Functionality for missing values in a time series

[Mullefa]

I am finding RcppRoll very convenient to use in conjunction with dplyr with one caveat: if I am doing rolling summaries over a numeric vector which is indexed by date (or a time period), then I may still want this to be used for calculating the rolling window (with value 0). It is analogous to using OLAP functions in SQL with range.

mutate_over() is my most recent attempt at implementing this functionality. @kevinushey @hadley I am wondering whether something like this will sit in RcppRoll or dplyr in the future?

[Mullefa]

@kevinushey @hadley on a related note, the user may want to expand the range so that it includes all dates within a given period.

For example, whilst mutate_over() (or another range based function) gives the user a way of calculating a rolling metric over a range of dates they have in their data frame, they may want the other values for e.g. visualisation.

To mitigate against this issue, I have created another function - regularise() - which allows the user to create a data frame with a 'full' index of dates . However, over large data frames it is noticeably slower than the RcppRoll and dplyr API's - I was wondering whether you think it would be possible to rewrite it using Rcpp (no meaningful experience with this)?

If this sounds a bit cryptic, hopefully the following example will illustrate my point:

devtools::install_github("Mullefa/dtsr")


library(dtsr)
library(ggvis)


roll_mean <- function(x, n) {
  out <- RcppRoll::roll_mean(x , n)
  c(dplyr::cummean(x[seq_len(n - 1)]), out)
}


# If the date doesn't appear in the data frame, say no sales occurred on that day.
sales_data <- data_frame(
    date = seq(as.Date("2014-01-01"), as.Date("2014-12-31"), by = 1),
    sales = sample(1:1000, length(date), replace = TRUE)
  ) %>%
  sample_n(250) %>%
  ts_df(date) %>%
  arrange


# This graph doesn't visualize the average weekly sales for dates on which there were no sales.
sales_data %>%
  mutate_over(avg_weekly_sales = roll_mean(sales, n = 7)) %>%
  ggvis(~date, ~avg_weekly_sales) %>%
  layer_bars


# Using regularise() followed by mutate(), this graph does.
sales_data %>%
  regularise %>%
  mutate(avg_weekly_sales = roll_mean(sales, n = 7)) %>%
  ggvis(~date, ~avg_weekly_sales) %>%
  layer_bars