vignettes/getting-started-with-modeltime.Rmd
getting-started-with-modeltime.Rmd
Forecasting with tidymodels
made easy! This short tutorial shows how you can use:
arima_reg()
, arima_boost()
, exp_smoothing()
, prophet_reg()
, prophet_boost()
, and morelinear_reg()
, mars()
, svm_rbf()
, rand_forest()
, boost_tree()
and moreTo perform classical time series analysis and machine learning in one framework! See “Model List” for the full list of modeltime
models.
Here’s the general process and where the functions fit.
The Modeltime Workflow
Just follow the modeltime
workflow, which is detailed in 6 convenient steps:
Let’s go through a guided tour to kick the tires on modeltime
.
Load libraries to complete this short tutorial.
library(tidymodels)
library(modeltime)
library(tidyverse)
library(lubridate)
library(timetk)
# This toggles plots from plotly (interactive) to ggplot (static)
interactive <- FALSE
# Data
m750 <- m4_monthly %>% filter(id == "M750")
We can visualize the dataset.
m750 %>%
plot_time_series(date, value, .interactive = interactive)
Let’s split the data into training and test sets using initial_time_split()
# Split Data 80/20
splits <- initial_time_split(m750, prop = 0.9)
We can easily create dozens of forecasting models by combining modeltime
and parsnip
. We can also use the workflows
interface for adding preprocessing! Your forecasting possibilities are endless. Let’s get a few basic models developed:
Important note: Handling Date Features
Modeltime models (e.g. arima_reg()
) are created with a date or date time feature in the model. You will see that most models include a formula like fit(value ~ date, data)
.
Parsnip models (e.g. linear_reg()
) typically should not have date features, but may contain derivatives of dates (e.g. month, year, etc). You will often see formulas like fit(value ~ as.numeric(date) + month(date), data)
.
First, we create a basic univariate ARIMA model using “Auto Arima” using arima_reg()
# Model 1: auto_arima ----
model_fit_arima_no_boost <- arima_reg() %>%
set_engine(engine = "auto_arima") %>%
fit(value ~ date, data = training(splits))
#> frequency = 12 observations per 1 year
Next, we create a boosted ARIMA using arima_boost()
. Boosting uses XGBoost to model the ARIMA errors. Note that model formula contains both a date feature and derivatives of date - ARIMA uses the date - XGBoost uses the derivatives of date as regressors
Normally I’d use a preprocessing workflow for the month features using a function like step_timeseries_signature()
from timetk
to help reduce the complexity of the parsnip formula interface.
# Model 2: arima_boost ----
model_fit_arima_boosted <- arima_boost(
min_n = 2,
learn_rate = 0.015
) %>%
set_engine(engine = "auto_arima_xgboost") %>%
fit(value ~ date + as.numeric(date) + factor(month(date, label = TRUE), ordered = F),
data = training(splits))
#> frequency = 12 observations per 1 year
Next, create an Error-Trend-Season (ETS) model using an Exponential Smoothing State Space model. This is accomplished with exp_smoothing()
.
# Model 3: ets ----
model_fit_ets <- exp_smoothing() %>%
set_engine(engine = "ets") %>%
fit(value ~ date, data = training(splits))
#> frequency = 12 observations per 1 year
We’ll create a prophet
model using prophet_reg()
.
# Model 4: prophet ----
model_fit_prophet <- prophet_reg() %>%
set_engine(engine = "prophet") %>%
fit(value ~ date, data = training(splits))
#> Disabling weekly seasonality. Run prophet with weekly.seasonality=TRUE to override this.
#> Disabling daily seasonality. Run prophet with daily.seasonality=TRUE to override this.
We can model time series linear regression (TSLM) using the linear_reg()
algorithm from parsnip
. The following derivatives of date are used:
as.numeric(date)
month(date)
# Model 5: lm ----
model_fit_lm <- linear_reg() %>%
set_engine("lm") %>%
fit(value ~ as.numeric(date) + factor(month(date, label = TRUE), ordered = FALSE),
data = training(splits))
We can model a Multivariate Adaptive Regression Spline model using mars()
. I’ve modified the process to use a workflow
to standardize the preprocessing of the features that are provided to the machine learning model (mars).
# Model 6: earth ----
model_spec_mars <- mars(mode = "regression") %>%
set_engine("earth")
recipe_spec <- recipe(value ~ date, data = training(splits)) %>%
step_date(date, features = "month", ordinal = FALSE) %>%
step_mutate(date_num = as.numeric(date)) %>%
step_normalize(date_num) %>%
step_rm(date)
wflw_fit_mars <- workflow() %>%
add_recipe(recipe_spec) %>%
add_model(model_spec_mars) %>%
fit(training(splits))
OK, with these 6 models, we’ll show how easy it is to forecast.
The next step is to add each of the models to a Modeltime Table using modeltime_table()
. This step does some basic checking to make sure each of the models are fitted and that organizes into a scalable structure called a “Modeltime Table” that is used as part of our forecasting workflow.
We have 6 models to add. A couple of notes before moving on:
modeltime_table()
will complain (throw an informative error) saying you need to fit()
the model.
models_tbl <- modeltime_table(
model_fit_arima_no_boost,
model_fit_arima_boosted,
model_fit_ets,
model_fit_prophet,
model_fit_lm,
wflw_fit_mars
)
models_tbl
#> # Modeltime Table
#> # A tibble: 6 x 3
#> .model_id .model .model_desc
#> <int> <list> <chr>
#> 1 1 <fit[+]> ARIMA(0,1,1)(0,1,1)[12]
#> 2 2 <fit[+]> ARIMA(0,1,1)(0,1,1)[12] W/ XGBOOST ERRORS
#> 3 3 <fit[+]> ETS(M,A,A)
#> 4 4 <fit[+]> PROPHET
#> 5 5 <fit[+]> LM
#> 6 6 <workflow> EARTH
Calibrating adds a new column, .calibration_data
, with the test predictions and residuals inside. A few notes on Calibration:
calibration_tbl <- models_tbl %>%
modeltime_calibrate(new_data = testing(splits))
calibration_tbl
#> # Modeltime Table
#> # A tibble: 6 x 5
#> .model_id .model .model_desc .type .calibration_da…
#> <int> <list> <chr> <chr> <list>
#> 1 1 <fit[+]> ARIMA(0,1,1)(0,1,1)[12] Test <tibble [31 × 4…
#> 2 2 <fit[+]> ARIMA(0,1,1)(0,1,1)[12] W/ XGBOOS… Test <tibble [31 × 4…
#> 3 3 <fit[+]> ETS(M,A,A) Test <tibble [31 × 4…
#> 4 4 <fit[+]> PROPHET Test <tibble [31 × 4…
#> 5 5 <fit[+]> LM Test <tibble [31 × 4…
#> 6 6 <workflow> EARTH Test <tibble [31 × 4…
There are 2 critical parts to an evaluation.
Visualizing the Test Error is easy to do using the interactive plotly visualization (just toggle the visibility of the models using the Legend).
calibration_tbl %>%
modeltime_forecast(
new_data = testing(splits),
actual_data = m750
) %>%
plot_modeltime_forecast(
.legend_max_width = 25, # For mobile screens
.interactive = interactive
)
From visualizing the test set forecast:
We can use modeltime_accuracy()
to collect common accuracy metrics. The default reports the following metrics using yardstick
functions:
mae()
mape()
mase()
smape()
rmse()
rsq()
These of course can be customized following the rules for creating new yardstick metrics, but the defaults are very useful. Refer to default_forecast_accuracy_metrics()
to learn more.
To make table-creation a bit easier, I’ve included table_modeltime_accuracy()
for outputing results in either interactive (reactable
) or static (gt
) tables.
calibration_tbl %>%
modeltime_accuracy() %>%
table_modeltime_accuracy(
.interactive = interactive
)
Accuracy Table | ||||||||
---|---|---|---|---|---|---|---|---|
.model_id | .model_desc | .type | mae | mape | mase | smape | rmse | rsq |
1 | ARIMA(0,1,1)(0,1,1)[12] | Test | 151.33 | 1.41 | 0.52 | 1.43 | 197.71 | 0.93 |
2 | ARIMA(0,1,1)(0,1,1)[12] W/ XGBOOST ERRORS | Test | 147.04 | 1.37 | 0.50 | 1.39 | 191.84 | 0.93 |
3 | ETS(M,A,A) | Test | 77.00 | 0.73 | 0.26 | 0.73 | 90.27 | 0.98 |
4 | PROPHET | Test | 177.51 | 1.70 | 0.61 | 1.70 | 234.65 | 0.88 |
5 | LM | Test | 629.12 | 6.01 | 2.15 | 5.81 | 657.19 | 0.91 |
6 | EARTH | Test | 709.83 | 6.59 | 2.42 | 6.86 | 782.82 | 0.55 |
From the accuracy metrics:
The final step is to refit the models to the full dataset using modeltime_refit()
and forecast them forward.
refit_tbl <- calibration_tbl %>%
modeltime_refit(data = m750)
refit_tbl %>%
modeltime_forecast(h = "3 years", actual_data = m750) %>%
plot_modeltime_forecast(
.legend_max_width = 25, # For mobile screens
.interactive = interactive
)
The models have all changed! (Yes - this is the point of refitting)
This is the (potential) benefit of refitting.
More often than not refitting is a good idea. Refitting:
min_n = 2
, learn_rate = 0.015
.
My Talk on High-Performance Time Series Forecasting
Time series is changing. Businesses now need 10,000+ time series forecasts every day. This is what I call a High-Performance Time Series Forecasting System (HPTSF) - Accurate, Robust, and Scalable Forecasting.
You need to provide a “High-Performance Time Series Forecasting System” (HPTSF System). High-Performance Forecasting Systems will save companies MILLIONS of dollars.
I teach how to build a HPTFS System in my High-Performance Time Series Forecasting Course. If interested in learning Scalable High-Performance Forecasting Strategies then take my course. You will learn:
Modeltime
- 30+ Models (Prophet, ARIMA, XGBoost, Random Forest, & many more)GluonTS
(Competition Winners)