healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse
Packages Steven P. Sanderson II, MPH - Date: 2025-11-15
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: 160,240
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 2025-11-13 21:11:20, the file was birthed on: 2022-07-02 23:58:17.511888, and at report knit time is 2.951322^{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 | 160240 |
| 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 | 117529 | 0.27 | 5 | 7 | 0 | 50 | 0 |
| r_arch | 117529 | 0.27 | 1 | 7 | 0 | 6 | 0 |
| r_os | 117529 | 0.27 | 7 | 19 | 0 | 24 | 0 |
| package | 0 | 1.00 | 7 | 13 | 0 | 8 | 0 |
| version | 0 | 1.00 | 5 | 17 | 0 | 62 | 0 |
| country | 14997 | 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 | 2025-11-13 | 2023-10-24 | 1810 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1125227.35 | 1491076.70 | 355 | 26719.75 | 309998 | 2355466 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 11323.88 | 21981.89 | 1 | 228.00 | 2910 | 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 | 2025-11-13 21:11:20 | 2023-10-24 08:51:59 | 100809 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 24 | 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
-146.72 -36.46 -11.40 27.26 818.81
Coefficients:
Estimate Std. Error
(Intercept) -1.991e+02 6.164e+01
date 1.206e-02 3.269e-03
lag(value, 1) 1.103e-01 2.336e-02
lag(value, 7) 8.760e-02 2.413e-02
lag(value, 14) 7.960e-02 2.410e-02
lag(value, 21) 7.705e-02 2.417e-02
lag(value, 28) 6.516e-02 2.409e-02
lag(value, 35) 5.804e-02 2.416e-02
lag(value, 42) 5.972e-02 2.434e-02
lag(value, 49) 6.524e-02 2.426e-02
month(date, label = TRUE).L -9.478e+00 5.084e+00
month(date, label = TRUE).Q 1.446e+00 5.015e+00
month(date, label = TRUE).C -1.562e+01 5.026e+00
month(date, label = TRUE)^4 -6.766e+00 5.012e+00
month(date, label = TRUE)^5 -7.748e+00 4.979e+00
month(date, label = TRUE)^6 4.436e-02 5.001e+00
month(date, label = TRUE)^7 -5.442e+00 4.894e+00
month(date, label = TRUE)^8 -4.665e+00 4.849e+00
month(date, label = TRUE)^9 2.436e+00 4.843e+00
month(date, label = TRUE)^10 8.164e-01 4.856e+00
month(date, label = TRUE)^11 -4.143e+00 4.843e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.154e+01 2.237e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.340e+00 2.317e+00
t value Pr(>|t|)
(Intercept) -3.230 0.001263 **
date 3.689 0.000232 ***
lag(value, 1) 4.722 2.52e-06 ***
lag(value, 7) 3.631 0.000291 ***
lag(value, 14) 3.303 0.000976 ***
lag(value, 21) 3.187 0.001461 **
lag(value, 28) 2.705 0.006895 **
lag(value, 35) 2.402 0.016396 *
lag(value, 42) 2.454 0.014235 *
lag(value, 49) 2.689 0.007232 **
month(date, label = TRUE).L -1.864 0.062442 .
month(date, label = TRUE).Q 0.288 0.773162
month(date, label = TRUE).C -3.107 0.001921 **
month(date, label = TRUE)^4 -1.350 0.177225
month(date, label = TRUE)^5 -1.556 0.119813
month(date, label = TRUE)^6 0.009 0.992924
month(date, label = TRUE)^7 -1.112 0.266308
month(date, label = TRUE)^8 -0.962 0.336172
month(date, label = TRUE)^9 0.503 0.614971
month(date, label = TRUE)^10 0.168 0.866507
month(date, label = TRUE)^11 -0.856 0.392348
fourier_vec(date, type = "sin", K = 1, period = 7) -5.160 2.76e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.168 0.001562 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.39 on 1738 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.2333, Adjusted R-squared: 0.2236
F-statistic: 24.04 on 22 and 1738 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( 5 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 10.2218309948049"
[1] "BEST method = 'lin' PATH MEMBER = c( 5 )"
[1] "BEST lin OBJECTIVE FUNCTION = 10.2218309948049"
[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 = 7.39054316240373"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 5 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.39054316240373"
[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 = 10.206755757149"
[1] "BEST method = 'both' PATH MEMBER = c( 5 )"
[1] "BEST both OBJECTIVE FUNCTION = 10.206755757149"
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( 1 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 106.512111718227"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 106.512111718227"
[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 = 118.629707052093"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 118.629707052093"
[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 = 116.305518755073"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 116.305518755073"
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( 1 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 1257.06718666215"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 1257.06718666215"
[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 = 146.664141528897"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 146.664141528897"
[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 = 239.369822109581"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 239.369822109581"
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( 16 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 12.7055468358819"
[1] "BEST method = 'lin' PATH MEMBER = c( 16 )"
[1] "BEST lin OBJECTIVE FUNCTION = 12.7055468358819"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 16 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 7.65296232676752"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 16 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 7.65296232676752"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 16 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 8.98694619703699"
[1] "BEST method = 'both' PATH MEMBER = c( 16 )"
[1] "BEST both OBJECTIVE FUNCTION = 8.98694619703699"
Package: healthyverse
[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 = 34.2412531933398"
[1] "BEST method = 'lin' PATH MEMBER = c( 3 )"
[1] "BEST lin OBJECTIVE FUNCTION = 34.2412531933398"
[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 = 6.31664591371713"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 3 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 6.31664591371713"
[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 = 9.7265148876117"
[1] "BEST method = 'both' PATH MEMBER = c( 3 )"
[1] "BEST both OBJECTIVE FUNCTION = 9.7265148876117"
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 = 155.832310375856"
[1] "BEST method = 'lin' PATH MEMBER = c( 2 )"
[1] "BEST lin OBJECTIVE FUNCTION = 155.832310375856"
[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 = 52.8697031611539"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 2 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 52.8697031611539"
[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 = 63.9055771101315"
[1] "BEST method = 'both' PATH MEMBER = c( 2 )"
[1] "BEST both OBJECTIVE FUNCTION = 63.9055771101315"
Package: tidyAML
[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 = 44.3696797877118"
[1] "BEST method = 'lin' PATH MEMBER = c( 10 )"
[1] "BEST lin OBJECTIVE FUNCTION = 44.3696797877118"
[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 = 8.18180113206421"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 10 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 8.18180113206421"
[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 = 13.1485414891716"
[1] "BEST method = 'both' PATH MEMBER = c( 10 )"
[1] "BEST both OBJECTIVE FUNCTION = 13.1485414891716"
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 = 11.1423185427416"
[1] "BEST method = 'lin' PATH MEMBER = c( 3 )"
[1] "BEST lin OBJECTIVE FUNCTION = 11.1423185427416"
[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 = 2.5148843004701"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 3 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 2.5148843004701"
[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 = 3.18847973177893"
[1] "BEST method = 'both' PATH MEMBER = c( 3 )"
[1] "BEST both OBJECTIVE FUNCTION = 3.18847973177893"
[[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 [1,802 × 50]> <tibble [28 × 50]> <split [1774|28]>
2 healthyR <tibble [1,793 × 50]> <tibble [28 × 50]> <split [1765|28]>
3 healthyR.ts <tibble [1,739 × 50]> <tibble [28 × 50]> <split [1711|28]>
4 healthyverse <tibble [1,710 × 50]> <tibble [28 × 50]> <split [1682|28]>
5 healthyR.ai <tibble [1,535 × 50]> <tibble [28 × 50]> <split [1507|28]>
6 TidyDensity <tibble [1,386 × 50]> <tibble [28 × 50]> <split [1358|28]>
7 tidyAML <tibble [993 × 50]> <tibble [28 × 50]> <split [965|28]>
8 RandomWalker <tibble [416 × 50]> <tibble [28 × 50]> <split [388|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 | 0.6263653 | 104.55831 | 0.6088444 | 154.6322 | 0.7702092 | 0.0961322 |
| healthyR.data | 2 | LM | Test | 0.7151764 | 188.25074 | 0.6951713 | 143.7417 | 0.8224735 | 0.0025983 |
| healthyR.data | 3 | EARTH | Test | 0.6702533 | 97.25623 | 0.6515048 | 177.2687 | 0.8105751 | 0.0090187 |
| healthyR.data | 4 | NNAR | Test | 0.7455827 | 200.02308 | 0.7247270 | 149.0294 | 0.8586736 | 0.0005147 |
| healthyR | 1 | ARIMA | Test | 0.7548705 | 112.63959 | 0.6863639 | 177.3977 | 1.0514138 | 0.0205642 |
| healthyR | 2 | LM | Test | 0.7937890 | 266.26544 | 0.7217504 | 139.8340 | 0.9717826 | 0.0450710 |
| healthyR | 3 | EARTH | Test | 0.8485882 | 193.38121 | 0.7715764 | 133.3505 | 1.2227113 | 0.0339109 |
| healthyR | 4 | NNAR | Test | 0.7625437 | 194.70382 | 0.6933407 | 153.9383 | 0.9678839 | 0.0621554 |
| healthyR.ts | 1 | ARIMA | Test | 0.7345513 | 121.27237 | 0.5803031 | 163.9835 | 1.0028664 | 0.0475689 |
| healthyR.ts | 2 | LM | Test | 0.9110817 | 303.32944 | 0.7197640 | 168.9436 | 1.1306189 | 0.0041613 |
| healthyR.ts | 3 | EARTH | Test | 0.7992055 | 153.53775 | 0.6313806 | 151.3809 | 1.0820263 | 0.0205395 |
| healthyR.ts | 4 | NNAR | Test | 1.0216640 | 399.77347 | 0.8071252 | 182.2319 | 1.2237781 | 0.0106494 |
| healthyverse | 1 | ARIMA | Test | 0.6073118 | 90.38111 | 0.8538870 | 144.7301 | 0.8133636 | 0.1167300 |
| healthyverse | 2 | LM | Test | 0.7784853 | 188.92126 | 1.0945588 | 149.4198 | 0.8981262 | 0.0000971 |
| healthyverse | 3 | EARTH | Test | 0.6803035 | 101.65719 | 0.9565141 | 166.6507 | 0.8788223 | 0.0596485 |
| healthyverse | 4 | NNAR | Test | 0.7519458 | 186.79748 | 1.0572440 | 137.7688 | 0.9146347 | 0.0005821 |
| healthyR.ai | 1 | ARIMA | Test | 0.5149260 | 97.04598 | 0.8438281 | 177.4301 | 0.7038524 | 0.0081149 |
| healthyR.ai | 2 | LM | Test | 0.6762455 | 257.01739 | 1.1081882 | 156.2170 | 0.8032432 | 0.0042448 |
| healthyR.ai | 3 | EARTH | Test | 0.6706520 | 203.41774 | 1.0990220 | 129.1312 | 0.8687384 | 0.0105133 |
| healthyR.ai | 4 | NNAR | Test | 0.6480671 | 217.92350 | 1.0620112 | 157.1391 | 0.7674413 | 0.0046300 |
| TidyDensity | 1 | ARIMA | Test | 0.9759969 | 345.25756 | 0.6076949 | 134.2522 | 1.0862754 | 0.0567669 |
| TidyDensity | 2 | LM | Test | 0.8581685 | 146.44390 | 0.5343302 | 147.1794 | 1.0800026 | 0.0090094 |
| TidyDensity | 3 | EARTH | Test | 1.1804694 | 569.46870 | 0.7350076 | 130.9809 | 1.3284643 | 0.0006641 |
| TidyDensity | 4 | NNAR | Test | 0.9309029 | 236.51608 | 0.5796175 | 143.2679 | 1.1131906 | 0.0044373 |
| tidyAML | 1 | ARIMA | Test | 0.7658128 | 146.59503 | 0.7532331 | 152.3139 | 1.0149978 | 0.0422946 |
| tidyAML | 2 | LM | Test | 0.7587890 | 158.63572 | 0.7463248 | 141.7573 | 0.9545947 | 0.0218389 |
| tidyAML | 3 | EARTH | Test | 1.9942053 | 801.17291 | 1.9614474 | 149.1246 | 2.3219629 | 0.0289668 |
| tidyAML | 4 | NNAR | Test | 0.6950095 | 136.22270 | 0.6835929 | 145.0702 | 0.8994436 | 0.0912264 |
| RandomWalker | 1 | ARIMA | Test | 0.7106099 | 103.58272 | 0.5423400 | 145.3718 | 0.8245327 | 0.1874503 |
| RandomWalker | 2 | LM | Test | 0.8391255 | 158.98362 | 0.6404235 | 147.4285 | 0.9188780 | 0.0004570 |
| RandomWalker | 3 | EARTH | Test | 0.8372669 | 154.72938 | 0.6390050 | 141.1055 | 0.9015402 | 0.0044252 |
| RandomWalker | 4 | NNAR | Test | 0.9913480 | 203.03491 | 0.7566002 | 157.1338 | 1.0891837 | 0.0796963 |
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 0.626 105. 0.609 155. 0.770 0.0961
2 healthyR 4 NNAR Test 0.763 195. 0.693 154. 0.968 0.0622
3 healthyR.ts 1 ARIMA Test 0.735 121. 0.580 164. 1.00 0.0476
4 healthyverse 1 ARIMA Test 0.607 90.4 0.854 145. 0.813 0.117
5 healthyR.ai 1 ARIMA Test 0.515 97.0 0.844 177. 0.704 0.00811
6 TidyDensity 2 LM Test 0.858 146. 0.534 147. 1.08 0.00901
7 tidyAML 4 NNAR Test 0.695 136. 0.684 145. 0.899 0.0912
8 RandomWalker 1 ARIMA Test 0.711 104. 0.542 145. 0.825 0.187
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 [1774|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [1765|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1711|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1682|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1507|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1358|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [965|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [388|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")
