Ensemble Algorithms for Time Series Forecasting with Modeltime
A modeltime extension that that implements
ensemble forecasting methods including model
averaging, weighted averaging, and stacking. Let’s go through a guided
tour to kick the tires on modeltime.ensemble.
We’ll perform the simplest type of forecasting: Using a simple average of the forecasted models.
Note that modeltime.ensemble has capabilities for more
sophisticated model ensembling using:
Load libraries to complete this short tutorial.
# Time Series ML
library(tidymodels)
library(modeltime)
library(modeltime.ensemble)
# Core
library(tidyverse)
library(timetk)
interactive <- FALSEWe’ll use the m750 dataset that comes with
modeltime.ensemble. We can visualize the dataset.
m750 %>%
plot_time_series(date, value, .color_var = id, .interactive = interactive)
We’ll split into a training and testing set.
splits <- time_series_split(m750, assess = "2 years", cumulative = TRUE)
splits %>%
tk_time_series_cv_plan() %>%
plot_time_series_cv_plan(date, value, .interactive = interactive)
Once the data has been collected, we can move into modeling.
We’ll create a Feature Engineering Recipe that can be applied to the data to create features that machine learning models can key in on. This will be most useful for the Elastic Net (Model 3).
recipe_spec <- recipe(value ~ date, training(splits)) %>%
step_timeseries_signature(date) %>%
step_rm(matches("(.iso$)|(.xts$)")) %>%
step_normalize(matches("(index.num$)|(_year$)")) %>%
step_dummy(all_nominal()) %>%
step_fourier(date, K = 1, period = 12)
recipe_spec %>% prep() %>% juice()#> # A tibble: 282 x 42
#> date value date_index~1 date_~2 date_~3 date_~4 date_~5 date_~6 date_~7
#> <date> <dbl> <dbl> <dbl> <int> <int> <int> <int> <int>
#> 1 1990-01-01 6370 -1.72 -1.66 1 1 1 1 0
#> 2 1990-02-01 6430 -1.71 -1.66 1 1 2 1 0
#> 3 1990-03-01 6520 -1.70 -1.66 1 1 3 1 0
#> 4 1990-04-01 6580 -1.69 -1.66 1 2 4 1 0
#> 5 1990-05-01 6620 -1.67 -1.66 1 2 5 1 0
#> 6 1990-06-01 6690 -1.66 -1.66 1 2 6 1 0
#> 7 1990-07-01 6000 -1.65 -1.66 2 3 7 1 0
#> 8 1990-08-01 5450 -1.64 -1.66 2 3 8 1 0
#> 9 1990-09-01 6480 -1.62 -1.66 2 3 9 1 0
#> 10 1990-10-01 6820 -1.61 -1.66 2 4 10 1 0
#> # ... with 272 more rows, 33 more variables: date_minute <int>,
#> # date_second <int>, date_hour12 <int>, date_am.pm <int>, date_wday <int>,
#> # date_mday <int>, date_qday <int>, date_yday <int>, date_mweek <int>,
#> # date_week <int>, date_week2 <int>, date_week3 <int>, date_week4 <int>,
#> # date_mday7 <int>, date_month.lbl_01 <dbl>, date_month.lbl_02 <dbl>,
#> # date_month.lbl_03 <dbl>, date_month.lbl_04 <dbl>, date_month.lbl_05 <dbl>,
#> # date_month.lbl_06 <dbl>, date_month.lbl_07 <dbl>, ...First, we’ll make an ARIMA model using Auto ARIMA.
model_spec_arima <- arima_reg() %>%
set_engine("auto_arima")
wflw_fit_arima <- workflow() %>%
add_model(model_spec_arima) %>%
add_recipe(recipe_spec %>% step_rm(all_predictors(), -date)) %>%
fit(training(splits))Next, we’ll make a Prophet Model.
model_spec_prophet <- prophet_reg() %>%
set_engine("prophet")
wflw_fit_prophet <- workflow() %>%
add_model(model_spec_prophet) %>%
add_recipe(recipe_spec %>% step_rm(all_predictors(), -date)) %>%
fit(training(splits))Third, we’ll make an Elastic Net Model using glmnet.
model_spec_glmnet <- linear_reg(
mixture = 0.9,
penalty = 4.36e-6
) %>%
set_engine("glmnet")
wflw_fit_glmnet <- workflow() %>%
add_model(model_spec_glmnet) %>%
add_recipe(recipe_spec %>% step_rm(date)) %>%
fit(training(splits))With the models created, we can can create an Ensemble Average Model using a simple Mean Average.
Create a Modeltime Table using the modeltime
package.
m750_models <- modeltime_table(
wflw_fit_arima,
wflw_fit_prophet,
wflw_fit_glmnet
)
m750_models#> # Modeltime Table
#> # A tibble: 3 x 3
#> .model_id .model .model_desc
#> <int> <list> <chr>
#> 1 1 <workflow> ARIMA(0,1,1)(0,1,1)[12]
#> 2 2 <workflow> PROPHET
#> 3 3 <workflow> GLMNETThen use ensemble_average() to turn that Modeltime Table
into a Modeltime Ensemble. This is a
fitted ensemble specification containing the ingredients to
forecast future data and be refitted on data sets using the 3
submodels.
ensemble_fit <- m750_models %>%
ensemble_average(type = "mean")
ensemble_fit#> -- Modeltime Ensemble -------------------------------------------
#> Ensemble of 3 Models (MEAN)
#>
#> # Modeltime Table
#> # A tibble: 3 x 3
#> .model_id .model .model_desc
#> <int> <list> <chr>
#> 1 1 <workflow> ARIMA(0,1,1)(0,1,1)[12]
#> 2 2 <workflow> PROPHET
#> 3 3 <workflow> GLMNETTo forecast, just follow the Modeltime Workflow.
# Calibration
calibration_tbl <- modeltime_table(
ensemble_fit
) %>%
modeltime_calibrate(testing(m750_splits))
# Forecast vs Test Set
calibration_tbl %>%
modeltime_forecast(
new_data = testing(m750_splits),
actual_data = m750
) %>%
plot_modeltime_forecast(.interactive = interactive)
Once satisfied with our ensemble model, we can
modeltime_refit() on the full data set and forecast forward
gaining the confidence intervals in the process.
refit_tbl <- calibration_tbl %>%
modeltime_refit(m750)
refit_tbl %>%
modeltime_forecast(
h = "2 years",
actual_data = m750
) %>%
plot_modeltime_forecast(.interactive = interactive)
This was a very short tutorial on the simplest type of forecasting, but there’s a lot more to learn.
Become the forecasting expert for your organization
High-Performance Time Series Course
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.
High-Performance Forecasting Systems will save companies by improving accuracy and scalability. Imagine what will happen to your career if you can provide your organization a “High-Performance Time Series Forecasting System” (HPTSF System).
I teach how to build a HPTFS System in my High-Performance Time Series Forecasting Course. You will learn:
Modeltime - 30+ Models (Prophet, ARIMA, XGBoost, Random
Forest, & many more)GluonTS
(Competition Winners)Become the Time Series Expert for your organization.