healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse
Packages Steven P. Sanderson II, MPH - Date: 2026-01-16
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: 165,653
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 2026-01-14 23:58:18, the file was birthed on: 2025-10-31 10:47:59.603742, and at report knit time is 1809.17 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 | 165653 |
| 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 | 121929 | 0.26 | 5 | 7 | 0 | 50 | 0 |
| r_arch | 121929 | 0.26 | 1 | 7 | 0 | 6 | 0 |
| r_os | 121929 | 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 | 15433 | 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 | 2026-01-14 | 2023-11-22 | 1872 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1123475.02 | 1485248.20 | 355 | 32599 | 310813 | 2347869 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 11218.23 | 21842.91 | 1 | 223 | 2808 | 11752 | 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 | 2026-01-14 23:58:18 | 2023-11-22 13:13:46 | 104721 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 12H 6M 31S | 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
-148.66 -36.89 -11.26 27.15 822.73
Coefficients:
Estimate Std. Error
(Intercept) -1.619e+02 5.812e+01
date 1.009e-02 3.078e-03
lag(value, 1) 1.071e-01 2.294e-02
lag(value, 7) 8.758e-02 2.369e-02
lag(value, 14) 8.153e-02 2.364e-02
lag(value, 21) 8.492e-02 2.372e-02
lag(value, 28) 6.545e-02 2.364e-02
lag(value, 35) 5.352e-02 2.364e-02
lag(value, 42) 6.618e-02 2.375e-02
lag(value, 49) 6.517e-02 2.366e-02
month(date, label = TRUE).L -9.087e+00 4.879e+00
month(date, label = TRUE).Q -8.334e-01 4.780e+00
month(date, label = TRUE).C -1.469e+01 4.819e+00
month(date, label = TRUE)^4 -7.030e+00 4.860e+00
month(date, label = TRUE)^5 -5.948e+00 4.846e+00
month(date, label = TRUE)^6 8.137e-01 4.883e+00
month(date, label = TRUE)^7 -4.203e+00 4.830e+00
month(date, label = TRUE)^8 -4.171e+00 4.807e+00
month(date, label = TRUE)^9 2.863e+00 4.821e+00
month(date, label = TRUE)^10 8.440e-01 4.837e+00
month(date, label = TRUE)^11 -4.084e+00 4.824e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.084e+01 2.179e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.009e+00 2.253e+00
t value Pr(>|t|)
(Intercept) -2.785 0.005405 **
date 3.278 0.001067 **
lag(value, 1) 4.670 3.23e-06 ***
lag(value, 7) 3.697 0.000225 ***
lag(value, 14) 3.449 0.000575 ***
lag(value, 21) 3.580 0.000352 ***
lag(value, 28) 2.769 0.005680 **
lag(value, 35) 2.264 0.023716 *
lag(value, 42) 2.786 0.005386 **
lag(value, 49) 2.754 0.005942 **
month(date, label = TRUE).L -1.863 0.062674 .
month(date, label = TRUE).Q -0.174 0.861598
month(date, label = TRUE).C -3.049 0.002329 **
month(date, label = TRUE)^4 -1.446 0.148227
month(date, label = TRUE)^5 -1.227 0.219796
month(date, label = TRUE)^6 0.167 0.867675
month(date, label = TRUE)^7 -0.870 0.384280
month(date, label = TRUE)^8 -0.868 0.385731
month(date, label = TRUE)^9 0.594 0.552637
month(date, label = TRUE)^10 0.174 0.861511
month(date, label = TRUE)^11 -0.847 0.397348
fourier_vec(date, type = "sin", K = 1, period = 7) -4.974 7.19e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.110 0.001898 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.18 on 1800 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.2254, Adjusted R-squared: 0.2159
F-statistic: 23.8 on 22 and 1800 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( 1 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 148.920842030399"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 148.920842030399"
[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 = 137.010188276091"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 137.010188276091"
[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 = 141.187057975961"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 141.187057975961"
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( 19 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 10.2934331461317"
[1] "BEST method = 'lin' PATH MEMBER = c( 19 )"
[1] "BEST lin OBJECTIVE FUNCTION = 10.2934331461317"
[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 = 15.6549664147421"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 19 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 15.6549664147421"
[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 = 12.0367547787774"
[1] "BEST method = 'both' PATH MEMBER = c( 19 )"
[1] "BEST both OBJECTIVE FUNCTION = 12.0367547787774"
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( 23 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 4.00552093915195"
[1] "BEST method = 'lin' PATH MEMBER = c( 23 )"
[1] "BEST lin OBJECTIVE FUNCTION = 4.00552093915195"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 23 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.14191586934216"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 23 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.14191586934216"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 23 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 5.28511443053936"
[1] "BEST method = 'both' PATH MEMBER = c( 23 )"
[1] "BEST both OBJECTIVE FUNCTION = 5.28511443053936"
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 = 9.63751037810354"
[1] "BEST method = 'lin' PATH MEMBER = c( 4 )"
[1] "BEST lin OBJECTIVE FUNCTION = 9.63751037810354"
[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 = 7.31628518228973"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 4 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.31628518228973"
[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 = 9.01612863248761"
[1] "BEST method = 'both' PATH MEMBER = c( 4 )"
[1] "BEST both OBJECTIVE FUNCTION = 9.01612863248761"
Package: healthyverse
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 20 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 8.04896123447652"
[1] "BEST method = 'lin' PATH MEMBER = c( 20 )"
[1] "BEST lin OBJECTIVE FUNCTION = 8.04896123447652"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 20 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 6.57848171599127"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 20 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 6.57848171599127"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 20 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 7.22543173337111"
[1] "BEST method = 'both' PATH MEMBER = c( 20 )"
[1] "BEST both OBJECTIVE FUNCTION = 7.22543173337111"
Package: RandomWalker
[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 = 3.33900060604145"
[1] "BEST method = 'lin' PATH MEMBER = c( 24 )"
[1] "BEST lin OBJECTIVE FUNCTION = 3.33900060604145"
[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 = 2.66143364362892"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 24 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 2.66143364362892"
[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 = 2.65487854009529"
[1] "BEST method = 'both' PATH MEMBER = c( 24 )"
[1] "BEST both OBJECTIVE FUNCTION = 2.65487854009529"
Package: tidyAML
[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 = 30.0339398004346"
[1] "BEST method = 'lin' PATH MEMBER = c( 4 )"
[1] "BEST lin OBJECTIVE FUNCTION = 30.0339398004346"
[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 = 17.754634277242"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 4 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 17.754634277242"
[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 = 23.8731210967479"
[1] "BEST method = 'both' PATH MEMBER = c( 4 )"
[1] "BEST both OBJECTIVE FUNCTION = 23.8731210967479"
Package: TidyDensity
[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 = 4.93436515624358"
[1] "BEST method = 'lin' PATH MEMBER = c( 3 )"
[1] "BEST lin OBJECTIVE FUNCTION = 4.93436515624358"
[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 = 2.58519664131672"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 3 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 2.58519664131672"
[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 = 2.84532088576053"
[1] "BEST method = 'both' PATH MEMBER = c( 3 )"
[1] "BEST both OBJECTIVE FUNCTION = 2.84532088576053"
[[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 [1,863 × 50]> <tibble [28 × 50]> <split [1835|28]>
2 healthyR <tibble [1,855 × 50]> <tibble [28 × 50]> <split [1827|28]>
3 healthyR.ts <tibble [1,794 × 50]> <tibble [28 × 50]> <split [1766|28]>
4 healthyverse <tibble [1,757 × 50]> <tibble [28 × 50]> <split [1729|28]>
5 healthyR.ai <tibble [1,597 × 50]> <tibble [28 × 50]> <split [1569|28]>
6 TidyDensity <tibble [1,448 × 50]> <tibble [28 × 50]> <split [1420|28]>
7 tidyAML <tibble [1,055 × 50]> <tibble [28 × 50]> <split [1027|28]>
8 RandomWalker <tibble [478 × 50]> <tibble [28 × 50]> <split [450|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 | 0.7334873 | 145.91329 | 0.6149974 | 148.59092 | 0.8544037 | 0.0017299 |
| healthyR.data | 2 | LM | Test | 0.7221575 | 164.72137 | 0.6054979 | 145.68455 | 0.8545065 | 0.0266586 |
| healthyR.data | 3 | EARTH | Test | 0.8074656 | 161.42420 | 0.6770251 | 156.99346 | 1.0103475 | 0.0053495 |
| healthyR.data | 4 | NNAR | Test | 0.7766743 | 143.73122 | 0.6512078 | 146.98974 | 0.9447595 | 0.0043560 |
| healthyR | 1 | ARIMA | Test | 0.5989769 | 191.06942 | 0.6994818 | 123.54546 | 0.7728230 | 0.0145669 |
| healthyR | 2 | LM | Test | 0.6538337 | 347.09433 | 0.7635432 | 113.39133 | 0.8276689 | 0.0071466 |
| healthyR | 3 | EARTH | Test | 1.9735113 | 1463.79961 | 2.3046552 | 183.39521 | 2.1717887 | 0.0057165 |
| healthyR | 4 | NNAR | Test | 0.6571171 | 369.29053 | 0.7673776 | 116.81450 | 0.8714677 | 0.0076451 |
| healthyR.ts | 1 | ARIMA | Test | 0.8754996 | 99.20711 | 0.8371220 | 151.71733 | 1.1691583 | 0.0046950 |
| healthyR.ts | 2 | LM | Test | 0.9506151 | 106.91124 | 0.9089448 | 130.42440 | 1.2840217 | 0.0544702 |
| healthyR.ts | 3 | EARTH | Test | 0.7586471 | 116.97101 | 0.7253917 | 98.41252 | 0.9886629 | 0.2569398 |
| healthyR.ts | 4 | NNAR | Test | 0.9515170 | 103.52824 | 0.9098071 | 139.64653 | 1.2963465 | 0.0134673 |
| healthyverse | 1 | ARIMA | Test | 0.8697332 | 91.76079 | 0.9994492 | 128.78178 | 1.0517810 | 0.0001699 |
| healthyverse | 2 | LM | Test | 0.9454449 | 120.34996 | 1.0864528 | 128.04375 | 1.1224576 | 0.0077246 |
| healthyverse | 3 | EARTH | Test | 0.8918691 | 178.04572 | 1.0248865 | 90.91918 | 1.0633622 | 0.4144917 |
| healthyverse | 4 | NNAR | Test | 0.9651861 | 110.16833 | 1.1091383 | 140.82241 | 1.1925664 | 0.0283160 |
| healthyR.ai | 1 | ARIMA | Test | 1.1321128 | 113.40491 | 1.1365372 | 182.05103 | 1.3188833 | 0.0396432 |
| healthyR.ai | 2 | LM | Test | 0.9887703 | 113.01504 | 0.9926345 | 146.69351 | 1.1307659 | 0.0626436 |
| healthyR.ai | 3 | EARTH | Test | 1.4833615 | 185.06157 | 1.4891585 | 182.00926 | 1.6564987 | 0.0247856 |
| healthyR.ai | 4 | NNAR | Test | 0.9939146 | 119.02638 | 0.9977989 | 147.00037 | 1.1444930 | 0.0619689 |
| TidyDensity | 1 | ARIMA | Test | 1.0603123 | 152.43770 | 0.5825665 | 178.17528 | 1.1820765 | 0.0788830 |
| TidyDensity | 2 | LM | Test | 1.0682531 | 325.08536 | 0.5869294 | 159.66293 | 1.1715393 | 0.0734074 |
| TidyDensity | 3 | EARTH | Test | 1.0270632 | 134.87003 | 0.5642985 | 143.65374 | 1.2483537 | 0.0299400 |
| TidyDensity | 4 | NNAR | Test | 1.0251686 | 245.64217 | 0.5632575 | 158.48336 | 1.1732880 | 0.0490875 |
| tidyAML | 1 | ARIMA | Test | 1.0018296 | 83.99809 | 0.9554943 | 150.99119 | 1.2136255 | 0.0011730 |
| tidyAML | 2 | LM | Test | 1.2189796 | 164.33243 | 1.1626010 | 145.72908 | 1.4483203 | 0.0298504 |
| tidyAML | 3 | EARTH | Test | 0.8571451 | 84.73636 | 0.8175016 | 103.31774 | 1.0718055 | 0.0144375 |
| tidyAML | 4 | NNAR | Test | 0.9049709 | 132.96131 | 0.8631154 | 116.01854 | 1.1365866 | 0.0056016 |
| RandomWalker | 1 | ARIMA | Test | 1.0328009 | 120.98485 | 0.6390495 | 180.13804 | 1.0989217 | 0.0079241 |
| RandomWalker | 2 | LM | Test | 1.1431245 | 146.13452 | 0.7073126 | 168.17749 | 1.2312323 | 0.0095345 |
| RandomWalker | 3 | EARTH | Test | 1.0545115 | 134.05821 | 0.6524830 | 170.74912 | 1.0964369 | 0.0012894 |
| RandomWalker | 4 | NNAR | Test | 1.1055340 | 148.69670 | 0.6840534 | 163.31336 | 1.1993101 | 0.0031517 |
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 0.733 146. 0.615 149. 0.854 1.73e-3
2 healthyR 1 ARIMA Test 0.599 191. 0.699 124. 0.773 1.46e-2
3 healthyR.ts 3 EARTH Test 0.759 117. 0.725 98.4 0.989 2.57e-1
4 healthyverse 1 ARIMA Test 0.870 91.8 0.999 129. 1.05 1.70e-4
5 healthyR.ai 2 LM Test 0.989 113. 0.993 147. 1.13 6.26e-2
6 TidyDensity 2 LM Test 1.07 325. 0.587 160. 1.17 7.34e-2
7 tidyAML 3 EARTH Test 0.857 84.7 0.818 103. 1.07 1.44e-2
8 RandomWalker 3 EARTH Test 1.05 134. 0.652 171. 1.10 1.29e-3
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 [1835|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [1827|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1766|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1729|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1569|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1420|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [1027|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [450|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")
