Autocorrelation plot for only negative values - r

Question

I would like to do an acf plot in R for only the negative values of a time series. I cannot do this by just subsetting the data for only negative values beforehand, because then the autocorrelation will remove arbitrary number of positive days in between the negative values and be unreasonably high, but rather, I would like to run the autocorrelation on the whole time series and then filter out the results given the first day is negative.
For example, in theory, I could make a data frame with the original series and all of the lagged time series in a data frame, then filter for the negative values in the original series, and then plot the correlations. However, I would like to automate this using the acf plot.
Here is an example of my time series:
> dput(exampleSeries)
c(0, 0, -0.000687, -0.004489, -0.005688, 0.000801, 0.005601,
0.004546, 0.003451, -0.000836, -0.002796, 0.005581, -0.003247,
-0.002416, 0.00122, 0.005337, -0.000195, -0.004255, -0.003097,
0.000751, -0.002037, 0.00837, -0.003965, -0.001786, 0.008497,
0.000693, 0.000824, 0.005681, 0.002274, 0.000773, 0.001141, 0.000652,
0.001559, -0.006201, 0.000479, -0.002041, 0.002757, -0.000736,
-2.1e-05, 0.000904, -0.000319, -0.000227, -0.006589, 0.000998,
0.00171, 0.000271, -0.004121, -0.002788, -9e-04, 0.001639, 0.004245,
-0.00267, -0.004738, 0.001192, 0.002175, 0.004666, 0.006005,
0.001218, -0.003188, -0.004363, 0.000462, -0.002241, -0.004806,
0.000463, 0.000795, -0.005715, 0.004635, -0.004286, -0.008908,
-0.001044, -0.000842, -0.00445, -0.006094, -0.001846, 0.005013,
-0.006599, 0.001914, 0.00221, 6.2e-05, -0.001391, 0.004369, -0.005739,
-0.003467, -0.002103, -0.000882, 0.001483, 0.003074, 0.00165,
-0.00035, -0.000573, -0.00316, -0.00102, -0.00144, 0.003421,
0.005436, 0.001994, 0.00619, 0.005319, 7.3e-05, 0.004513)

Answer 1

I tried to implement your description.
correl <- function(x, lag.max = 10){
library(dplyr)
m <- matrix(ncol = lag.max, nrow = length(x))
for(i in 1:lag.max){
m[,i] <- lag(x, i)
}
m <- m[x<0,]
res <- apply(m, 2, function(y) cor(y, x[x<0], use = "complete.obs"))
barplot(res)
}
correl(exampleSeries)

Answer 2

Maybe just write your own function? Something like:
negativeACF <- function(x, num.lags = 10)
{
n <- length(x)
acfs <- sapply(0:num.lags, function(i) cor(x[-i:-1], x[(-n-1+i):-n]))
names(acfs) <- 0:num.lags
acfs[acfs < 0]
}
results <- negativeACF(exampleSeries, num.lags=20)
barplot(results)

Answer 3

Yea I ended up writing my own functions and just replacing the values in the R acf object with my own values that are just the correlations. So:
genACF <- function(series, my.acf, lag.max = NULL, neg){
x <- na.fail(as.ts(series))
x.freq <- frequency(x)
x <- as.matrix(x)
if (!is.numeric(x))
stop("'x' must be numeric")
sampleT <- as.integer(nrow(x))
nser <- as.integer(ncol(x))
if (is.null(lag.max))
lag.max <- floor(10 * (log10(sampleT) - log10(nser)))
lag.max <- as.integer(min(lag.max, sampleT - 1L))
if (is.na(lag.max) || lag.max < 0)
stop("'lag.max' must be at least 0")
if(neg){
indices <- which(series < 0)
}else{
indices <- which(series > 0)
}
series <- scale(series, scale = FALSE)
series.zoo <- zoo(series)
for(i in 0:lag.max){
lag.series <- lag(series.zoo, k = -i, na.pad = TRUE)
temp.corr <- cor(series.zoo[indices], lag.series[indices], use = 'complete.obs', method = 'pearson')
my.acf[i+1] <- temp.corr
}
my.acf[1] <- 0
return(my.acf)
}
plotMyACF <- function(series, main, type = 'correlation', neg = TRUE){
series.acf <- acf(series, plot = FALSE)
my.acf <- genACF(series, series.acf$acf, neg = neg)
series.acf$acf <- my.acf
plot(series.acf, xlim = c(1, dim(series.acf$acf)[1] - (type == 'correlation')), xaxt = "n", main = main)
if (dim(series.acf$acf)[1] < 25){
axis(1, at = 1:(dim(series.acf$acf)[1] - 1))
}else{
axis(1)
}
}
And I get something like this: