healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse
Packages Steven P. Sanderson II, MPH - Date: 2025-10-17
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: 157,000
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 2025-10-15 23:22:23, the file was birthed on: 2024-08-07 07:35:44.428716, and at report knit time is 1.042778^{4} 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 | 157000 |
| 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 | 114833 | 0.27 | 5 | 5 | 0 | 48 | 0 |
| r_arch | 114833 | 0.27 | 3 | 7 | 0 | 5 | 0 |
| r_os | 114833 | 0.27 | 7 | 15 | 0 | 23 | 0 |
| package | 0 | 1.00 | 7 | 13 | 0 | 8 | 0 |
| version | 0 | 1.00 | 5 | 17 | 0 | 62 | 0 |
| country | 14762 | 0.91 | 2 | 2 | 0 | 165 | 0 |
Variable type: Date
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date | 0 | 1 | 2020-11-23 | 2025-10-15 | 2023-10-02 | 1781 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1124750.1 | 1494610.50 | 355 | 16879 | 308131 | 2360708 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 11328.6 | 21946.94 | 1 | 204 | 2928 | 11961 | 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 | 2025-10-15 23:22:23 | 2023-10-02 21:25:16 | 98643 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 20.5 | 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
-147.81 -36.24 -10.98 27.04 815.22
Coefficients:
Estimate Std. Error
(Intercept) -2.122e+02 6.232e+01
date 1.266e-02 3.303e-03
lag(value, 1) 1.159e-01 2.346e-02
lag(value, 7) 8.967e-02 2.429e-02
lag(value, 14) 8.200e-02 2.433e-02
lag(value, 21) 6.750e-02 2.440e-02
lag(value, 28) 7.696e-02 2.431e-02
lag(value, 35) 6.790e-02 2.452e-02
lag(value, 42) 5.527e-02 2.467e-02
lag(value, 49) 6.815e-02 2.454e-02
month(date, label = TRUE).L -8.077e+00 5.109e+00
month(date, label = TRUE).Q 2.233e+00 5.019e+00
month(date, label = TRUE).C -1.634e+01 5.042e+00
month(date, label = TRUE)^4 -8.302e+00 5.067e+00
month(date, label = TRUE)^5 -9.550e+00 5.035e+00
month(date, label = TRUE)^6 -1.258e+00 5.073e+00
month(date, label = TRUE)^7 -6.039e+00 4.964e+00
month(date, label = TRUE)^8 -4.596e+00 4.916e+00
month(date, label = TRUE)^9 2.721e+00 4.869e+00
month(date, label = TRUE)^10 1.021e+00 4.852e+00
month(date, label = TRUE)^11 -4.160e+00 4.832e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.122e+01 2.256e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 6.814e+00 2.345e+00
t value Pr(>|t|)
(Intercept) -3.405 0.000677 ***
date 3.832 0.000131 ***
lag(value, 1) 4.941 8.51e-07 ***
lag(value, 7) 3.691 0.000230 ***
lag(value, 14) 3.371 0.000767 ***
lag(value, 21) 2.766 0.005731 **
lag(value, 28) 3.166 0.001572 **
lag(value, 35) 2.769 0.005685 **
lag(value, 42) 2.240 0.025215 *
lag(value, 49) 2.777 0.005548 **
month(date, label = TRUE).L -1.581 0.114060
month(date, label = TRUE).Q 0.445 0.656363
month(date, label = TRUE).C -3.242 0.001211 **
month(date, label = TRUE)^4 -1.638 0.101521
month(date, label = TRUE)^5 -1.897 0.058026 .
month(date, label = TRUE)^6 -0.248 0.804244
month(date, label = TRUE)^7 -1.216 0.223965
month(date, label = TRUE)^8 -0.935 0.350011
month(date, label = TRUE)^9 0.559 0.576245
month(date, label = TRUE)^10 0.210 0.833423
month(date, label = TRUE)^11 -0.861 0.389407
fourier_vec(date, type = "sin", K = 1, period = 7) -4.973 7.26e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 2.906 0.003706 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.23 on 1709 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.2392, Adjusted R-squared: 0.2294
F-statistic: 24.42 on 22 and 1709 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( 3 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 8.14355679426738"
[1] "BEST method = 'lin' PATH MEMBER = c( 3 )"
[1] "BEST lin OBJECTIVE FUNCTION = 8.14355679426738"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 3 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 11.5188963808431"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 3 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 11.5188963808431"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 3 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 8.08916319720732"
[1] "BEST method = 'both' PATH MEMBER = c( 3 )"
[1] "BEST both OBJECTIVE FUNCTION = 8.08916319720732"
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( 5 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 53.8707732608271"
[1] "BEST method = 'lin' PATH MEMBER = c( 5 )"
[1] "BEST lin OBJECTIVE FUNCTION = 53.8707732608271"
[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 = 35.6602438776922"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 5 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 35.6602438776922"
[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 = 66.2911720840684"
[1] "BEST method = 'both' PATH MEMBER = c( 5 )"
[1] "BEST both OBJECTIVE FUNCTION = 66.2911720840684"
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( 22 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 7.55984402656023"
[1] "BEST method = 'lin' PATH MEMBER = c( 22 )"
[1] "BEST lin OBJECTIVE FUNCTION = 7.55984402656023"
[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 = 7.40948053856251"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 22 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.40948053856251"
[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 = 6.39078789196786"
[1] "BEST method = 'both' PATH MEMBER = c( 22 )"
[1] "BEST both OBJECTIVE FUNCTION = 6.39078789196786"
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( 2 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 65.8538420444278"
[1] "BEST method = 'lin' PATH MEMBER = c( 2 )"
[1] "BEST lin OBJECTIVE FUNCTION = 65.8538420444278"
[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 = 24.402423908038"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 2 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 24.402423908038"
[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 = 26.3221262460319"
[1] "BEST method = 'both' PATH MEMBER = c( 2 )"
[1] "BEST both OBJECTIVE FUNCTION = 26.3221262460319"
Package: healthyverse
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 10 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 13.965442386704"
[1] "BEST method = 'lin' PATH MEMBER = c( 10 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.965442386704"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 10 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 12.0106291038189"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 10 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 12.0106291038189"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 10 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 9.06103294267949"
[1] "BEST method = 'both' PATH MEMBER = c( 10 )"
[1] "BEST both OBJECTIVE FUNCTION = 9.06103294267949"
Package: RandomWalker
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 25 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 3.78036881955424"
[1] "BEST method = 'lin' PATH MEMBER = c( 25 )"
[1] "BEST lin OBJECTIVE FUNCTION = 3.78036881955424"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 25 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.19230486208912"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 25 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.19230486208912"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 25 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 5.83321048649789"
[1] "BEST method = 'both' PATH MEMBER = c( 25 )"
[1] "BEST both OBJECTIVE FUNCTION = 5.83321048649789"
Package: tidyAML
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 18 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 43.6978859742687"
[1] "BEST method = 'lin' PATH MEMBER = c( 18 )"
[1] "BEST lin OBJECTIVE FUNCTION = 43.6978859742687"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 18 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 28.3202531119387"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 18 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 28.3202531119387"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 18 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 41.8573100943215"
[1] "BEST method = 'both' PATH MEMBER = c( 18 )"
[1] "BEST both OBJECTIVE FUNCTION = 41.8573100943215"
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 = 34.1253496735053"
[1] "BEST method = 'lin' PATH MEMBER = c( 11 )"
[1] "BEST lin OBJECTIVE FUNCTION = 34.1253496735053"
[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 = 11.0860039979295"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 11 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 11.0860039979295"
[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 = 24.9634405673427"
[1] "BEST method = 'both' PATH MEMBER = c( 11 )"
[1] "BEST both OBJECTIVE FUNCTION = 24.9634405673427"
[[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 [1,773 × 50]> <tibble [28 × 50]> <split [1745|28]>
2 healthyR <tibble [1,764 × 50]> <tibble [28 × 50]> <split [1736|28]>
3 healthyR.ts <tibble [1,710 × 50]> <tibble [28 × 50]> <split [1682|28]>
4 healthyverse <tibble [1,681 × 50]> <tibble [28 × 50]> <split [1653|28]>
5 healthyR.ai <tibble [1,506 × 50]> <tibble [28 × 50]> <split [1478|28]>
6 TidyDensity <tibble [1,357 × 50]> <tibble [28 × 50]> <split [1329|28]>
7 tidyAML <tibble [964 × 50]> <tibble [28 × 50]> <split [936|28]>
8 RandomWalker <tibble [387 × 50]> <tibble [28 × 50]> <split [359|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 | 0.6445486 | 101.61201 | 0.6942080 | 162.4485 | 0.8137877 | 0.0532867 |
| healthyR.data | 2 | LM | Test | 0.8037314 | 177.53076 | 0.8656551 | 143.2744 | 0.9723433 | 0.0049995 |
| healthyR.data | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| healthyR.data | 4 | NNAR | Test | 0.7536167 | 129.62546 | 0.8116793 | 148.2688 | 0.9491970 | 0.0032517 |
| healthyR | 1 | ARIMA | Test | 0.5319473 | 146.89514 | 0.7896331 | 173.0737 | 0.6663800 | 0.0085956 |
| healthyR | 2 | LM | Test | 0.7295269 | 242.82709 | 1.0829242 | 154.5287 | 0.9451825 | 0.0283342 |
| healthyR | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| healthyR | 4 | NNAR | Test | 0.5898167 | 187.32180 | 0.8755355 | 141.1373 | 0.8274473 | 0.0378338 |
| healthyR.ts | 1 | ARIMA | Test | 0.5060458 | 94.93315 | 0.7209496 | 158.1647 | 0.6437125 | 0.0001573 |
| healthyR.ts | 2 | LM | Test | 0.6270048 | 138.22964 | 0.8932765 | 137.1143 | 0.7816929 | 0.0036640 |
| healthyR.ts | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| healthyR.ts | 4 | NNAR | Test | 0.7022620 | 186.87738 | 1.0004933 | 148.3743 | 0.8428415 | 0.0170186 |
| healthyverse | 1 | ARIMA | Test | 0.6740553 | 118.03375 | 0.8020991 | 157.9121 | 0.8100768 | 0.0353154 |
| healthyverse | 2 | LM | Test | 0.6767318 | 177.67356 | 0.8052841 | 152.3307 | 0.8110548 | 0.1242553 |
| healthyverse | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| healthyverse | 4 | NNAR | Test | 0.6145310 | 169.13696 | 0.7312676 | 140.2969 | 0.7320920 | 0.0884013 |
| healthyR.ai | 1 | ARIMA | Test | 0.4909486 | 106.39598 | 0.8693612 | 151.2732 | 0.6219120 | 0.0042982 |
| healthyR.ai | 2 | LM | Test | 0.4617832 | 109.77640 | 0.8177158 | 134.3170 | 0.6255310 | 0.0548925 |
| healthyR.ai | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| healthyR.ai | 4 | NNAR | Test | 0.5650957 | 156.28646 | 1.0006593 | 140.5847 | 0.7142858 | 0.0000382 |
| TidyDensity | 1 | ARIMA | Test | 1.2718747 | 268.08817 | 0.9695626 | 119.1108 | 1.5500526 | 0.0043474 |
| TidyDensity | 2 | LM | Test | 1.5543726 | 199.40353 | 1.1849135 | 161.3009 | 1.9254038 | 0.0117932 |
| TidyDensity | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| TidyDensity | 4 | NNAR | Test | 1.3054563 | 202.86336 | 0.9951622 | 127.1320 | 1.5870346 | 0.0544057 |
| tidyAML | 1 | ARIMA | Test | 0.9514485 | 119.71351 | 0.9244233 | 157.5239 | 1.3696256 | 0.0015414 |
| tidyAML | 2 | LM | Test | 1.0359531 | 199.53931 | 1.0065276 | 156.9688 | 1.5004747 | 0.0002959 |
| tidyAML | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| tidyAML | 4 | NNAR | Test | 1.0032804 | 180.36467 | 0.9747829 | 155.8998 | 1.4252395 | 0.0252275 |
| RandomWalker | 1 | ARIMA | Test | 0.7356171 | 230.06898 | 0.5664817 | 148.1566 | 0.8301111 | 0.2861879 |
| RandomWalker | 2 | LM | Test | 0.7855958 | 598.35566 | 0.6049692 | 145.0738 | 0.9353710 | 0.0029209 |
| RandomWalker | 3 | NULL | NA | NA | NA | NA | NA | NA | NA |
| RandomWalker | 4 | NNAR | Test | 0.8927815 | 430.40204 | 0.6875103 | 154.1877 | 1.1016064 | 0.0510776 |
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… 1 ARIMA Test 0.645 102. 0.694 162. 0.814 5.33e-2
2 healthyR 1 ARIMA Test 0.532 147. 0.790 173. 0.666 8.60e-3
3 healthyR.ts 1 ARIMA Test 0.506 94.9 0.721 158. 0.644 1.57e-4
4 healthyverse 4 NNAR Test 0.615 169. 0.731 140. 0.732 8.84e-2
5 healthyR.ai 1 ARIMA Test 0.491 106. 0.869 151. 0.622 4.30e-3
6 TidyDensity 1 ARIMA Test 1.27 268. 0.970 119. 1.55 4.35e-3
7 tidyAML 1 ARIMA Test 0.951 120. 0.924 158. 1.37 1.54e-3
8 RandomWalker 1 ARIMA Test 0.736 230. 0.566 148. 0.830 2.86e-1
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 [1745|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [1736|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1682|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1653|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1478|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1329|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [936|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [359|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")
