healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse Packages
Steven P. Sanderson II, MPH - Date: 2026-09-04
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: 187,725
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 2026-09-02 23:49:26, the file was birthed on: 2025-10-31 10:47:59.603742, and at report knit time is 7353.02 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 | 187725 |
| 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 | 141017 | 0.25 | 5 | 17 | 0 | 54 | 0 |
| r_arch | 141017 | 0.25 | 1 | 7 | 0 | 7 | 0 |
| r_os | 141017 | 0.25 | 7 | 33 | 0 | 38 | 0 |
| package | 0 | 1.00 | 7 | 13 | 0 | 8 | 0 |
| version | 0 | 1.00 | 5 | 17 | 0 | 63 | 0 |
| country | 18362 | 0.90 | 2 | 2 | 0 | 172 | 0 |
Variable type: Date
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| date | 0 | 1 | 2020-11-23 | 2026-09-02 | 2024-03-01 | 2103 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1139803.70 | 1472299.93 | 355 | 46549 | 329166 | 2353487 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 12519.74 | 26381.33 | 1 | 162 | 2741 | 12270 | 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 | 2026-09-02 23:49:26 | 2024-03-01 14:56:14 | 121100 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 12H 15M 9S | 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
-152.07 -38.70 -11.75 28.34 826.41
Coefficients:
Estimate Std. Error
(Intercept) -1.195e+02 4.745e+01
date 7.998e-03 2.499e-03
lag(value, 1) 9.828e-02 2.174e-02
lag(value, 7) 7.077e-02 2.234e-02
lag(value, 14) 7.703e-02 2.218e-02
lag(value, 21) 8.775e-02 2.225e-02
lag(value, 28) 8.186e-02 2.219e-02
lag(value, 35) 3.585e-02 2.223e-02
lag(value, 42) 6.130e-02 2.232e-02
lag(value, 49) 7.409e-02 2.227e-02
month(date, label = TRUE).L -8.646e+00 4.730e+00
month(date, label = TRUE).Q 2.126e+00 4.618e+00
month(date, label = TRUE).C -1.476e+01 4.684e+00
month(date, label = TRUE)^4 -9.781e+00 4.713e+00
month(date, label = TRUE)^5 -5.561e+00 4.663e+00
month(date, label = TRUE)^6 1.710e-01 4.706e+00
month(date, label = TRUE)^7 -2.125e+00 4.624e+00
month(date, label = TRUE)^8 -4.111e+00 4.603e+00
month(date, label = TRUE)^9 2.298e-02 4.625e+00
month(date, label = TRUE)^10 -1.203e-02 4.559e+00
month(date, label = TRUE)^11 -3.839e-01 4.480e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.039e+01 2.058e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.110e+00 2.118e+00
t value Pr(>|t|)
(Intercept) -2.518 0.011878 *
date 3.201 0.001393 **
lag(value, 1) 4.520 6.54e-06 ***
lag(value, 7) 3.168 0.001560 **
lag(value, 14) 3.473 0.000526 ***
lag(value, 21) 3.944 8.28e-05 ***
lag(value, 28) 3.690 0.000230 ***
lag(value, 35) 1.613 0.106885
lag(value, 42) 2.747 0.006066 **
lag(value, 49) 3.327 0.000892 ***
month(date, label = TRUE).L -1.828 0.067718 .
month(date, label = TRUE).Q 0.460 0.645305
month(date, label = TRUE).C -3.152 0.001644 **
month(date, label = TRUE)^4 -2.075 0.038069 *
month(date, label = TRUE)^5 -1.192 0.233209
month(date, label = TRUE)^6 0.036 0.971022
month(date, label = TRUE)^7 -0.460 0.645920
month(date, label = TRUE)^8 -0.893 0.371862
month(date, label = TRUE)^9 0.005 0.996036
month(date, label = TRUE)^10 -0.003 0.997895
month(date, label = TRUE)^11 -0.086 0.931724
fourier_vec(date, type = "sin", K = 1, period = 7) -5.046 4.90e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.357 0.000803 ***
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 60.14 on 2031 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.1952, Adjusted R-squared: 0.1865
F-statistic: 22.39 on 22 and 2031 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( 17 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 7.40034368051499"
[1] "BEST method = 'lin' PATH MEMBER = c( 17 )"
[1] "BEST lin OBJECTIVE FUNCTION = 7.40034368051499"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 17 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 12.3032789629617"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 17 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 12.3032789629617"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 17 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 9.83801555832493"
[1] "BEST method = 'both' PATH MEMBER = c( 17 )"
[1] "BEST both OBJECTIVE FUNCTION = 9.83801555832493"
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( 22 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 7.30502443316129"
[1] "BEST method = 'lin' PATH MEMBER = c( 22 )"
[1] "BEST lin OBJECTIVE FUNCTION = 7.30502443316129"
[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 = 11.4960639391382"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 22 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 11.4960639391382"
[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 = 9.87740414601748"
[1] "BEST method = 'both' PATH MEMBER = c( 22 )"
[1] "BEST both OBJECTIVE FUNCTION = 9.87740414601748"
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( 19 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 13.1149974194161"
[1] "BEST method = 'lin' PATH MEMBER = c( 19 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.1149974194161"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 19 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 10.9626734930599"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 19 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 10.9626734930599"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 19 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 10.6340786267887"
[1] "BEST method = 'both' PATH MEMBER = c( 19 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.6340786267887"
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( 14 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 4.21607474427729"
[1] "BEST method = 'lin' PATH MEMBER = c( 14 )"
[1] "BEST lin OBJECTIVE FUNCTION = 4.21607474427729"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 14 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 8.17239972642686"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 14 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 8.17239972642686"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 14 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 5.33081335098511"
[1] "BEST method = 'both' PATH MEMBER = c( 14 )"
[1] "BEST both OBJECTIVE FUNCTION = 5.33081335098511"
Package: healthyverse
[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 = 13.2828230806306"
[1] "BEST method = 'lin' PATH MEMBER = c( 5 )"
[1] "BEST lin OBJECTIVE FUNCTION = 13.2828230806306"
[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 = 5.47182779653506"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 5 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.47182779653506"
[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 = 6.56280050364373"
[1] "BEST method = 'both' PATH MEMBER = c( 5 )"
[1] "BEST both OBJECTIVE FUNCTION = 6.56280050364373"
Package: RandomWalker
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 1 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 2852.85126060735"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 2852.85126060735"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 1 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 53.332218302048"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 53.332218302048"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 1 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 125.457577789155"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 125.457577789155"
Package: tidyAML
[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 = 8.34678394198462"
[1] "BEST method = 'lin' PATH MEMBER = c( 22 )"
[1] "BEST lin OBJECTIVE FUNCTION = 8.34678394198462"
[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 = 5.101084070555"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 22 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.101084070555"
[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 = 5.16596192691766"
[1] "BEST method = 'both' PATH MEMBER = c( 22 )"
[1] "BEST both OBJECTIVE FUNCTION = 5.16596192691766"
Package: TidyDensity
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 19 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 5.05761890765119"
[1] "BEST method = 'lin' PATH MEMBER = c( 19 )"
[1] "BEST lin OBJECTIVE FUNCTION = 5.05761890765119"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 19 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 5.1635164464915"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 19 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.1635164464915"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 19 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 4.42973117107477"
[1] "BEST method = 'both' PATH MEMBER = c( 19 )"
[1] "BEST both OBJECTIVE FUNCTION = 4.42973117107477"
[[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 [2,091 × 50]> <tibble [28 × 50]> <split [2063|28]>
2 healthyR <tibble [2,085 × 50]> <tibble [28 × 50]> <split [2057|28]>
3 healthyR.ts <tibble [2,021 × 50]> <tibble [28 × 50]> <split [1993|28]>
4 healthyverse <tibble [1,901 × 50]> <tibble [28 × 50]> <split [1873|28]>
5 healthyR.ai <tibble [1,826 × 50]> <tibble [28 × 50]> <split [1798|28]>
6 TidyDensity <tibble [1,679 × 50]> <tibble [28 × 50]> <split [1651|28]>
7 tidyAML <tibble [1,282 × 50]> <tibble [28 × 50]> <split [1254|28]>
8 RandomWalker <tibble [706 × 50]> <tibble [28 × 50]> <split [678|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 | 0.9472332 | 650.70475 | 0.7150394 | 114.02803 | 1.2126982 | 0.0299258 |
| healthyR.data | 2 | LM | Test | 0.8975402 | 554.79536 | 0.6775276 | 127.12341 | 1.1451449 | 0.0018779 |
| healthyR.data | 3 | EARTH | Test | 2.3307883 | 2507.56373 | 1.7594458 | 133.86034 | 2.6245785 | 0.0064303 |
| healthyR.data | 4 | NNAR | Test | 0.8781185 | 565.07871 | 0.6628667 | 125.24932 | 1.1159613 | 0.0300751 |
| healthyR | 1 | ARIMA | Test | 0.7157275 | 275.95276 | 0.6915785 | 117.12457 | 0.8892630 | 0.0397126 |
| healthyR | 2 | LM | Test | 0.7548899 | 243.95437 | 0.7294195 | 125.08418 | 0.9835815 | 0.0000755 |
| healthyR | 3 | EARTH | Test | 1.9038195 | 1392.16696 | 1.8395837 | 133.50801 | 2.1315352 | 0.0611022 |
| healthyR | 4 | NNAR | Test | 0.6856910 | 145.93915 | 0.6625555 | 129.05222 | 0.9196714 | 0.0192564 |
| healthyR.ts | 1 | ARIMA | Test | 0.6217551 | 475.58831 | 0.6936846 | 116.85805 | 0.7844280 | 0.0588297 |
| healthyR.ts | 2 | LM | Test | 0.7062205 | 462.58862 | 0.7879215 | 172.43298 | 0.8518597 | 0.0061547 |
| healthyR.ts | 3 | EARTH | Test | 0.7709714 | 678.48872 | 0.8601634 | 122.57547 | 0.9251031 | 0.0521985 |
| healthyR.ts | 4 | NNAR | Test | 0.6268848 | 241.85230 | 0.6994078 | 160.70228 | 0.8027713 | 0.0260555 |
| healthyverse | 1 | ARIMA | Test | 0.4163765 | 56.92084 | 0.8293629 | 30.40768 | 0.4873999 | 0.0109569 |
| healthyverse | 2 | LM | Test | 0.8253289 | 95.43262 | 1.6439380 | 74.47565 | 0.9592074 | 0.0033460 |
| healthyverse | 3 | EARTH | Test | 0.3905476 | 61.50322 | 0.7779153 | 28.34866 | 0.4630734 | 0.0069388 |
| healthyverse | 4 | NNAR | Test | 1.0098992 | 84.01675 | 2.0115759 | 99.68465 | 1.1211961 | 0.0036483 |
| healthyR.ai | 1 | ARIMA | Test | 0.8892639 | 166.80442 | 1.0391073 | 138.70966 | 1.0393641 | 0.2612336 |
| healthyR.ai | 2 | LM | Test | 0.9458971 | 163.21690 | 1.1052833 | 153.96185 | 1.0805653 | 0.0182220 |
| healthyR.ai | 3 | EARTH | Test | 1.6939776 | 389.96209 | 1.9794175 | 141.86818 | 1.9612350 | 0.2379208 |
| healthyR.ai | 4 | NNAR | Test | 0.9260953 | 168.09401 | 1.0821449 | 152.73519 | 1.0679374 | 0.0000070 |
| TidyDensity | 1 | ARIMA | Test | 0.9128076 | 209.79950 | 0.8604351 | 131.09093 | 1.1142145 | 0.1799499 |
| TidyDensity | 2 | LM | Test | 0.8153827 | 151.64563 | 0.7685999 | 151.50131 | 0.9303623 | 0.0202111 |
| TidyDensity | 3 | EARTH | Test | 2.0563785 | 550.45399 | 1.9383934 | 149.25656 | 2.2921492 | 0.0720139 |
| TidyDensity | 4 | NNAR | Test | 0.8079948 | 152.57305 | 0.7616360 | 156.43974 | 0.9068954 | 0.0147979 |
| tidyAML | 1 | ARIMA | Test | 0.9476315 | 156.41259 | 1.1682035 | 169.02693 | 1.1294280 | 0.1675692 |
| tidyAML | 2 | LM | Test | 1.0135208 | 191.37611 | 1.2494293 | 164.75606 | 1.1881313 | 0.0375460 |
| tidyAML | 3 | EARTH | Test | 0.8907107 | 91.88401 | 1.0980336 | 174.41349 | 1.0671884 | 0.2517910 |
| tidyAML | 4 | NNAR | Test | 0.9536134 | 148.46110 | 1.1755778 | 169.88144 | 1.1248895 | 0.0181122 |
| RandomWalker | 1 | ARIMA | Test | 0.8088110 | 261.65475 | 0.8501192 | 122.12844 | 0.9505473 | 0.0084982 |
| RandomWalker | 2 | LM | Test | 0.6633518 | 106.06526 | 0.6972309 | 158.93288 | 0.7968009 | 0.0015149 |
| RandomWalker | 3 | EARTH | Test | 1.2362083 | 426.79747 | 1.2993448 | 128.28033 | 1.4514051 | 0.0843063 |
| RandomWalker | 4 | NNAR | Test | 0.7446074 | 108.02187 | 0.7826366 | 155.79040 | 0.9260425 | 0.1184855 |
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… 4 NNAR Test 0.878 565. 0.663 125. 1.12 0.0301
2 healthyR 1 ARIMA Test 0.716 276. 0.692 117. 0.889 0.0397
3 healthyR.ts 1 ARIMA Test 0.622 476. 0.694 117. 0.784 0.0588
4 healthyverse 3 EARTH Test 0.391 61.5 0.778 28.3 0.463 0.00694
5 healthyR.ai 1 ARIMA Test 0.889 167. 1.04 139. 1.04 0.261
6 TidyDensity 4 NNAR Test 0.808 153. 0.762 156. 0.907 0.0148
7 tidyAML 3 EARTH Test 0.891 91.9 1.10 174. 1.07 0.252
8 RandomWalker 2 LM Test 0.663 106. 0.697 159. 0.797 0.00151
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 [2063|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [2057|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1993|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1873|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1798|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1651|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [1254|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [678|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")
