Fix date operations for R 4.3; speed up summarization by ~60%

Extracting date() directly from datetimes vs. converting via as_date()
saves a LOT of time.

Model building actually seems quite fast compared to other summaries,
so doesn't seem necessary to limit the time period just yet.

R/stocks.R

@@ -105,7 +105,7 @@ get_price_data <- function(data) {
   prices_5d <- last_n_days(data, 5)[, list(time, ticker, curr)]
   prices_1m <- last_n_days(data, 30)[, list(time, ticker, curr)]
   # Aggregate by day
-  prices_all <- data[, list(curr = max(curr)), by = list(time = as_date(time), ticker)]
+  prices_all <- data[, list(curr = max(curr)), by = list(time = date(time), ticker)]
 
   peaks <- peaks_by_day(data)
 
@@ -163,7 +163,7 @@ last_n_days <- function(data, n, full_days = FALSE) {
   time_day <- 3600*24
   start_time <- max(data$time) - n * time_day
   if (full_days) {
-    start_time <- as_date(start_time)
+    start_time <- floor_date(start_time)
   }
   data[time >= start_time]
 }

R/summary.R

@@ -75,11 +75,11 @@ summarize_all <- function(data) {
 # Get price highs by day
 high_by_day <- function(data, fill = TRUE) {
   highs <- data[data[, list(.I = .I[which.max(curr)]),
-                     by = list(ticker, date = as_date(time))]$.I]
+                     by = list(ticker, date = date(time))]$.I]
 
   if (fill) {
     # Fill in missing days
-    highs[, date := as_date(time)]
+    highs[, date := date(time)]
     all_dates <- highs[, list(date = seq(min(date), max(date), by = "days")), by = ticker]
     highs <- highs[all_dates, on = list(date, ticker)]
     highs[is.na(curr), time := as_datetime(format(date))]
@@ -148,7 +148,7 @@ summarize_volume_price <- function(data) {
 # Summarize market volume (shares bought) by day
 summarize_volume_day <- function(data) {
   summ <- data[, list(volume = total_volume(volume)),
-               by = list(date = as_date(time), ticker)]
+               by = list(date = date(time), ticker)]
   summ <- summ[, list(volume = sum(volume)), by = date]
   summ
 }

R/utils.R

@@ -5,9 +5,20 @@ as_datetime <- function(x, tz = TZ_NST) {
 }
 
 as_date <- function(x, tz = TZ_NST) {
+  # Note: as.Date ignores tz for character strings
   as.Date(x, tz = tz)
 }
 
+# Get the date of an existing date-time object. Use this over as_date() for
+# date-time objects as date() will be much faster.
+date <- function(x) {
+  lubridate::date(x)
+}
+
+floor_date <- function(x, unit = "day") {
+  lubridate::floor_date(x, unit = unit)
+}
+
 to_iso_string <- function(time, tz = "UTC") {
   strftime(time, tz = tz, format = "%Y-%m-%dT%H:%M:%SZ")
 }

tests/testthat/test-predict.R

@@ -1,6 +1,6 @@
 test_that("predict_prices", {
   data <- read_stock_data(stock_data_example())
-  predicted <- predict_prices(data[time <= as_date("2018-07-15")])
+  predicted <- predict_prices(data[time <= as_datetime("2018-07-15")])
 
   p <- predicted[curr %in% c(6, 10, 15, 30, 60, 90), p]
   expected_p <- c(8.03, 7.18, 4.46, 0.348, 0.176, 0.052)

tests/testthat/test-stocks.R

@@ -211,7 +211,7 @@ test_that("get_price_data", {
   expect_equal(prices[["1d"]], data[1:2, ])
   expect_equal(prices[["5d"]], data[1:3, ])
   expect_equal(prices[["1m"]], data[1:4, ])
-  expect_equal(prices[["all"]], data[, list(time = as_date(time), ticker, curr)])
+  expect_equal(prices[["all"]], data[, list(time = date(time), ticker, curr)])
   expect_equal(prices$peaks, peaks_by_day(data))
 })
 
@@ -316,6 +316,14 @@ test_that("last_n_days", {
     ))
   ))
 
+  expect_equal(last_n_days(data, 3), data.table(
+    time = as_datetime(c(
+      "2018-03-06 8:24:58",
+      "2018-03-07",
+      "2018-03-07 08:24:59"
+    ))
+  ))
+
   expect_equal(last_n_days(data, 3, TRUE), data.table(
     time = as_datetime(c(
       "2018-03-04",

tests/testthat/test-utils.R

@@ -6,10 +6,23 @@ test_that("as_datetime converts time strings", {
 
 test_that("as_date converts date strings", {
   date <- as_date("2018-07-03")
-  expected <- as.Date("2018-07-03", tz = TZ_NST)
+  expected <- as.Date("2018-07-03")
   expect_equal(date, expected)
 })
 
+test_that("date gets the date of date-times", {
+  time <- as_datetime("2018-07-01T17:30:00Z")
+  expect_equal(date(time), as.Date("2018-07-01"))
+  date <- as_date("2018-07-03")
+  expect_equal(date(date), as.Date("2018-07-03"))
+})
+
+test_that("floor_date rounds dates down", {
+  time <- floor_date(as_datetime("2018-07-01T17:30:00Z"))
+  expected <- as_datetime("2018-07-01")
+  expect_equal(time, expected)
+})
+
 test_that("to_iso_string converts datetimes", {
   time <- as_datetime("2018-07-01T17:30:00Z")
   str <- to_iso_string(time)