healthyverse_tsa
Time Series Analysis, Modeling and Forecasting of the Healthyverse
Packages Steven P. Sanderson II, MPH - Date: 2025-12-16
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: 163,105
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 2025-12-14 23:15:21, the file was birthed on: 2022-07-02 23:58:17.511888, and at report knit time is 3.025928^{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 | 163105 |
| 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 | 119680 | 0.27 | 5 | 7 | 0 | 50 | 0 |
| r_arch | 119680 | 0.27 | 1 | 7 | 0 | 6 | 0 |
| r_os | 119680 | 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 | 15296 | 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 | 2025-12-14 | 2023-11-08 | 1841 |
Variable type: numeric
| skim_variable | n_missing | complete_rate | mean | sd | p0 | p25 | p50 | p75 | p100 | hist |
|---|---|---|---|---|---|---|---|---|---|---|
| size | 0 | 1 | 1124358.31 | 1487385.92 | 355 | 27733 | 310317 | 2352303 | 5677952 | ▇▁▂▁▁ |
| ip_id | 0 | 1 | 11311.04 | 21951.29 | 1 | 235 | 2889 | 11933 | 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 | 2025-12-14 23:15:21 | 2023-11-08 12:45:09 | 103015 |
Variable type: Timespan
| skim_variable | n_missing | complete_rate | min | max | median | n_unique |
|---|---|---|---|---|---|---|
| time | 0 | 1 | 0 | 59 | 12H 6M 50S | 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
-146.90 -36.40 -11.24 26.95 820.08
Coefficients:
Estimate Std. Error
(Intercept) -1.848e+02 6.004e+01
date 1.130e-02 3.182e-03
lag(value, 1) 1.090e-01 2.315e-02
lag(value, 7) 8.856e-02 2.387e-02
lag(value, 14) 7.757e-02 2.382e-02
lag(value, 21) 8.167e-02 2.390e-02
lag(value, 28) 6.767e-02 2.380e-02
lag(value, 35) 5.296e-02 2.383e-02
lag(value, 42) 6.567e-02 2.397e-02
lag(value, 49) 6.141e-02 2.388e-02
month(date, label = TRUE).L -1.043e+01 4.997e+00
month(date, label = TRUE).Q 7.413e-01 4.909e+00
month(date, label = TRUE).C -1.566e+01 4.940e+00
month(date, label = TRUE)^4 -5.919e+00 4.925e+00
month(date, label = TRUE)^5 -6.561e+00 4.887e+00
month(date, label = TRUE)^6 1.330e+00 4.903e+00
month(date, label = TRUE)^7 -4.435e+00 4.840e+00
month(date, label = TRUE)^8 -4.022e+00 4.814e+00
month(date, label = TRUE)^9 2.786e+00 4.827e+00
month(date, label = TRUE)^10 9.308e-01 4.844e+00
month(date, label = TRUE)^11 -4.072e+00 4.830e+00
fourier_vec(date, type = "sin", K = 1, period = 7) -1.116e+01 2.208e+00
fourier_vec(date, type = "cos", K = 1, period = 7) 7.305e+00 2.285e+00
t value Pr(>|t|)
(Intercept) -3.077 0.002120 **
date 3.552 0.000392 ***
lag(value, 1) 4.709 2.68e-06 ***
lag(value, 7) 3.710 0.000214 ***
lag(value, 14) 3.257 0.001149 **
lag(value, 21) 3.416 0.000649 ***
lag(value, 28) 2.843 0.004523 **
lag(value, 35) 2.222 0.026421 *
lag(value, 42) 2.740 0.006197 **
lag(value, 49) 2.572 0.010199 *
month(date, label = TRUE).L -2.087 0.037007 *
month(date, label = TRUE).Q 0.151 0.879982
month(date, label = TRUE).C -3.169 0.001553 **
month(date, label = TRUE)^4 -1.202 0.229635
month(date, label = TRUE)^5 -1.343 0.179587
month(date, label = TRUE)^6 0.271 0.786180
month(date, label = TRUE)^7 -0.916 0.359627
month(date, label = TRUE)^8 -0.835 0.403638
month(date, label = TRUE)^9 0.577 0.563869
month(date, label = TRUE)^10 0.192 0.847630
month(date, label = TRUE)^11 -0.843 0.399349
fourier_vec(date, type = "sin", K = 1, period = 7) -5.056 4.73e-07 ***
fourier_vec(date, type = "cos", K = 1, period = 7) 3.196 0.001416 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 59.25 on 1769 degrees of freedom
(49 observations deleted due to missingness)
Multiple R-squared: 0.2294, Adjusted R-squared: 0.2198
F-statistic: 23.93 on 22 and 1769 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 = 423.837625315235"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 423.837625315235"
[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 = 107.004634711231"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 107.004634711231"
[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 = 247.900790186791"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 247.900790186791"
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( 3 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 23.4232449134407"
[1] "BEST method = 'lin' PATH MEMBER = c( 3 )"
[1] "BEST lin OBJECTIVE FUNCTION = 23.4232449134407"
[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 = 13.6721692244789"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 3 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 13.6721692244789"
[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 = 16.7292893798583"
[1] "BEST method = 'both' PATH MEMBER = c( 3 )"
[1] "BEST both OBJECTIVE FUNCTION = 16.7292893798583"
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( 2 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 357.990367175258"
[1] "BEST method = 'lin' PATH MEMBER = c( 2 )"
[1] "BEST lin OBJECTIVE FUNCTION = 357.990367175258"
[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 = 5.88513067948722"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 2 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.88513067948722"
[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 = 7.48265578371466"
[1] "BEST method = 'both' PATH MEMBER = c( 2 )"
[1] "BEST both OBJECTIVE FUNCTION = 7.48265578371466"
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( 19 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 17.9263765210634"
[1] "BEST method = 'lin' PATH MEMBER = c( 19 )"
[1] "BEST lin OBJECTIVE FUNCTION = 17.9263765210634"
[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 = 13.9850168481418"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 19 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 13.9850168481418"
[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 = 17.5837536385415"
[1] "BEST method = 'both' PATH MEMBER = c( 19 )"
[1] "BEST both OBJECTIVE FUNCTION = 17.5837536385415"
Package: healthyverse
[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 = 9.29712183104311"
[1] "BEST method = 'lin' PATH MEMBER = c( 4 )"
[1] "BEST lin OBJECTIVE FUNCTION = 9.29712183104311"
[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 = 9.2883784774975"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 4 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 9.2883784774975"
[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 = 8.50357086525901"
[1] "BEST method = 'both' PATH MEMBER = c( 4 )"
[1] "BEST both OBJECTIVE FUNCTION = 8.50357086525901"
Package: RandomWalker
[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 = 8.11415196738458"
[1] "BEST method = 'lin' PATH MEMBER = c( 23 )"
[1] "BEST lin OBJECTIVE FUNCTION = 8.11415196738458"
[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 = 5.73752266180947"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 23 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 5.73752266180947"
[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 = 5.683176355396"
[1] "BEST method = 'both' PATH MEMBER = c( 23 )"
[1] "BEST both OBJECTIVE FUNCTION = 5.683176355396"
Package: tidyAML
[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 = 162.988215442403"
[1] "BEST method = 'lin' PATH MEMBER = c( 1 )"
[1] "BEST lin OBJECTIVE FUNCTION = 162.988215442403"
[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 = 33.7071960258099"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 1 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 33.7071960258099"
[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 = 34.1699118765355"
[1] "BEST method = 'both' PATH MEMBER = c( 1 )"
[1] "BEST both OBJECTIVE FUNCTION = 34.1699118765355"
Package: TidyDensity
[1] "CURRNET METHOD: lin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'lin' , seasonal.factor = c( 7 ) ...)"
[1] "CURRENT lin OBJECTIVE FUNCTION = 12.5443842203178"
[1] "BEST method = 'lin' PATH MEMBER = c( 7 )"
[1] "BEST lin OBJECTIVE FUNCTION = 12.5443842203178"
[1] "CURRNET METHOD: nonlin"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'nonlin' , seasonal.factor = c( 7 ) ...)"
[1] "CURRENT nonlin OBJECTIVE FUNCTION = 10.8829543373614"
[1] "BEST method = 'nonlin' PATH MEMBER = c( 7 )"
[1] "BEST nonlin OBJECTIVE FUNCTION = 10.8829543373614"
[1] "CURRNET METHOD: both"
[1] "COPY LATEST PARAMETERS DIRECTLY FOR NNS.ARMA() IF ERROR:"
[1] "NNS.ARMA(... method = 'both' , seasonal.factor = c( 7 ) ...)"
[1] "CURRENT both OBJECTIVE FUNCTION = 12.3864044839651"
[1] "BEST method = 'both' PATH MEMBER = c( 7 )"
[1] "BEST both OBJECTIVE FUNCTION = 12.3864044839651"
[[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 [1,833 × 50]> <tibble [28 × 50]> <split [1805|28]>
2 healthyR <tibble [1,824 × 50]> <tibble [28 × 50]> <split [1796|28]>
3 healthyR.ts <tibble [1,769 × 50]> <tibble [28 × 50]> <split [1741|28]>
4 healthyverse <tibble [1,740 × 50]> <tibble [28 × 50]> <split [1712|28]>
5 healthyR.ai <tibble [1,566 × 50]> <tibble [28 × 50]> <split [1538|28]>
6 TidyDensity <tibble [1,417 × 50]> <tibble [28 × 50]> <split [1389|28]>
7 tidyAML <tibble [1,024 × 50]> <tibble [28 × 50]> <split [996|28]>
8 RandomWalker <tibble [447 × 50]> <tibble [28 × 50]> <split [419|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 | 0.6405336 | 109.74864 | 0.7309688 | 129.53379 | 0.7638890 | 0.0450898 |
| healthyR.data | 2 | LM | Test | 0.6104739 | 185.07381 | 0.6966650 | 105.71054 | 0.7521770 | 0.0005239 |
| healthyR.data | 3 | EARTH | Test | 0.6727750 | 274.14020 | 0.7677622 | 96.25974 | 0.8052706 | 0.0190628 |
| healthyR.data | 4 | NNAR | Test | 0.6625972 | 249.08712 | 0.7561474 | 112.66758 | 0.7802426 | 0.0006441 |
| healthyR | 1 | ARIMA | Test | 0.5709736 | 157.43893 | 0.7879975 | 141.05371 | 0.7146578 | 0.0065708 |
| healthyR | 2 | LM | Test | 0.5620079 | 433.94519 | 0.7756240 | 110.84721 | 0.6845832 | 0.0515447 |
| healthyR | 3 | EARTH | Test | 0.7766663 | 852.79396 | 1.0718729 | 102.51307 | 0.9519201 | 0.0432813 |
| healthyR | 4 | NNAR | Test | 0.5552993 | 264.59067 | 0.7663655 | 125.46593 | 0.6625720 | 0.0852044 |
| healthyR.ts | 1 | ARIMA | Test | 0.6059940 | 123.61115 | 0.7302510 | 139.92696 | 0.7585825 | 0.0300536 |
| healthyR.ts | 2 | LM | Test | 0.7244076 | 180.77247 | 0.8729448 | 134.60775 | 0.8826946 | 0.0041798 |
| healthyR.ts | 3 | EARTH | Test | 0.5433493 | 128.00634 | 0.6547612 | 100.61207 | 0.7170904 | 0.0006068 |
| healthyR.ts | 4 | NNAR | Test | 0.7792081 | 193.97392 | 0.9389820 | 146.57874 | 0.9446083 | 0.0043421 |
| healthyverse | 1 | ARIMA | Test | 0.7371031 | 88.73857 | 0.7556244 | 132.88478 | 0.9706330 | 0.0154141 |
| healthyverse | 2 | LM | Test | 0.8041318 | 137.21844 | 0.8243373 | 121.14916 | 0.9553127 | 0.0000056 |
| healthyverse | 3 | EARTH | Test | 0.7387231 | 86.95169 | 0.7572850 | 134.27604 | 0.9665230 | NA |
| healthyverse | 4 | NNAR | Test | 0.7250974 | 116.86338 | 0.7433169 | 114.56132 | 0.9083967 | 0.0406100 |
| healthyR.ai | 1 | ARIMA | Test | 0.8229141 | 102.41629 | 0.9402023 | 182.19284 | 0.9937573 | 0.1593240 |
| healthyR.ai | 2 | LM | Test | 1.0635789 | 190.22607 | 1.2151685 | 152.91763 | 1.2835648 | 0.0880664 |
| healthyR.ai | 3 | EARTH | Test | 2.4077687 | 864.57993 | 2.7509428 | 125.54800 | 2.8680473 | 0.2612615 |
| healthyR.ai | 4 | NNAR | Test | 1.0403103 | 193.21271 | 1.1885835 | 159.48283 | 1.2415946 | 0.0314801 |
| TidyDensity | 1 | ARIMA | Test | 1.0328558 | 228.74199 | 0.6930347 | 162.64268 | 1.1725809 | 0.1270217 |
| TidyDensity | 2 | LM | Test | 0.9056677 | 136.71350 | 0.6076929 | 162.99855 | 1.0856311 | 0.0080809 |
| TidyDensity | 3 | EARTH | Test | 0.8997841 | 129.61684 | 0.6037451 | 181.71804 | 1.0701766 | 0.0023764 |
| TidyDensity | 4 | NNAR | Test | 1.0069810 | 163.37266 | 0.6756730 | 152.29781 | 1.1918485 | 0.0001570 |
| tidyAML | 1 | ARIMA | Test | 0.5713255 | 112.73479 | 0.7336173 | 147.46575 | 0.7400101 | 0.0143055 |
| tidyAML | 2 | LM | Test | 0.6397208 | 272.07826 | 0.8214412 | 137.83704 | 0.7888353 | 0.0243861 |
| tidyAML | 3 | EARTH | Test | 0.6579586 | 282.71295 | 0.8448596 | 113.79097 | 0.7770538 | 0.1710669 |
| tidyAML | 4 | NNAR | Test | 0.7122028 | 330.05564 | 0.9145125 | 130.99565 | 0.8327698 | 0.0000257 |
| RandomWalker | 1 | ARIMA | Test | 0.7012012 | 178.10163 | 0.7231308 | 157.56714 | 0.7942708 | 0.2007504 |
| RandomWalker | 2 | LM | Test | 0.7628786 | 163.95798 | 0.7867371 | 156.60547 | 0.9044253 | 0.0003460 |
| RandomWalker | 3 | EARTH | Test | 0.8511978 | 215.84725 | 0.8778185 | 174.62093 | 0.9065964 | 0.0345232 |
| RandomWalker | 4 | NNAR | Test | 0.8565578 | 222.85139 | 0.8833461 | 179.35241 | 0.9923136 | 0.0403869 |
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… 2 LM Test 0.610 185. 0.697 106. 0.752 5.24e-4
2 healthyR 4 NNAR Test 0.555 265. 0.766 125. 0.663 8.52e-2
3 healthyR.ts 3 EARTH Test 0.543 128. 0.655 101. 0.717 6.07e-4
4 healthyverse 4 NNAR Test 0.725 117. 0.743 115. 0.908 4.06e-2
5 healthyR.ai 1 ARIMA Test 0.823 102. 0.940 182. 0.994 1.59e-1
6 TidyDensity 3 EARTH Test 0.900 130. 0.604 182. 1.07 2.38e-3
7 tidyAML 1 ARIMA Test 0.571 113. 0.734 147. 0.740 1.43e-2
8 RandomWalker 1 ARIMA Test 0.701 178. 0.723 158. 0.794 2.01e-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 [1805|28]> <mdl_tm_t [1 × 5]>
2 healthyR <tibble> <tibble> <split [1796|28]> <mdl_tm_t [1 × 5]>
3 healthyR.ts <tibble> <tibble> <split [1741|28]> <mdl_tm_t [1 × 5]>
4 healthyverse <tibble> <tibble> <split [1712|28]> <mdl_tm_t [1 × 5]>
5 healthyR.ai <tibble> <tibble> <split [1538|28]> <mdl_tm_t [1 × 5]>
6 TidyDensity <tibble> <tibble> <split [1389|28]> <mdl_tm_t [1 × 5]>
7 tidyAML <tibble> <tibble> <split [996|28]> <mdl_tm_t [1 × 5]>
8 RandomWalker <tibble> <tibble> <split [419|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")
