rolling correlation

[wcornwell]

Hi @kevinushey,

Thanks for this package. We needed to calculate a rolling correlation between two variables, so wrote this function using your tools. If you'd like, I can make it a PR, but also happy to just leave it for people to find:

roll_cor <- function(x, y, window_size) {
  n <- length(x)
  
  # Calculate rolling sums
  sum_x <- RcppRoll::roll_sum(x,
                              n = window_size,
                              fill = NA,
                              align = "right")
  sum_y <- RcppRoll::roll_sum(y,
                              n = window_size,
                              fill = NA,
                              align = "right")
  
  # Calculate rolling sums of squares
  sum_x_sq <- RcppRoll::roll_sum(x ^ 2,
                                 n = window_size,
                                 fill = NA,
                                 align = "right")
  sum_y_sq <- RcppRoll::roll_sum(y ^ 2,
                                 n = window_size,
                                 fill = NA,
                                 align = "right")
  
  # Calculate rolling sum of products
  sum_xy <- RcppRoll::roll_sum(x * y,
                               n = window_size,
                               fill = NA,
                               align = "right")
  
  # Calculate means
  mean_x <- sum_x / window_size #could use roll_mean here
  mean_y <- sum_y / window_size #could use roll_mean here
  
  # Calculate covariance
  covariance <- (sum_xy - window_size * mean_x * mean_y) / window_size
  
  # Calculate variances
  variance_x <- (sum_x_sq - window_size * mean_x ^ 2) / window_size
  variance_y <- (sum_y_sq - window_size * mean_y ^ 2) / window_size
  
  # Calculate correlation
  correlation <- covariance / sqrt(variance_x * variance_y)
  
  return(correlation)
}