价格水平百分比。如果长度为500或1000,它运行得很快,但正如您所看到的,有45K个观测值,而对于整个数据,它的速度非常慢。我可以应用任何plyr功能吗?欢迎提出任何其他建议 贸易数据是这样的: > str(trade) 'data.frame': 45571 obs. of 5 variables: $ time : chr "2013-10-20 22:00:00.489" "2013-10-20 22:00:00
> str(trade)
'data.frame': 45571 obs. of 5 variables:
$ time : chr "2013-10-20 22:00:00.489" "2013-10-20 22:00:00.807" "2013-10-20 22:00:00.811" "2013-10-20 22:00:00.811" ...
$ prc : num 121 121 121 121 121 ...
$ siz : int 1 4 1 2 3 3 2 2 3 4 ...
$ aggress : chr "B" "B" "B" "B" ...
$ time.pos: POSIXlt, format: "2013-10-20 22:00:00.489" "2013-10-20 22:00:00.807" "2013-10-20 22:00:00.811" "2013-10-20 22:00:00.811" ...
这就是新列trade$time.pos之后数据的样子
trade$time.pos <- strptime(trade$time, format="%Y-%m-%d %H:%M:%OS")
> head(trade)
time prc siz aggress time.pos
1 2013-10-20 22:00:00.489 121.3672 1 B 2013-10-20 22:00:00.489
2 2013-10-20 22:00:00.807 121.3750 4 B 2013-10-20 22:00:00.807
3 2013-10-20 22:00:00.811 121.3750 1 B 2013-10-20 22:00:00.811
4 2013-10-20 22:00:00.811 121.3750 2 B 2013-10-20 22:00:00.811
5 2013-10-20 22:00:00.811 121.3750 3 B 2013-10-20 22:00:00.811
6 2013-10-20 22:00:00.811 121.3750 3 B 2013-10-20 22:00:00.811
#t_15_index function returns the indices of the trades that were executed in last 15 minutes from the current trade(t-15 to t).
t_15_index <- function(data_vector,index) {
which(data_vector[index] - data_vector[1:index]<=15*60)
}
get_percentile <- function(data) {
len_d <- dim(trade)[1]
price_percentile = vector(length=len_d)
for(i in 1: len_d) {
t_15 = t_15_index(trade$time.pos,i)
#ecdf(rep(..)) gets the empirical distribution of the the trade size on a particular trade-price level
price_dist = ecdf(rep(trade$prc[t_15],trade$siz[t_15]))
#percentile of the current price level depending on current (t-15 to t) subset of data
price_percentile[i] = price_dist(trade$prc[i])
}
trade$price_percentile = price_percentile
trade
}
res_trade = get_percentile(trade)
trade$time.pos头(交易)
时间prc siz侵略时间.pos
1 2013-10-20 22:00:00.489 121.3672 1b 2013-10-20 22:00:00.489
2 2013-10-20 22:00:00.807121.37504 B 2013-10-20 22:00:00.807
2013-10-20 22:00:00.811121.3750 1b 2013-10-20 22:00:00.811
4 2013-10-20 22:00:00.811121.3750 2b 2013-10-20 22:00:00.811
5 2013-10-20 22:00:00.811121.3750 B 2013-10-20 22:00:00.811
6 2013-10-20 22:00:00.811121.3750 B 2013-10-20 22:00:00.811
#t_15_index函数返回当前交易(t-15到t)最后15分钟内执行的交易的指数。
t_15_index可能有一种加速滚动应用程序的方法,但是由于窗口大小的变化,我认为标准工具(例如rollappy
)不起作用,尽管可能一些更熟悉它们的人会有想法。同时,您可以优化百分比计算。您可以直接计算一个合适的近似值,而不是使用ecdf
,它创建一个具有所有相关开销的函数:
> vec <- rnorm(10000, 0, 3)
> val <- 5
> max(which(sort(vec) < val)) / length(vec)
[1] 0.9543
> ecdf(vec)(val)
[1] 0.9543
> microbenchmark(max(which(sort(vec) < val)) / length(vec))
Unit: milliseconds
expr min lq median uq max neval
max(which(sort(vec) < val))/length(vec) 1.093434 1.105231 1.116364 1.141204 1.449141 100
> microbenchmark(ecdf(vec)(val))
Unit: milliseconds
expr min lq median uq max neval
ecdf(vec)(val) 2.552946 2.808041 3.043579 3.439269 4.208202 100
>vec val max(其中(排序(vec)ecdf(vec)(val)
[1] 0.9543
>微基准(最大值(排序(vec)微基准(ecdf(vec)(val))
单位:毫秒
expr最小lq中值uq最大neval
ecdf(vec)(val)2.552946 2.808041 3.043579 3.439269 4.208202 100
大约提高了2.5倍。小样本的改进更大。好吧,这个问题让我感兴趣。以下是我所做的事情:
用自定义百分位数计算替换ecdf
将time.pos更改为数字(因为它无论如何都是以秒为单位的),因为与[.POSIXct
vs[
将t_15_index
更改为只回溯到上一个最早的时间戳,因为数据已排序,因此我们不需要回溯到索引1
这就是结果:
> system.time(res2 <- get_percentile2(trade))
user system elapsed
14.458 0.768 15.215
> system.time(res1 <- get_percentile(trade))
user system elapsed
110.851 17.974 128.736
大约8.5倍的改进。请注意,如果每15分钟间隔的项目较少,则此改进会更大。这将在24小时内填满45K点。因此,如果您的45K点实际超过几天,则您可以期待更多改进。以下是代码:
t_15_index2 <- function(data_vector,index, min.index) {
which(data_vector[index] - data_vector[min.index:index]<=minutes*60) + min.index - 1L
}
get_percentile2 <- function(trade) {
len_d <- dim(trade)[1]
price_percentile = vector(length=len_d)
min.index <- 1
for(i in 1: len_d) {
t_15 = t_15_index2(trade$time.pos.2,i, min.index)
vec <- rep(trade$prc[t_15],trade$siz[t_15])
price_percentile[i] <- max(0, which(sort(vec) <= trade$prc[i])) / length(vec)
min.index <- t_15[[1]]
}
trade$price_percentile = price_percentile
trade
}
最后,如果你这样做的话,你也可以想出聪明的方法来重新计算百分位数,因为你知道你的15分钟包含了什么,增加了什么,删除了什么
无法100%确定执行FIFO 15分钟窗口所需的簿记是否最终会克服执行FIFO 15分钟窗口所带来的好处。这里有一个快速尝试,可以更有效地查找15分钟前发生的时间索引:
# Create some sample data (modified from BrodieG)
set.seed(1)
ticks <- 45000
start <- as.numeric(as.POSIXct("2013-01-01"))
end <- as.numeric(as.POSIXct("2013-01-02"))
times <- as.POSIXct(runif(ticks, start, end), origin=as.POSIXct("1970-01-01"))
trade <- data.frame(
time = sort(times),
prc = 100 + rnorm(ticks, 0, 5),
siz = sample(1:10, ticks, rep = T)
)
# For vector of times, find the index of the first time that was at least
# fifteen minutes before the current time. Assumes times are sorted
minutes_ago <- function(time, minutes = 15) {
secs <- minutes * 60
time <- as.numeric(time)
out <- integer(length(time))
before <- 1
for(i in seq_along(out)) {
while(time[before] < time[i] - secs) {
before <- before + 1
}
out[i] <- before
}
out
}
system.time(minutes_ago(trade$time))
# Takes about 0.2s on my machine
library(Rcpp)
cppFunction("IntegerVector minutes_ago2(NumericVector time, int minutes = 15) {
int secs = minutes * 60;
int n = time.size();
IntegerVector out(n);
int before = 0;
for (int i = 0; i < n; ++i) {
# Could do even better here with a binary search
while(time[before] < time[i] - secs) {
before++;
}
out[i] = before + 1;
}
return out;
}")
system.time(minutes_ago2(trade$time, 10))
# Takes less than < 0.001
all.equal(minutes_ago(trade$time, 15), minutes_ago2(trade$time, 15))
#创建一些示例数据(从BrodieG修改)
种子(1)
ticks如果您想在dplyr中使用ecdf,请在mutate中使用seq_-along/sapply以获得更快的结果
y <- x %>% mutate(percentile.score = sapply(seq_along(score), function(i){sum(score[1:i] <= score[i])/i}))
y%变异(percentile.score=sappy)(seq_沿着(score),函数(i){sum(score[1:i]get_percentile
接受一个输入,data
,并且在函数中不引用数据
。如果您有时间学习一些东西,您可以查看数据表
包,它实现了一个带有参数roll
的滚动联接,我很确定这将比usi快数倍ngplyr
这个答案()如果你想尝试Rcpp,你应该想办法。在C++中,你必须找出代码> ECDF < /Cord>。在你正确地整理数据之后,编译的函数可能会在毫秒内工作。如果有人在寻找一个快速的函数来查找一个区间内的点,请检查<代码> FiffeldIs/COD>。如果您在一次扫描中计算了所有t_15
值,则会做得更好-假设它已排序,您应该只需要通过data@hadley,您的意思是存储所有可能的15分钟滚动间隔(我估计大约有20毫米)?然后才对它们进行操作?现在,原始方法和我的调整都会扫描数据一次。请看我的答案。您在t_15_索引中扫描了多次(虽然它不是整个向量,但效率仍然相当低)。我明白您的意思。t_15索引构建步骤大约需要1.3秒(只需运行t_15_index2
),因此在非Rcpp版本中可以节省约1秒。不幸的是,与17秒的总运行时间相比,这是很小的(这与我现在使用的机器有关,与我编写原始代码时使用的机器不同).+1用于教育。@BrodieG我认为对于真正快速的代码,您可以使用相同的策略来计算滚动ecdf,但这将非常复杂
t_15_index <- function(data_vector,index) {
which(data_vector[index] - data_vector[1:index]<=minutes*60)
}
get_percentile <- function(trade) {
len_d <- dim(trade)[1]
price_percentile = vector(length=len_d)
for(i in 1: len_d) {
t_15 = t_15_index(trade$time.pos,i)
price_dist = ecdf(rep(trade$prc[t_15],trade$siz[t_15]))
price_percentile[i] = price_dist(trade$prc[i])
}
trade$price_percentile = price_percentile
trade
}
# Version that pulls whole 2000 entries each time
sub.vec <- numeric(2000)
system.time(r1 <- for(i in seq_len(length(vec) - 2000)) sub.vec <- vec[i:(i+1999)])
# user system elapsed
# 17.507 4.723 22.211
# Version that overwrites 1 value at a time keeping the most recent 2K
sub.vec <- numeric(2001) # need to make this slightly larger because of 2000 %% 2000 == 0
system.time(r2 <- for(i in seq_len(length(vec) - 2000)) sub.vec[[(i %% 2000) + 1]] <- vec[[i]])
# user system elapsed
# 2.642 0.009 2.650
all.equal(r1, tail(r2, -1L))
# [1] TRUE
# Create some sample data (modified from BrodieG)
set.seed(1)
ticks <- 45000
start <- as.numeric(as.POSIXct("2013-01-01"))
end <- as.numeric(as.POSIXct("2013-01-02"))
times <- as.POSIXct(runif(ticks, start, end), origin=as.POSIXct("1970-01-01"))
trade <- data.frame(
time = sort(times),
prc = 100 + rnorm(ticks, 0, 5),
siz = sample(1:10, ticks, rep = T)
)
# For vector of times, find the index of the first time that was at least
# fifteen minutes before the current time. Assumes times are sorted
minutes_ago <- function(time, minutes = 15) {
secs <- minutes * 60
time <- as.numeric(time)
out <- integer(length(time))
before <- 1
for(i in seq_along(out)) {
while(time[before] < time[i] - secs) {
before <- before + 1
}
out[i] <- before
}
out
}
system.time(minutes_ago(trade$time))
# Takes about 0.2s on my machine
library(Rcpp)
cppFunction("IntegerVector minutes_ago2(NumericVector time, int minutes = 15) {
int secs = minutes * 60;
int n = time.size();
IntegerVector out(n);
int before = 0;
for (int i = 0; i < n; ++i) {
# Could do even better here with a binary search
while(time[before] < time[i] - secs) {
before++;
}
out[i] = before + 1;
}
return out;
}")
system.time(minutes_ago2(trade$time, 10))
# Takes less than < 0.001
all.equal(minutes_ago(trade$time, 15), minutes_ago2(trade$time, 15))
y <- x %>% mutate(percentile.score = sapply(seq_along(score), function(i){sum(score[1:i] <= score[i])/i}))