healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse
Packages Steven P. Sanderson II, MPH - Date: 2026-01-07
Introduction
This analysis follows a Nested Modeltime Workflow from modeltime
along with using the NNS package. I use this to monitor the
downloads of all of my packages:
Get Data
glimpse(downloads_tbl)
Rows: 165,021
Columns: 11
$ date <date> 2020-11-23, 2020-11-23, 2020-11-23, 2020-11-23, 2020-11-23,…
$ time <Period> 15H 36M 55S, 11H 26M 39S, 23H 34M 44S, 18H 39M 32S, 9H 0M…
$ date_time <dttm> 2020-11-23 15:36:55, 2020-11-23 11:26:39, 2020-11-23 23:34:…
$ size <int> 4858294, 4858294, 4858301, 4858295, 361, 4863722, 4864794, 4…
$ r_version <chr> NA, "4.0.3", "3.5.3", "3.5.2", NA, NA, NA, NA, NA, NA, NA, N…
$ r_arch <chr> NA, "x86_64", "x86_64", "x86_64", NA, NA, NA, NA, NA, NA, NA…
$ r_os <chr> NA, "mingw32", "mingw32", "linux-gnu", NA, NA, NA, NA, NA, N…
$ package <chr> "healthyR.data", "healthyR.data", "healthyR.data", "healthyR…
$ version <chr> "1.0.0", "1.0.0", "1.0.0", "1.0.0", "1.0.0", "1.0.0", "1.0.0…
$ country <chr> "US", "US", "US", "GB", "US", "US", "DE", "HK", "JP", "US", …
$ ip_id <int> 2069, 2804, 78827, 27595, 90474, 90474, 42435, 74, 7655, 638…
The last day in the data set is 2026-01-05 23:44:26, the file was birthed on: 2025-10-31 10:47:59.603742, and at report knit time is 1592.94 hours old. Happy analyzing!
Now that we have our data lets take a look at it using the skimr
package.
skim(downloads_tbl)
| Name | downloads_tbl |
| Number of rows | 165021 |
| Number of columns | 11 |
| _______________________ | |
| Column type frequency: | |
| character | 6 |
| Date | 1 |
| numeric | 2 |
| POSIXct | 1 |
| Timespan | 1 |
| ________________________ | |
| Group variables | None |
Data summary
Variable type: character
| skim_variable | n_missing | complete_rate | min | max | empty | n_unique | whitespace |
|---|---|---|---|---|---|---|---|
| r_version | 121380 | 0.26 | 5 | 7 | 0 | 50 | 0 |
| r_arch | 121380 | 0.26 | 1 | 7 | 0 | 6 | 0 |
| r_os | 121380 | 0.26 | 7 | 19 | 0 | 24 | 0 |
| package | 0 | 1.00 | 7 | 13 | 0 | 8 | 0 |
| version | 0 | 1.00 | 5 | 17 | 0 | 62 | 0 |
| country | 15390 | 0.91 | 2 | 2 | 0 | 166 | 0 |
Variable type: Date
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date | 0 | 1 | 2020-11-23 | 2026-01-05 | 2023-11-17 | 1863 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1123594.48 | 1485769.2 | 355 | 31085 | 310611 | 2348316 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 11236.17 | 21865.3 | 1 | 233 | 2823 | 11797 | 299146 | ▇▁▁▁▁ |
Variable type: POSIXct
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date_time | 0 | 1 | 2020-11-23 09:00:41 | 2026-01-05 23:44:26 | 2023-11-17 12:33:09 | 104285 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 12H 6M 41S | 60 |
We can see that the following columns are missing a lot of data and for
us are most likely not useful anyways, so we will drop them
c(r_version, r_arch, r_os)
Plots
Now lets take a look at a time-series plot of the total daily downloads by package. We will use a log scale and place a vertical line at each version release for each package.
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]
[[6]]
[[7]]
[[8]]
Now lets take a look at some time series decomposition graphs.
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]
[[6]]
[[7]]
[[8]]
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]
[[6]]
[[7]]
[[8]]
Seasonal Diagnostics:
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]
[[6]]
[[7]]
[[8]]
ACF and PACF Diagnostics:
[[1]]
[[2]]
[[3]]
[[4]]
[[5]]
[[6]]
[[7]]
[[8]]
Feature Engineering
Now that we have our basic data and a shot of what it looks like, let’s
add some features to our data which can be very helpful in modeling.
Lets start by making a tibble that is aggregated by the day and
package, as we are going to be interested in forecasting the next 4
weeks or 28 days for each package. First lets get our base data.
Call:
stats::lm(formula = .formula, data = df)
Residuals:
Min 1Q Median 3Q Max
-148.31 -36.65 -11.17 27.15 820.41
Coefficients:
Estimate Std. Error
(Intercept) -1.779e+02 5.877e+01
date 1.093e-02 3.113e-03
lag(value, 1) 1.067e-01 2.301e-02
lag(value, 7) 8.761e-02 2.373e-02
lag(value, 14) 7.701e-02 2.369e-02
lag(value, 21) 8.547e-02 2.375e-02
lag(value, 28) 6.853e-02 2.367e-02
lag(value, 35) 5.199e-02 2.368e-02
lag(value, 42) 6.507e-02 2.378e-02
lag(value, 49) 6.545e-02 2.368e-02
month(date, label = TRUE).L -1.039e+01 4.918e+00
month(date, label = TRUE).Q 2.934e-01 4.810e+00
month(date, label = TRUE).C -1.574e+01 4.845e+00
month(date, label = TRUE)^4 -6.250e+00 4.873e+00
month(date, label = TRUE)^5 -6.543e+00 4.854e+00
month(date, label = TRUE)^6 1.167e+00 4.885e+00
month(date, label = TRUE)^7 -4.438e+00 4.830e+00
month(date, label = TRUE)^8 -4.101e+00 4.807e+00
month(date, label = TRUE)^9 2.817e+00 4.820e+00
month(date, label = TRUE)^10 9.131e-01 4.837e+00
month(date, label = TRUE)^11 -4.103e+00 4.823e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.110e+01 2.188e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.116e+00 2.263e+00
t value Pr(>|t|)
(Intercept) -3.026 0.002509 **
date 3.512 0.000457 ***
lag(value, 1) 4.636 3.82e-06 ***
lag(value, 7) 3.692 0.000229 ***
lag(value, 14) 3.251 0.001173 **
lag(value, 21) 3.598 0.000329 ***
lag(value, 28) 2.895 0.003838 **
lag(value, 35) 2.196 0.028234 *
lag(value, 42) 2.736 0.006278 **
lag(value, 49) 2.764 0.005773 **
month(date, label = TRUE).L -2.112 0.034862 *
month(date, label = TRUE).Q 0.061 0.951372
month(date, label = TRUE).C -3.249 0.001179 **
month(date, label = TRUE)^4 -1.283 0.199831
month(date, label = TRUE)^5 -1.348 0.177812
month(date, label = TRUE)^6 0.239 0.811193
month(date, label = TRUE)^7 -0.919 0.358358
month(date, label = TRUE)^8 -0.853 0.393627
month(date, label = TRUE)^9 0.584 0.559041
month(date, label = TRUE)^10 0.189 0.850278
month(date, label = TRUE)^11 -0.851 0.395028
fourier_vec(date, type = "sin", K = 1, period = 7) -5.074 4.30e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.144 0.001692 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.16 on 1791 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.2281, Adjusted R-squared: 0.2186
F-statistic: 24.05 on 22 and 1791 DF, p-value: < 2.2e-16
NNS Forecasting
This is something I have been wanting to try for a while. The NNS
package is a great package for forecasting time series data.
library(NNS)
data_list <- base_data |>
select(package, value) |>
group_split(package)
data_list |>
imap(
\(x, idx) {
obj <- x
x <- obj |> pull(value) |> tail(7*52)
train_set_size <- length(x) - 56
pkg <- obj |> pluck(1) |> unique()
# sf <- NNS.seas(x, modulo = 7, plot = FALSE)$periods
seas <- t(
sapply(
1:25,
function(i) c(
i,
sqrt(
mean((
NNS.ARMA(x,
h = 28,
training.set = train_set_size,
method = "lin",
seasonal.factor = i,
plot=FALSE
) - tail(x, 28)) ^ 2)))
)
)
colnames(seas) <- c("Period", "RMSE")
sf <- seas[which.min(seas[, 2]), 1]
cat(paste0("Package: ", pkg, "\n"))
NNS.ARMA.optim(
variable = x,
h = 28,
training.set = train_set_size,
#seasonal.factor = seq(12, 60, 7),
seasonal.factor = sf,
pred.int = 0.95,
plot = TRUE
)
title(
sub = paste0("\n",
"Package: ", pkg, " - NNS Optimization")
)
}
)
Package: healthyR
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 16 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 15.4152984346734"
[1] "BEST method = 'lin' PATH MEMBER = c( 16 )"
[1] "BEST lin OBJECTIVE FUNCTION = 15.4152984346734"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 16 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 14.9783221158427"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 16 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 14.9783221158427"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 16 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 15.9810155175667"
[1] "BEST method = 'both' PATH MEMBER = c( 16 )"
[1] "BEST both OBJECTIVE FUNCTION = 15.9810155175667"
Package: healthyR.ai
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 15.0155686528509"
[1] "BEST method = 'lin' PATH MEMBER = c( 4 )"
[1] "BEST lin OBJECTIVE FUNCTION = 15.0155686528509"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 11.0949617565685"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 4 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 11.0949617565685"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 14.0115820661357"
[1] "BEST method = 'both' PATH MEMBER = c( 4 )"
[1] "BEST both OBJECTIVE FUNCTION = 14.0115820661357"
Package: healthyR.data
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 8 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 17.1515591207442"
[1] "BEST method = 'lin' PATH MEMBER = c( 8 )"
[1] "BEST lin OBJECTIVE FUNCTION = 17.1515591207442"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 8 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 4.18655009539107"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 8 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 4.18655009539107"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 8 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 4.88280572046642"
[1] "BEST method = 'both' PATH MEMBER = c( 8 )"
[1] "BEST both OBJECTIVE FUNCTION = 4.88280572046642"
Package: healthyR.ts
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 13.1812470208737"
[1] "BEST method = 'lin' PATH MEMBER = c( 4 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.1812470208737"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.1949797444864"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 4 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.1949797444864"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 4 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 15.3919558667678"
[1] "BEST method = 'both' PATH MEMBER = c( 4 )"
[1] "BEST both OBJECTIVE FUNCTION = 15.3919558667678"
Package: healthyverse
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 2 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 12.2768225695489"
[1] "BEST method = 'lin' PATH MEMBER = c( 2 )"
[1] "BEST lin OBJECTIVE FUNCTION = 12.2768225695489"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 2 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 60.0571785006502"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 2 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 60.0571785006502"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 2 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 23.0419899836871"
[1] "BEST method = 'both' PATH MEMBER = c( 2 )"
[1] "BEST both OBJECTIVE FUNCTION = 23.0419899836871"
Package: RandomWalker
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 22 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 4.71435603451493"
[1] "BEST method = 'lin' PATH MEMBER = c( 22 )"
[1] "BEST lin OBJECTIVE FUNCTION = 4.71435603451493"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 22 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.16653785319029"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 22 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.16653785319029"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 22 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 6.69913022981683"
[1] "BEST method = 'both' PATH MEMBER = c( 22 )"
[1] "BEST both OBJECTIVE FUNCTION = 6.69913022981683"
Package: tidyAML
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 5 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 29.1310826811074"
[1] "BEST method = 'lin' PATH MEMBER = c( 5 )"
[1] "BEST lin OBJECTIVE FUNCTION = 29.1310826811074"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 5 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 5.09309826104778"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 5 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.09309826104778"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 5 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 4.20360441318497"
[1] "BEST method = 'both' PATH MEMBER = c( 5 )"
[1] "BEST both OBJECTIVE FUNCTION = 4.20360441318497"
Package: TidyDensity
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 11 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 13.1656287060795"
[1] "BEST method = 'lin' PATH MEMBER = c( 11 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.1656287060795"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 11 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 3.68785664839474"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 11 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 3.68785664839474"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 11 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 4.59617566633061"
[1] "BEST method = 'both' PATH MEMBER = c( 11 )"
[1] "BEST both OBJECTIVE FUNCTION = 4.59617566633061"
[[1]]
NULL
[[2]]
NULL
[[3]]
NULL
[[4]]
NULL
[[5]]
NULL
[[6]]
NULL
[[7]]
NULL
[[8]]
NULL
Pre-Processing
Now we are going to do some basic pre-processing.
data_padded_tbl <- base_data %>%
pad_by_time(
.date_var = date,
.pad_value = 0
)
# Get log interval and standardization parameters
log_params <- liv(data_padded_tbl$value, limit_lower = 0, offset = 1, silent = TRUE)
limit_lower <- log_params$limit_lower
limit_upper <- log_params$limit_upper
offset <- log_params$offset
data_liv_tbl <- data_padded_tbl %>%
# Get log interval transform
mutate(value_trans = liv(value, limit_lower = 0, offset = 1, silent = TRUE)$log_scaled)
# Get Standardization Params
std_params <- standard_vec(data_liv_tbl$value_trans, silent = TRUE)
std_mean <- std_params$mean
std_sd <- std_params$sd
data_transformed_tbl <- data_liv_tbl %>%
group_by(package) %>%
# get standardization
mutate(value_trans = standard_vec(value_trans, silent = TRUE)$standard_scaled) %>%
tk_augment_fourier(
.date_var = date,
.periods = c(7, 14, 30, 90, 180),
.K = 2
) %>%
tk_augment_timeseries_signature(
.date_var = date
) %>%
ungroup() %>%
select(-c(value, -year.iso))
Since this is panel data we can follow one of two different modeling strategies. We can search for a global model in the panel data or we can use nested forecasting finding the best model for each of the time series. Since we only have 5 panels, we will use nested forecasting.
To do this we will use the nest_timeseries and
split_nested_timeseries functions to create a nested tibble.
horizon <- 4*7
nested_data_tbl <- data_transformed_tbl %>%
# 0. Filter out column where package is NA
filter(!is.na(package)) %>%
# 1. Extending: We'll predict n days into the future.
extend_timeseries(
.id_var = package,
.date_var = date,
.length_future = horizon
) %>%
# 2. Nesting: We'll group by id, and create a future dataset
# that forecasts n days of extended data and
# an actual dataset that contains n*2 days
nest_timeseries(
.id_var = package,
.length_future = horizon
#.length_actual = horizon*2
) %>%
# 3. Splitting: We'll take the actual data and create splits
# for accuracy and confidence interval estimation of n das (test)
# and the rest is training data
split_nested_timeseries(
.length_test = horizon
)
nested_data_tbl
# A tibble: 8 × 4
package .actual_data .future_data .splits
<fct> <list> <list> <list>
1 healthyR.data <tibble [1,854 × 50]> <tibble [28 × 50]> <split [1826|28]>
2 healthyR <tibble [1,846 × 50]> <tibble [28 × 50]> <split [1818|28]>
3 healthyR.ts <tibble [1,788 × 50]> <tibble [28 × 50]> <split [1760|28]>
4 healthyverse <tibble [1,752 × 50]> <tibble [28 × 50]> <split [1724|28]>
5 healthyR.ai <tibble [1,588 × 50]> <tibble [28 × 50]> <split [1560|28]>
6 TidyDensity <tibble [1,439 × 50]> <tibble [28 × 50]> <split [1411|28]>
7 tidyAML <tibble [1,046 × 50]> <tibble [28 × 50]> <split [1018|28]>
8 RandomWalker <tibble [469 × 50]> <tibble [28 × 50]> <split [441|28]>
Now it is time to make some recipes and models using the modeltime workflow.
Modeltime Workflow
Recipe Object
recipe_base <- recipe(
value_trans ~ .
, data = extract_nested_test_split(nested_data_tbl)
)
recipe_base
recipe_date <- recipe(
value_trans ~ date
, data = extract_nested_test_split(nested_data_tbl)
)
Models
# Models ------------------------------------------------------------------
# Auto ARIMA --------------------------------------------------------------
model_spec_arima_no_boost <- arima_reg() %>%
set_engine(engine = "auto_arima")
wflw_auto_arima <- workflow() %>%
add_recipe(recipe = recipe_date) %>%
add_model(model_spec_arima_no_boost)
# NNETAR ------------------------------------------------------------------
model_spec_nnetar <- nnetar_reg(
mode = "regression"
, seasonal_period = "auto"
) %>%
set_engine("nnetar")
wflw_nnetar <- workflow() %>%
add_recipe(recipe = recipe_base) %>%
add_model(model_spec_nnetar)
# TSLM --------------------------------------------------------------------
model_spec_lm <- linear_reg() %>%
set_engine("lm")
wflw_lm <- workflow() %>%
add_recipe(recipe = recipe_base) %>%
add_model(model_spec_lm)
# MARS --------------------------------------------------------------------
model_spec_mars <- mars(mode = "regression") %>%
set_engine("earth")
wflw_mars <- workflow() %>%
add_recipe(recipe = recipe_date) %>%
add_model(model_spec_mars)
Nested Modeltime Tables
nested_modeltime_tbl <- modeltime_nested_fit(
# Nested Data
nested_data = nested_data_tbl,
control = control_nested_fit(
verbose = TRUE,
allow_par = FALSE
),
# Add workflows
wflw_auto_arima,
wflw_lm,
wflw_mars,
wflw_nnetar
)
nested_modeltime_tbl <- nested_modeltime_tbl[!is.na(nested_modeltime_tbl$package),]
Model Accuracy
nested_modeltime_tbl %>%
extract_nested_test_accuracy() %>%
filter(!is.na(package)) %>%
knitr::kable()
| package | .model_id | .model_desc | .type | mae | mape | mase | smape | rmse | rsq |
|---|---|---|---|---|---|---|---|---|---|
| healthyR.data | 1 | ARIMA | Test | 0.7700455 | 156.5418 | 0.6499645 | 146.2637 | 0.8726351 | 0.0540641 |
| healthyR.data | 2 | LM | Test | 0.6955837 | 173.5825 | 0.5871143 | 131.4360 | 0.8119165 | 0.0975661 |
| healthyR.data | 3 | EARTH | Test | 0.9753583 | 309.4283 | 0.8232608 | 126.4238 | 1.1612631 | 0.0056600 |
| healthyR.data | 4 | NNAR | Test | 0.7644509 | 177.2692 | 0.6452423 | 137.4837 | 0.9157157 | 0.0006987 |
| healthyR | 1 | ARIMA | Test | 0.5436782 | 286.2946 | 0.7427131 | 121.9137 | 0.6980680 | 0.1180019 |
| healthyR | 2 | LM | Test | 0.5904421 | 451.8492 | 0.8065967 | 109.5139 | 0.7497147 | 0.0051117 |
| healthyR | 3 | EARTH | Test | 0.8376259 | 868.6402 | 1.1442720 | 115.2661 | 0.9675747 | 0.0997740 |
| healthyR | 4 | NNAR | Test | 0.5370739 | 239.9227 | 0.7336910 | 117.4048 | 0.7119163 | 0.0346538 |
| healthyR.ts | 1 | ARIMA | Test | 0.7099362 | 130.6125 | 0.7712800 | 140.6835 | 1.0012655 | 0.0032415 |
| healthyR.ts | 2 | LM | Test | 0.7520962 | 122.2661 | 0.8170830 | 122.1258 | 1.0517861 | 0.0014138 |
| healthyR.ts | 3 | EARTH | Test | 0.6428525 | 134.8351 | 0.6983998 | 95.4240 | 0.8968520 | 0.0847293 |
| healthyR.ts | 4 | NNAR | Test | 0.7554584 | 123.7695 | 0.8207357 | 127.1870 | 1.0666258 | 0.0005618 |
| healthyverse | 1 | ARIMA | Test | 0.6972805 | 104.3177 | 0.7581001 | 109.4398 | 0.8683722 | 0.0013685 |
| healthyverse | 2 | LM | Test | 0.7664275 | 127.5640 | 0.8332783 | 120.0347 | 0.9063872 | 0.0154571 |
| healthyverse | 3 | EARTH | Test | 0.9256873 | 231.2822 | 1.0064293 | 101.5756 | 1.0751317 | 0.0867023 |
| healthyverse | 4 | NNAR | Test | 0.7580834 | 125.0793 | 0.8242064 | 128.7876 | 0.9457266 | 0.0264734 |
| healthyR.ai | 1 | ARIMA | Test | 1.0462267 | 138.9485 | 0.9292592 | 175.6361 | 1.2647297 | 0.0042927 |
| healthyR.ai | 2 | LM | Test | 0.8650978 | 147.0916 | 0.7683805 | 136.3783 | 1.0302146 | 0.0820251 |
| healthyR.ai | 3 | EARTH | Test | 1.5112775 | 380.1440 | 1.3423177 | 109.7202 | 1.9404137 | 0.0706054 |
| healthyR.ai | 4 | NNAR | Test | 0.9105342 | 167.4718 | 0.8087371 | 149.4882 | 1.0664483 | 0.0768396 |
| TidyDensity | 1 | ARIMA | Test | 1.0994889 | 178.0195 | 0.5689558 | 159.5115 | 1.2160380 | 0.1208148 |
| TidyDensity | 2 | LM | Test | 1.1263779 | 192.8379 | 0.5828701 | 156.2024 | 1.2604858 | 0.0527263 |
| TidyDensity | 3 | EARTH | Test | 1.1401044 | 150.6733 | 0.5899732 | 124.8304 | 1.4422800 | 0.0031176 |
| TidyDensity | 4 | NNAR | Test | 1.1254646 | 203.0156 | 0.5823975 | 156.8293 | 1.2448875 | 0.0545559 |
| tidyAML | 1 | ARIMA | Test | 0.8072872 | 107.7089 | 0.9493717 | 131.2575 | 1.0037174 | 0.0037743 |
| tidyAML | 2 | LM | Test | 0.8772240 | 185.9247 | 1.0316176 | 136.4019 | 1.0609471 | 0.0555635 |
| tidyAML | 3 | EARTH | Test | 1.2227186 | 188.9284 | 1.4379200 | 176.6407 | 1.5188868 | 0.1626243 |
| tidyAML | 4 | NNAR | Test | 0.7911418 | 213.4876 | 0.9303847 | 120.9961 | 0.8969707 | 0.0745942 |
| RandomWalker | 1 | ARIMA | Test | 0.8671046 | 121.4217 | 0.6026929 | 175.2116 | 0.9837434 | 0.0600833 |
| RandomWalker | 2 | LM | Test | 0.9465890 | 137.3449 | 0.6579397 | 159.7801 | 1.0977088 | 0.0026896 |
| RandomWalker | 3 | EARTH | Test | 1.0901726 | 217.3735 | 0.7577395 | 124.0966 | 1.4414515 | 0.0172453 |
| RandomWalker | 4 | NNAR | Test | 0.9547820 | 205.1002 | 0.6636343 | 157.0835 | 1.0857085 | 0.0035698 |
Plot Models
nested_modeltime_tbl %>%
extract_nested_test_forecast() %>%
group_by(package) %>%
filter_by_time(.date_var = .index, .start_date = max(.index) - 60) %>%
ungroup() %>%
plot_modeltime_forecast(
.interactive = FALSE,
.conf_interval_show = FALSE,
.facet_scales = "free"
) +
theme_minimal() +
facet_wrap(~ package, nrow = 3) +
theme(legend.position = "bottom")
Best Model
best_nested_modeltime_tbl <- nested_modeltime_tbl %>%
modeltime_nested_select_best(
metric = "rmse",
minimize = TRUE,
filter_test_forecasts = TRUE
)
best_nested_modeltime_tbl %>%
extract_nested_best_model_report()
# Nested Modeltime Table
# A tibble: 8 × 10
package .model_id .model_desc .type mae mape mase smape rmse rsq
<fct> <int> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
1 healthyR.da… 2 LM Test 0.696 174. 0.587 131. 0.812 0.0976
2 healthyR 1 ARIMA Test 0.544 286. 0.743 122. 0.698 0.118
3 healthyR.ts 3 EARTH Test 0.643 135. 0.698 95.4 0.897 0.0847
4 healthyverse 1 ARIMA Test 0.697 104. 0.758 109. 0.868 0.00137
5 healthyR.ai 2 LM Test 0.865 147. 0.768 136. 1.03 0.0820
6 TidyDensity 1 ARIMA Test 1.10 178. 0.569 160. 1.22 0.121
7 tidyAML 4 NNAR Test 0.791 213. 0.930 121. 0.897 0.0746
8 RandomWalker 1 ARIMA Test 0.867 121. 0.603 175. 0.984 0.0601
best_nested_modeltime_tbl %>%
extract_nested_test_forecast() %>%
#filter(!is.na(.model_id)) %>%
group_by(package) %>%
filter_by_time(.date_var = .index, .start_date = max(.index) - 60) %>%
ungroup() %>%
plot_modeltime_forecast(
.interactive = FALSE,
.conf_interval_alpha = 0.2,
.facet_scales = "free"
) +
facet_wrap(~ package, nrow = 3) +
theme_minimal() +
theme(legend.position = "bottom")
Refitting and Future Forecast
Now that we have the best models, we can make our future forecasts.
nested_modeltime_refit_tbl <- best_nested_modeltime_tbl %>%
modeltime_nested_refit(
control = control_nested_refit(verbose = TRUE)
)
nested_modeltime_refit_tbl
# Nested Modeltime Table
# A tibble: 8 × 5
package .actual_data .future_data .splits .modeltime_tables
<fct> <list> <list> <list> <list>
1 healthyR.data <tibble> <tibble> <split [1826|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [1818|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1760|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1724|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1560|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1411|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [1018|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [441|28]> <mdl_tm_t [1 × 5]>
nested_modeltime_refit_tbl %>%
extract_nested_future_forecast() %>%
group_by(package) %>%
mutate(across(.value:.conf_hi, .fns = ~ standard_inv_vec(
x = .,
mean = std_mean,
sd = std_sd
)$standard_inverse_value)) %>%
mutate(across(.value:.conf_hi, .fns = ~ liiv(
x = .,
limit_lower = limit_lower,
limit_upper = limit_upper,
offset = offset
)$rescaled_v)) %>%
filter_by_time(.date_var = .index, .start_date = max(.index) - 60) %>%
ungroup() %>%
plot_modeltime_forecast(
.interactive = FALSE,
.conf_interval_alpha = 0.2,
.facet_scales = "free"
) +
facet_wrap(~ package, nrow = 3) +
theme_minimal() +
theme(legend.position = "bottom")
