healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse Packages
Steven P. Sanderson II, MPH - Date: 2026-10-26
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: 193,935
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 2026-10-24 22:59:11, the file was birthed on: 2025-10-31 10:47:59.603742, and at report knit time is 8600.19 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 | 193935 |
| 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 | 146562 | 0.24 | 1 | 17 | 0 | 55 | 0 |
| r_arch | 146562 | 0.24 | 1 | 7 | 0 | 8 | 0 |
| r_os | 146562 | 0.24 | 6 | 33 | 0 | 43 | 0 |
| package | 0 | 1.00 | 7 | 13 | 0 | 8 | 0 |
| version | 0 | 1.00 | 5 | 17 | 0 | 63 | 0 |
| country | 18659 | 0.90 | 2 | 2 | 0 | 173 | 0 |
Variable type: Date
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date | 0 | 1 | 2020-11-23 | 2026-10-24 | 2024-04-03 | 2155 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1162548.23 | 1478697.54 | 355 | 47146 | 361718 | 2367747 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 13537.88 | 28874.95 | 1 | 189 | 2818 | 12780 | 429286 | ▇▁▁▁▁ |
Variable type: POSIXct
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date_time | 0 | 1 | 2020-11-23 09:00:41 | 2026-10-24 22:59:11 | 2024-04-03 16:47:06 | 126892 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 12H 16M 1S | 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
-148.97 -38.42 -11.66 28.31 827.84
Coefficients:
Estimate Std. Error
(Intercept) -1.344e+02 4.544e+01
date 8.814e-03 2.391e-03
lag(value, 1) 9.934e-02 2.148e-02
lag(value, 7) 7.133e-02 2.206e-02
lag(value, 14) 7.530e-02 2.192e-02
lag(value, 21) 8.711e-02 2.196e-02
lag(value, 28) 7.941e-02 2.191e-02
lag(value, 35) 3.630e-02 2.193e-02
lag(value, 42) 6.041e-02 2.198e-02
lag(value, 49) 6.701e-02 2.192e-02
month(date, label = TRUE).L -7.924e+00 4.668e+00
month(date, label = TRUE).Q 2.013e+00 4.593e+00
month(date, label = TRUE).C -1.535e+01 4.604e+00
month(date, label = TRUE)^4 -1.048e+01 4.637e+00
month(date, label = TRUE)^5 -5.554e+00 4.617e+00
month(date, label = TRUE)^6 8.831e-01 4.617e+00
month(date, label = TRUE)^7 -1.481e+00 4.543e+00
month(date, label = TRUE)^8 -3.816e+00 4.486e+00
month(date, label = TRUE)^9 -2.316e-01 4.466e+00
month(date, label = TRUE)^10 -3.321e-01 4.453e+00
month(date, label = TRUE)^11 -5.263e-01 4.447e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.055e+01 2.019e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.274e+00 2.075e+00
t value Pr(>|t|)
(Intercept) -2.957 0.003142 **
date 3.687 0.000232 ***
lag(value, 1) 4.625 3.97e-06 ***
lag(value, 7) 3.234 0.001240 **
lag(value, 14) 3.436 0.000602 ***
lag(value, 21) 3.966 7.56e-05 ***
lag(value, 28) 3.624 0.000298 ***
lag(value, 35) 1.655 0.098019 .
lag(value, 42) 2.749 0.006035 **
lag(value, 49) 3.057 0.002268 **
month(date, label = TRUE).L -1.698 0.089720 .
month(date, label = TRUE).Q 0.438 0.661330
month(date, label = TRUE).C -3.334 0.000870 ***
month(date, label = TRUE)^4 -2.260 0.023932 *
month(date, label = TRUE)^5 -1.203 0.229129
month(date, label = TRUE)^6 0.191 0.848314
month(date, label = TRUE)^7 -0.326 0.744386
month(date, label = TRUE)^8 -0.851 0.395077
month(date, label = TRUE)^9 -0.052 0.958639
month(date, label = TRUE)^10 -0.075 0.940562
month(date, label = TRUE)^11 -0.118 0.905801
fourier_vec(date, type = "sin", K = 1, period = 7) -5.223 1.94e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.506 0.000464 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.9 on 2083 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.1955, Adjusted R-squared: 0.187
F-statistic: 23.01 on 22 and 2083 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( 13 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 9.09281919742698"
[1] "BEST method = 'lin' PATH MEMBER = c( 13 )"
[1] "BEST lin OBJECTIVE FUNCTION = 9.09281919742698"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.57746074323282"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 13 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.57746074323282"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 10.0039933385296"
[1] "BEST method = 'both' PATH MEMBER = c( 13 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.0039933385296"
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( 7 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 24.7929677406685"
[1] "BEST method = 'lin' PATH MEMBER = c( 7 )"
[1] "BEST lin OBJECTIVE FUNCTION = 24.7929677406685"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 7 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 5.95536327517294"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 7 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.95536327517294"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 7 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 10.0353702203128"
[1] "BEST method = 'both' PATH MEMBER = c( 7 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.0353702203128"
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( 13 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 14.1087626062112"
[1] "BEST method = 'lin' PATH MEMBER = c( 13 )"
[1] "BEST lin OBJECTIVE FUNCTION = 14.1087626062112"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 10.8799602647175"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 13 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 10.8799602647175"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 10.6527281146528"
[1] "BEST method = 'both' PATH MEMBER = c( 13 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.6527281146528"
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( 13 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 14.0877453490794"
[1] "BEST method = 'lin' PATH MEMBER = c( 13 )"
[1] "BEST lin OBJECTIVE FUNCTION = 14.0877453490794"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 9.9813364401029"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 13 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 9.9813364401029"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 13 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 12.7991628296033"
[1] "BEST method = 'both' PATH MEMBER = c( 13 )"
[1] "BEST both OBJECTIVE FUNCTION = 12.7991628296033"
Package: healthyverse
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 24 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 27.887832200682"
[1] "BEST method = 'lin' PATH MEMBER = c( 24 )"
[1] "BEST lin OBJECTIVE FUNCTION = 27.887832200682"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 24 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 22.2172312880256"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 24 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 22.2172312880256"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 24 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 33.1015176582596"
[1] "BEST method = 'both' PATH MEMBER = c( 24 )"
[1] "BEST both OBJECTIVE FUNCTION = 33.1015176582596"
Package: RandomWalker
[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 = 5.99376680274813"
[1] "BEST method = 'lin' PATH MEMBER = c( 2 )"
[1] "BEST lin OBJECTIVE FUNCTION = 5.99376680274813"
[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 = 11.4252779594901"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 2 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 11.4252779594901"
[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 = 24.5575213671155"
[1] "BEST method = 'both' PATH MEMBER = c( 2 )"
[1] "BEST both OBJECTIVE FUNCTION = 24.5575213671155"
Package: tidyAML
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 21 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 13.8276792014672"
[1] "BEST method = 'lin' PATH MEMBER = c( 21 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.8276792014672"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 21 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 9.42086911859023"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 21 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 9.42086911859023"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 21 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 10.4709846688237"
[1] "BEST method = 'both' PATH MEMBER = c( 21 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.4709846688237"
Package: TidyDensity
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 9 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 14.2309236240723"
[1] "BEST method = 'lin' PATH MEMBER = c( 9 )"
[1] "BEST lin OBJECTIVE FUNCTION = 14.2309236240723"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 9 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 5.24929873879259"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 9 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.24929873879259"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 9 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 4.72948244047585"
[1] "BEST method = 'both' PATH MEMBER = c( 9 )"
[1] "BEST both OBJECTIVE FUNCTION = 4.72948244047585"
[[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 [2,143 × 50]> <tibble [28 × 50]> <split [2115|28]>
2 healthyR <tibble [2,137 × 50]> <tibble [28 × 50]> <split [2109|28]>
3 healthyR.ts <tibble [2,073 × 50]> <tibble [28 × 50]> <split [2045|28]>
4 healthyverse <tibble [1,945 × 50]> <tibble [28 × 50]> <split [1917|28]>
5 healthyR.ai <tibble [1,878 × 50]> <tibble [28 × 50]> <split [1850|28]>
6 TidyDensity <tibble [1,731 × 50]> <tibble [28 × 50]> <split [1703|28]>
7 tidyAML <tibble [1,334 × 50]> <tibble [28 × 50]> <split [1306|28]>
8 RandomWalker <tibble [758 × 50]> <tibble [28 × 50]> <split [730|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 | 0.8651133 | 113.0289 | 1.0154985 | 187.25304 | 1.0615497 | 0.0261935 |
| healthyR.data | 2 | LM | Test | 1.0890138 | 226.3037 | 1.2783204 | 184.27829 | 1.2690802 | 0.0000173 |
| healthyR.data | 3 | EARTH | Test | 0.8550263 | 487.4628 | 1.0036581 | 101.30996 | 1.0354231 | 0.0355169 |
| healthyR.data | 4 | NNAR | Test | 1.0773934 | 234.1398 | 1.2646800 | 174.65558 | 1.2593109 | 0.0969285 |
| healthyR | 1 | ARIMA | Test | 0.6919739 | 157.2021 | 0.9520086 | 154.19997 | 0.8825884 | 0.0185845 |
| healthyR | 2 | LM | Test | 0.8316224 | 250.9897 | 1.1441353 | 151.71484 | 1.0031851 | 0.0030209 |
| healthyR | 3 | EARTH | Test | 1.0054360 | 454.9649 | 1.3832657 | 125.90868 | 1.1743392 | 0.0006680 |
| healthyR | 4 | NNAR | Test | 0.6858478 | 157.6953 | 0.9435804 | 142.71807 | 0.8634280 | 0.0025810 |
| healthyR.ts | 1 | ARIMA | Test | 0.6363540 | 118.7825 | 0.8038915 | 164.69268 | 0.8622758 | 0.0330353 |
| healthyR.ts | 2 | LM | Test | 0.6694607 | 251.6952 | 0.8457144 | 150.24200 | 0.8773853 | 0.0344034 |
| healthyR.ts | 3 | EARTH | Test | 0.9872314 | 1068.8247 | 1.2471468 | 128.38547 | 1.1150418 | 0.0018060 |
| healthyR.ts | 4 | NNAR | Test | 0.6788221 | 249.3023 | 0.8575404 | 141.03236 | 0.8632452 | 0.0017984 |
| healthyverse | 1 | ARIMA | Test | 0.6974928 | 162.0295 | 0.8007411 | 73.82861 | 0.8731024 | 0.0435063 |
| healthyverse | 2 | LM | Test | 0.6874345 | 104.4969 | 0.7891938 | 92.36817 | 0.8798812 | 0.0333636 |
| healthyverse | 3 | EARTH | Test | 0.7129231 | 180.6754 | 0.8184554 | 73.24776 | 0.8993790 | 0.0009680 |
| healthyverse | 4 | NNAR | Test | 0.8200156 | 100.2923 | 0.9414006 | 119.70159 | 1.0183693 | 0.0014117 |
| healthyR.ai | 1 | ARIMA | Test | 0.6648459 | 156.1649 | 0.8415350 | 115.83029 | 0.7824593 | 0.0003253 |
| healthyR.ai | 2 | LM | Test | 0.9370796 | 171.0687 | 1.1861176 | 159.42885 | 1.1177953 | 0.0006216 |
| healthyR.ai | 3 | EARTH | Test | 1.1301316 | 496.4953 | 1.4304749 | 109.49169 | 1.3189933 | 0.0345781 |
| healthyR.ai | 4 | NNAR | Test | 0.7855307 | 131.7372 | 0.9942930 | 161.08670 | 0.9434796 | 0.0784469 |
| TidyDensity | 1 | ARIMA | Test | 0.5891822 | 128.7947 | 0.8729015 | 170.06557 | 0.7179079 | 0.0911543 |
| TidyDensity | 2 | LM | Test | 0.6640120 | 441.7173 | 0.9837653 | 131.61722 | 0.8352347 | 0.0331056 |
| TidyDensity | 3 | EARTH | Test | 0.5978309 | 146.7833 | 0.8857150 | 151.05347 | 0.7362319 | 0.0279920 |
| TidyDensity | 4 | NNAR | Test | 0.7926655 | 577.0494 | 1.1743717 | 128.06156 | 0.9958266 | 0.0252674 |
| tidyAML | 1 | ARIMA | Test | 0.6732975 | 209.0142 | 0.7809990 | 104.14576 | 0.8558773 | 0.0209386 |
| tidyAML | 2 | LM | Test | 0.7778321 | 221.6391 | 0.9022551 | 147.40018 | 0.9013967 | 0.0207695 |
| tidyAML | 3 | EARTH | Test | 1.2837890 | 494.6810 | 1.4891455 | 117.77041 | 1.5194251 | 0.0024572 |
| tidyAML | 4 | NNAR | Test | 0.6461430 | 193.2903 | 0.7495008 | 112.08761 | 0.8047833 | 0.0585898 |
| RandomWalker | 1 | ARIMA | Test | 0.4971538 | 157.6406 | 0.6374945 | 135.05533 | 0.6762756 | 0.0340740 |
| RandomWalker | 2 | LM | Test | 0.5267698 | 155.9874 | 0.6754708 | 128.76688 | 0.7110380 | 0.0004186 |
| RandomWalker | 3 | EARTH | Test | 0.7637426 | 441.1055 | 0.9793382 | 137.78441 | 0.8875769 | 0.0577803 |
| RandomWalker | 4 | NNAR | Test | 0.6156878 | 260.8952 | 0.7894894 | 132.90741 | 0.8137641 | 0.0025054 |
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… 3 EARTH Test 0.855 487. 1.00 101. 1.04 3.55e-2
2 healthyR 4 NNAR Test 0.686 158. 0.944 143. 0.863 2.58e-3
3 healthyR.ts 1 ARIMA Test 0.636 119. 0.804 165. 0.862 3.30e-2
4 healthyverse 1 ARIMA Test 0.697 162. 0.801 73.8 0.873 4.35e-2
5 healthyR.ai 1 ARIMA Test 0.665 156. 0.842 116. 0.782 3.25e-4
6 TidyDensity 1 ARIMA Test 0.589 129. 0.873 170. 0.718 9.12e-2
7 tidyAML 4 NNAR Test 0.646 193. 0.750 112. 0.805 5.86e-2
8 RandomWalker 1 ARIMA Test 0.497 158. 0.637 135. 0.676 3.41e-2
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 [2115|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [2109|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [2045|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1917|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1850|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1703|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [1306|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [730|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")
