r - Day to day rolling correlations by matching name -

let have data frame below. (the data set have not small this.)

library(lubridate)  x <- data.frame(   date = c(rep(ymd(20160601), 4), rep(ymd(20160602), 3), rep(ymd(20160603), 3)),   name = c("a", "b", "c", "d", "a", "b", "c", "b", "c", "d"),   observation = sample(1:10) )  #          date name observation # 1  2016-06-01             10 # 2  2016-06-01    b           7 # 3  2016-06-01    c           3 # 4  2016-06-01    d           2 # 5  2016-06-02              8 # 6  2016-06-02    b           6 # 7  2016-06-02    c           4 # 8  2016-06-03    b           5 # 9  2016-06-03    c           1 # 10 2016-06-03    d           9 

i want find day day correlation of observations of matching names, i.e., date 2016-06-02, want find correlation between <8, 6, 4> , <10, 7, 3> because a, b, , c common in both 2016-06-02 , 2016-06-01. can such (there better ways this):

filter(x, date %in% ymd(20160601)) %>%   left_join(filter(x, date %in% ymd(20160602)), = "name") %>%   transmute(     date = ymd(20160602),     correlation = cor(observation.x, observation.y, use = "complete.obs")) %>%   `[`(1, )  #         date correlation # 1 2016-06-02   0.9966159 

but how do entire data frame using window functions, data frame consists of dates , correlation previous date? prefer dplyr/rcpproll solution!

dplyr doesn't have rolling merges. assuming need 1 (not clear op, since sample data doesn't have holes), can do:

library(data.table) dt = as.data.table(x) # or setdt convert in place  dt[, date := as.date(date)] # not clear op if have dates or datetimes                             # let's make sure it's dates  dt[.(name = name, old.date = date - 1, obs = observation),      on = c(name = 'name', date = 'old.date'), roll = t][    , cor(obs, observation, use = 'pairwise.complete.obs'), = date] #         date         v1 #1: 2016-06-01         na #2: 2016-06-02  0.9966159 #3: 2016-06-03 -0.5000000