Grid seach on ARIMA model in R - r

Question

I'm trying to make grid search for my ARIMA model working and I need additional help with it.
I have the following data:
head(train)
Date Count
<date> <int>
1 2016-06-15 21
2 2016-06-16 21
3 2016-06-17 12
4 2016-06-18 20
5 2016-06-19 29
6 2016-06-20 30
Train data Date variable ranges from 2016-06-15 to 2019-06-30 with 1111 observations in total
Train data Count variable ranges from min=3 to max=154 with mean=23.83 and sd=13.84.
I was able to define hyper parameters and create 36 ARIMA models with the following code:
#Create ts data
ts_train = xts(train[, -1], order.by = as.POSIXct(train$Date), frequency = 365)
#ARIMA model tune
#tibble helper function
to_tibble <- function(forecast_object){
point_estimate <- forecast_object$mean %>%
as_tsibble() %>%
rename(point_estimate = value,
date = index)
upper <- forecast_object$upper %>%
as_tsibble() %>%
spread(key, value) %>%
rename(date = index,
upper80 = `80%`,
upper95 = `95%`)
lower <- forecast_object$lower %>%
as_tsibble() %>%
spread(key, value) %>%
rename(date = index,
lower80 = `80%`,
lower95 = `95%`)
reduce(list(point_estimate, upper, lower), full_join)
}
#Trend hyper parameters
order_list <- list("p" = seq(0, 2),
"d" = seq(0, 1),
"q" = seq(0, 2)) %>%
cross() %>%
map(lift(c))
#Seasonal hyper parameteres
season_list <- list("P" = seq(0, 2),
"D" = seq(0, 1),
"Q" = seq(0, 2),
"period" = 365) %>%
cross() %>%
map(lift(c))
#Coerce vectors to tibbles
orderdf <- tibble("order" = order_list)
seasondf <- tibble("season" = season_list)
#Create grid of hyper-parameters
hyper_parameters_df <- crossing(orderdf, seasondf)
#Run grid search of ARIMA models
tic <- Sys.time()
models_df <- hyper_parameters_df %>%
mutate(models = map2(.x = order,
.y = season,
~possibly(arima, otherwise = NULL)(x = ts_train,
order = .x, seasonal = .y)))
running_time <- Sys.time() - tic
running_time
#Drop models which couldn't compute ARIMA
final_models = models_df %>% drop_na()
nrows <- nrow(final_models)
And than I get an error when I try to calculate RMSE across my test data with the following code:
final_models <- final_models %>%
mutate(forecast = map(models, ~possibly(forecast, otherwise = NULL)(., h = 183))) %>%
mutate(point_forecast = map(forecast, ~.$`mean`)) %>%
mutate(true_value = rerun(nrows, test)) %>%
mutate(rmse = map2_dbl(point_forecast, true_value,
~sqrt(mean((.x - .y) ** 2))))
I get one error and one warning message:
Error in .x - .y : non-numeric argument to binary operator
In addition: Warning message:
In mean((.x - .y)^2) :
Incompatible methods ("Ops.ts", "Ops.data.frame") for "-"
Can someone please help me with that?
Here is my test data if it's needed to create dummy data:
head(test)
Date Count
<date> <int>
1 2019-07-02 20
2 2019-07-03 28
3 2019-07-04 35
4 2019-07-05 34
5 2019-07-06 60
6 2019-07-07 63
Test data Date variable ranges from 2019-07-01 to 2019-12-31 with 184 observations in total
Train data Count variable ranges from min=6 to max=63 with mean=21.06 and sd=9.89.

Answer 1

The problem is that when you are computing the RMSE you are using time series rather than vectors. So, you have to change the class of both predictions and true values to numeric.
Here is my solution:
# Load libraries
library(fpp2)
library(dplyr)
library(xts)
library(purrr)
library(tidyr)
# Create sample dataset
dates <- seq.Date(as.Date("2019-07-02"), by = "day", length.out = length(WWWusage))
train <- data.frame(Date = dates, Count = WWWusage)
# Get test dataset using drift method
test <- forecast::rwf(WWWusage, h = 183, drift = TRUE)$mean
#Create ts data
ts_train = xts(train[, -1], order.by = as.POSIXct(train$Date), frequency = 365)
#ARIMA model tune
#tibble helper function
to_tibble <- function(forecast_object){
point_estimate <- forecast_object$mean %>%
as_tsibble() %>%
rename(point_estimate = value,
date = index)
upper <- forecast_object$upper %>%
as_tsibble() %>%
spread(key, value) %>%
rename(date = index,
upper80 = `80%`,
upper95 = `95%`)
lower <- forecast_object$lower %>%
as_tsibble() %>%
spread(key, value) %>%
rename(date = index,
lower80 = `80%`,
lower95 = `95%`)
reduce(list(point_estimate, upper, lower), full_join)
}
#Trend hyper parameters
order_list <- list("p" = seq(0, 2),
"d" = seq(0, 1),
"q" = seq(0, 2)) %>%
cross() %>%
map(lift(c))
#Seasonal hyper parameteres
season_list <- list("P" = seq(0, 2),
"D" = seq(0, 1),
"Q" = seq(0, 2),
"period" = 365) %>%
cross() %>%
map(lift(c))
#Coerce vectors to tibbles
orderdf <- tibble("order" = order_list)
seasondf <- tibble("season" = season_list)
#Create grid of hyper-parameters
hyper_parameters_df <- crossing(orderdf, seasondf)
#Run grid search of ARIMA models
tic <- Sys.time()
models_df <- hyper_parameters_df %>%
mutate(models =
map2(.x = order,
.y = season,
~possibly(arima, otherwise = NULL)(x = ts_train, order = .x, seasonal = .y)))
running_time <- Sys.time() - tic
running_time
#Drop models which couldn't compute ARIMA
final_models = models_df %>% drop_na()
nrows <- nrow(final_models)
# Estimate RSME for each candidate
# Note: you have to make sure that both .x and .y are numeric
final_models2 <- final_models %>%
mutate(forecast = map(models, ~possibly(forecast, otherwise = NULL)(., h = 183))) %>%
mutate(point_forecast = map(forecast, ~.$`mean`)) %>%
mutate(true_value = rerun(nrows, test)) %>%
mutate(rmse = map2_dbl(point_forecast, true_value,
~sqrt(mean((as.numeric(.x) - as.numeric(.y)) ** 2))))