JustGitting / tsibble

Tidy Temporal Data Frames and Tools

tsibble

/ˈt͡sɪbəl/

The tsibble package provides a data class of tbl_ts to store and manage temporal-context data frames in a tidy manner. A tsibble consists of a time index, keys and other measured variables in a data-centric format, which is built on top of the tibble.

Installation

You could install the stable version on CRAN:

install.packages("tsibble")

You could install the development version from Github using

# install.packages("devtools")
devtools::install_github("tidyverts/tsibble", build_vignettes = TRUE)

Get started

Coerce to a tsibble with as_tsibble()

The weather data included in the package nycflights13 is used as an example to illustrate. The “index” variable is the time_hour containing the date-times, and the “key” is the origin as weather stations created via id(). The key(s) together with the index uniquely identifies each observation, which gives a valid tsibble. Other columns can be considered as measured variables.

library(tsibble)
weather <- nycflights13::weather %>% 
  select(origin, time_hour, temp, humid, precip)
weather_tsbl <- as_tsibble(weather, key = id(origin), index = time_hour)
weather_tsbl
#> # A tsibble: 26,130 x 5 [1HOUR]
#> # Keys:      origin [3]
#>   origin time_hour            temp humid precip
#>   <chr>  <dttm>              <dbl> <dbl>  <dbl>
#> 1 EWR    2013-01-01 00:00:00  37.0  54.0     0.
#> 2 EWR    2013-01-01 01:00:00  37.0  54.0     0.
#> 3 EWR    2013-01-01 02:00:00  37.9  52.1     0.
#> 4 EWR    2013-01-01 03:00:00  37.9  54.5     0.
#> 5 EWR    2013-01-01 04:00:00  37.9  57.0     0.
#> # ... with 2.612e+04 more rows

The key is not constrained to a single variable, but expressive of nested and crossed data structures. This incorporates univariate, multivariate, hierarchical and grouped time series into the tsibble framework. See ?tsibble and vignette("intro-tsibble") for details.

fill_na() to turn implicit missing values into explicit missing values

Often there are implicit missing cases in temporal data. If the observations are made at regular time interval, we could turn these implicit missings to be explicit simply using fill_na(). Meanwhile, fill NAs in by 0 for precipitation (precip). It is quite common to replaces NAs with its previous observation for each origin in time series analysis, which is easily done using fill() from tidyr.

full_weather <- weather_tsbl %>%
  fill_na(precip = 0) %>% 
  group_by(origin) %>% 
  fill(temp, humid, .direction = "down")
full_weather
#> # A tsibble: 26,208 x 5 [1HOUR]
#> # Keys:      origin [3]
#> # Groups:    origin [3]
#>   origin time_hour            temp humid precip
#>   <chr>  <dttm>              <dbl> <dbl>  <dbl>
#> 1 EWR    2013-01-01 00:00:00  37.0  54.0     0.
#> 2 EWR    2013-01-01 01:00:00  37.0  54.0     0.
#> 3 EWR    2013-01-01 02:00:00  37.9  52.1     0.
#> 4 EWR    2013-01-01 03:00:00  37.9  54.5     0.
#> 5 EWR    2013-01-01 04:00:00  37.9  57.0     0.
#> # ... with 2.62e+04 more rows

fill_na() also handles filling NA by values or functions, and preserves time zones for date-times.

tsummarise() to summarise over calendar periods

tsummarise() and its scoped variants (including _all(), _at(), _if()) are introduced to aggregate interested variables over calendar periods. tsummarise() goes hand in hand with the index functions including as.Date(), yearmonth(), and yearquarter(), as well as other friends from lubridate, such as year(), ceiling_date(), floor_date() and round_date(). For example, it would be of interest in computing average temperature and total precipitation per month, by applying yearmonth() to the hourly time index.

full_weather %>%
  group_by(origin) %>%
  tsummarise(
    year_month = yearmonth(time_hour), # monthly aggregates
    avg_temp = mean(temp, na.rm = TRUE),
    ttl_precip = sum(precip, na.rm = TRUE)
  )
#> # A tsibble: 36 x 4 [1MONTH]
#> # Keys:      origin [3]
#>   origin year_month avg_temp ttl_precip
#>   <chr>       <mth>    <dbl>      <dbl>
#> 1 EWR      2013 Jan     35.6       2.70
#> 2 EWR      2013 Feb     34.1       2.76
#> 3 EWR      2013 Mar     40.0       1.94
#> 4 EWR      2013 Apr     52.9       1.05
#> 5 EWR      2013 May     63.1       2.76
#> # ... with 31 more rows

tsummarise() can also help with regularising a tsibble of irregular time space.

A family of window functions: slide(), tile(), stretch()

Temporal data often involves moving window calculations. Several functions in the tsibble allow for different variations of moving windows using purrr-like syntax:

• slide(): sliding window with overlapping observations.
• tile(): tiling window without overlapping observations.
• stretch(): fixing an initial window and expanding to include more observations.

For example, a moving average of window size 3 is carried out on hourly temperatures for each group (origin).

full_weather %>% 
  group_by(origin) %>% 
  mutate(temp_ma = slide(temp, ~ mean(., na.rm = TRUE), size = 3))
#> # A tsibble: 26,208 x 6 [1HOUR]
#> # Keys:      origin [3]
#> # Groups:    origin [3]
#>   origin time_hour            temp humid precip temp_ma
#>   <chr>  <dttm>              <dbl> <dbl>  <dbl>   <dbl>
#> 1 EWR    2013-01-01 00:00:00  37.0  54.0     0.    NA  
#> 2 EWR    2013-01-01 01:00:00  37.0  54.0     0.    NA  
#> 3 EWR    2013-01-01 02:00:00  37.9  52.1     0.    37.3
#> 4 EWR    2013-01-01 03:00:00  37.9  54.5     0.    37.6
#> 5 EWR    2013-01-01 04:00:00  37.9  57.0     0.    37.9
#> # ... with 2.62e+04 more rows

Reexported functions from the tidyverse

It can be noticed that the tsibble seamlessly works with dplyr verbs. Use ?tsibble::reexports for a full list of re-exported functions.

• dplyr:
  • arrange(), filter(), slice()
  • mutate()/transmute(), select(), summarise()/summarize() with an additional argument drop = FALSE to drop tbl_ts and coerce to tbl_df
  • rename()
  • *_join()
  • group_by(), ungroup()
  • 🚫 distinct()
• tidyr: fill()
• tibble: glimpse(), as_tibble()/as.tibble()
• rlang: !!, !!!

Related work

• zoo: regular and irregular time series with methods.
• xts: extensible time series.
• tibbletime: time-aware tibbles.
• padr: padding of missing records in time series.