Introducing Modeltime Ensemble: Time Series Forecast Stacking

Written by Matt Dancho on October 13, 2020



I’m SUPER EXCITED to introduce modeltime.ensemble, the time series ensemble extension to modeltime. This tutorial introduces our new R Package, Modeltime Ensemble, which makes it easy to perform stacked forecasts that improve forecast accuracy. If you like what you see, I have an Advanced Time Series Course where you will become the time-series expert for your organization by learning modeltime, modeltime.ensemble, and timetk.


Forecasting and Time Series Software Announcements

Articles on the modeltime and timetk forecasting and time series ecosystem.


Like these articles?
👉 Register to stay in the know 👈
on new cutting-edge R software like modeltime.


Modeltime Ensemble
The time series ensemble extension to Modeltime

Three months ago I introduced modeltime, a new R package that speeds up forecasting experimentation and model selection with Machine Learning (e.g. XGBoost, GLMNET, Prophet, Prophet Boost, ARIMA, and ARIMA Boost).

Fast-forward to now. I’m thrilled to announce the first extension to Modeltime: modeltime.ensemble.

Modeltime Ensemble is a cutting-edge package that integrates 3 competition-winning time series ensembling strategies:

  1. Super Learners (Meta-Learners): Use modeltime_fit_resamples() and ensemble_model_spec() to create super learners (models that learn from the predictions of sub-models)

  2. Weighted Ensembles: Use ensemble_weighted() to create weighted ensembles.

  3. Average Ensembles: Use ensemble_average() to build simple average and median ensembles.

High-Performance Forecasting Stacks

Using these modeltime.ensemble, you can build high-performance forecasting stacks. Here’s a Multi-Level Stack, which won the Kaggle Grupo Bimbo Inventory Demand Forecasting Competition (I teach this technique in my High-Performance Time Series Forecasting Course).

Time Series Forecast Stacking

The Multi-Level Stacked Ensemble that won the Kaggle Grupo Bimbo Inventory Demand Challenge

Ensemble Tutorial
Forecasting Product Sales with Average Ensembles

Today, I’ll cover forecasting Product Sales with Average and Weighted Ensembles, which are fast to implement and can have good performance (although super-learner’s tend to have better performance).

Weighted Stacking with Modeltime Ensemble

Weighted Stacking with Modeltime Ensemble

Ensemble Key Concepts:

The idea is that we have several sub-models (Level 1) that make predictions. We can then take these predictions and combine them using a simple average (mean), median average, or a weighted average:

  • Simple Average: Weights all models with the same proportion. Selects the average for each timestamp. Use ensemble_average(type = "mean").
  • Median Average: No weighting. Selects prediction using the centered value for each time stamp. Use ensemble_average(type = "median").
  • Weighted Average: User defines the weights (loadings). Applies a weighted average at each of the timestamps. Use ensemble_weighted(loadings = c(1, 2, 3, 4)).

More Advanced Ensembles:

The average and weighted ensembles are the simplest approaches to ensembling. One method that Modeltime Ensemble has integrated is Super Learners. We won’t cover these in this tutorial. But, I teach them in my High-Performance Time Series Course. 💪

Getting Started
Let’s kick the tires on modeltime.ensemble

Install modeltime.ensemble.

install.packages("modeltime.ensemble")

Load the following libraries.

# Time Series Modeling and Machine Learning
library(tidymodels)
library(modeltime)
library(modeltime.ensemble)

# Time Series and Data Wrangling
library(timetk)
library(tidyverse)

Get Your Data
Forecasting Product Sales

Our Business objective is to forecast the next 12-weeks of Product Sales given 2-year sales history.

We’ll start with a walmart_sales_weekly time series data set that includes Walmart Product Transactions from several stores, which is a small sample of the dataset from Kaggle Walmart Recruiting - Store Sales Forecasting. We’ll simplify the data set to a univariate time series with columns, “Date” and “Weekly_Sales” from Store 1 and Department 1.

store_1_1_tbl <- walmart_sales_weekly %>%
    filter(id == "1_1") %>%
    select(Date, Weekly_Sales)

store_1_1_tbl
## # A tibble: 143 x 2
##    Date       Weekly_Sales
##    <date>            <dbl>
##  1 2010-02-05       24924.
##  2 2010-02-12       46039.
##  3 2010-02-19       41596.
##  4 2010-02-26       19404.
##  5 2010-03-05       21828.
##  6 2010-03-12       21043.
##  7 2010-03-19       22137.
##  8 2010-03-26       26229.
##  9 2010-04-02       57258.
## 10 2010-04-09       42961.
## # … with 133 more rows

Next, visualize the dataset with the plot_time_series() function. Toggle .interactive = TRUE to get a plotly interactive plot. FALSE returns a ggplot2 static plot.

store_1_1_tbl %>%
    plot_time_series(Date, Weekly_Sales, .smooth_period = "3 months", .interactive = FALSE)

plot of chunk unnamed-chunk-4

Seasonality Evaluation

Let’s do a quick seasonality evaluation to hone in on important features using plot_seasonal_diagnostics().

store_1_1_tbl %>%
    plot_seasonal_diagnostics(
        Date, Weekly_Sales,
        .feature_set = c("week", "month.lbl"),
        .interactive = FALSE
    )

plot of chunk unnamed-chunk-5

We can see that certain weeks and months of the year have higher sales. These anomalies are likely due to events. The Kaggle Competition informed competitors that Super Bowl, Labor Day, Thanksgiving, and Christmas were special holidays. To approximate the events, week number and month may be good features. Let’s come back to this when we preprocess our data.

Train / Test
Split your time series into training and testing sets

Give the objective to forecast 12 weeks of product sales, we use time_series_split() to make a train/test set consisting of 12-weeks of test data (hold out) and the rest for training.

  • Setting assess = "12 weeks" tells the function to use the last 12-weeks of data as the testing set.
  • Setting cumulative = TRUE tells the sampling to use all of the prior data as the training set.
splits <- store_1_1_tbl %>%
    time_series_split(assess = "12 weeks", cumulative = TRUE)

Next, visualize the train/test split.

  • tk_time_series_cv_plan(): Converts the splits object to a data frame
  • plot_time_series_cv_plan(): Plots the time series sampling data using the “date” and “value” columns.
splits %>%
    tk_time_series_cv_plan() %>%
    plot_time_series_cv_plan(Date, Weekly_Sales, .interactive = FALSE)

plot of chunk unnamed-chunk-7

Feature Engineering

We’ll make a number of calendar features using recipes. Most of the heavy lifting is done by timetk::step_timeseries_signature(), which generates a series of common time series features. We remove the ones that won’t help. After dummying we have 74 total columns, 72 of which are engineered calendar features.

recipe_spec <- recipe(Weekly_Sales ~ Date, store_1_1_tbl) %>%
    step_timeseries_signature(Date) %>%
    step_rm(matches("(iso$)|(xts$)|(day)|(hour)|(min)|(sec)|(am.pm)")) %>%
    step_mutate(Date_week = factor(Date_week, ordered = TRUE)) %>%
    step_dummy(all_nominal()) %>%
    step_normalize(contains("index.num"), Date_year)

recipe_spec %>% prep() %>% juice()
## # A tibble: 143 x 74
##    Date       Weekly_Sales Date_index.num Date_year Date_half Date_quarter
##    <date>            <dbl>          <dbl>     <dbl>     <int>        <int>
##  1 2010-02-05       24924.          -1.71     -1.21         1            1
##  2 2010-02-12       46039.          -1.69     -1.21         1            1
##  3 2010-02-19       41596.          -1.67     -1.21         1            1
##  4 2010-02-26       19404.          -1.64     -1.21         1            1
##  5 2010-03-05       21828.          -1.62     -1.21         1            1
##  6 2010-03-12       21043.          -1.59     -1.21         1            1
##  7 2010-03-19       22137.          -1.57     -1.21         1            1
##  8 2010-03-26       26229.          -1.54     -1.21         1            1
##  9 2010-04-02       57258.          -1.52     -1.21         1            2
## 10 2010-04-09       42961.          -1.50     -1.21         1            2
## # … with 133 more rows, and 68 more variables: Date_month <int>,
## #   Date_mweek <int>, Date_week2 <int>, Date_week3 <int>, Date_week4 <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>, Date_month.lbl_08 <dbl>, Date_month.lbl_09 <dbl>,
## #   Date_month.lbl_10 <dbl>, Date_month.lbl_11 <dbl>, Date_week_01 <dbl>,
## #   Date_week_02 <dbl>, Date_week_03 <dbl>, Date_week_04 <dbl>,
## #   Date_week_05 <dbl>, Date_week_06 <dbl>, Date_week_07 <dbl>,
## #   Date_week_08 <dbl>, Date_week_09 <dbl>, Date_week_10 <dbl>,
## #   Date_week_11 <dbl>, Date_week_12 <dbl>, Date_week_13 <dbl>,
## #   Date_week_14 <dbl>, Date_week_15 <dbl>, Date_week_16 <dbl>,
## #   Date_week_17 <dbl>, Date_week_18 <dbl>, Date_week_19 <dbl>,
## #   Date_week_20 <dbl>, Date_week_21 <dbl>, Date_week_22 <dbl>,
## #   Date_week_23 <dbl>, Date_week_24 <dbl>, Date_week_25 <dbl>,
## #   Date_week_26 <dbl>, Date_week_27 <dbl>, Date_week_28 <dbl>,
## #   Date_week_29 <dbl>, Date_week_30 <dbl>, Date_week_31 <dbl>,
## #   Date_week_32 <dbl>, Date_week_33 <dbl>, Date_week_34 <dbl>,
## #   Date_week_35 <dbl>, Date_week_36 <dbl>, Date_week_37 <dbl>,
## #   Date_week_38 <dbl>, Date_week_39 <dbl>, Date_week_40 <dbl>,
## #   Date_week_41 <dbl>, Date_week_42 <dbl>, Date_week_43 <dbl>,
## #   Date_week_44 <dbl>, Date_week_45 <dbl>, Date_week_46 <dbl>,
## #   Date_week_47 <dbl>, Date_week_48 <dbl>, Date_week_49 <dbl>,
## #   Date_week_50 <dbl>, Date_week_51 <dbl>, Date_week_52 <dbl>

Make Sub-Models
Let’s make some sub-models with Modeltime

Now for the fun part! Let’s make some models using functions from modeltime and parsnip.

Auto ARIMA

Here’s the basic Auto ARIMA Model.

  • Model Spec: arima_reg() <– This sets up your general model algorithm and key parameters
  • Set Engine: set_engine("auto_arima") <– This selects the specific package-function to use and you can add any function-level arguments here.
  • Fit Model: fit(Weekly_Sales ~ Date, training(splits)) <– All Modeltime Models require a date column to be a regressor.
model_fit_arima <- arima_reg(seasonal_period = 52) %>%
    set_engine("auto_arima") %>%
    fit(Weekly_Sales ~ Date, training(splits))

model_fit_arima
## parsnip model object
## 
## Fit time:  206ms 
## Series: outcome 
## ARIMA(0,0,1)(0,1,0)[52] 
## 
## Coefficients:
##          ma1
##       0.6704
## s.e.  0.0767
## 
## sigma^2 estimated as 60063672:  log likelihood=-819.37
## AIC=1642.74   AICc=1642.9   BIC=1647.48

Elastic Net

Making an Elastic NET model is easy to do. Just set up your model spec using linear_reg() and set_engine("glmnet"). Note that we have not fitted the model yet (as we did in previous steps).

model_spec_glmnet <- linear_reg(penalty = 0.01, mixture = 0.5) %>%
    set_engine("glmnet")

Next, make a fitted workflow:

  • Start with a workflow()
  • Add a Model Spec: add_model(model_spec_glmnet)
  • Add Preprocessing: add_recipe(recipe_spec %>% step_rm(date)) <– Note that I’m removing the “date” column since Machine Learning algorithms don’t typically know how to deal with date or date-time features
  • Fit the Workflow: fit(training(splits))
wflw_fit_glmnet <- workflow() %>%
    add_model(model_spec_glmnet) %>%
    add_recipe(recipe_spec %>% step_rm(Date)) %>%
    fit(training(splits))

XGBoost

We can fit a XGBoost Model using a similar process as the Elastic Net.

model_spec_xgboost <- boost_tree() %>%
    set_engine("xgboost")

set.seed(123)
wflw_fit_xgboost <- workflow() %>%
    add_model(model_spec_xgboost) %>%
    add_recipe(recipe_spec %>% step_rm(Date)) %>%
    fit(training(splits))

NNETAR

We can use a NNETAR model. Note that add_recipe() uses the full recipe (with the Date column) because this is a Modeltime Model.

model_spec_nnetar <- nnetar_reg(
        seasonal_period = 52,
        non_seasonal_ar = 4,
        seasonal_ar     = 1
    ) %>%
    set_engine("nnetar")

set.seed(123)
wflw_fit_nnetar <- workflow() %>%
    add_model(model_spec_nnetar) %>%
    add_recipe(recipe_spec) %>%
    fit(training(splits))

Prophet w/ Regressors

We’ll build a Prophet Model with Regressors. This uses the Facebook Prophet forecasting algorithm and supplies all of the 72 features as regressors to the model. Note - Because this is a Modeltime Model we need to have a Date Feature in the recipe.

model_spec_prophet <- prophet_reg(
      seasonality_yearly = TRUE
    ) %>%
    set_engine("prophet") 

wflw_fit_prophet <- workflow() %>%
    add_model(model_spec_prophet) %>%
    add_recipe(recipe_spec) %>%
    fit(training(splits))

Sub-Model Evaluation

Let’s take a look at our progress so far. We have 5 models. We’ll put them into a Modeltime Table to organize them using modeltime_table().

submodels_tbl <- modeltime_table(
    model_fit_arima,
    wflw_fit_glmnet,
    wflw_fit_xgboost,
    wflw_fit_nnetar,
    wflw_fit_prophet
)

submodels_tbl
## # Modeltime Table
## # A tibble: 5 x 3
##   .model_id .model     .model_desc            
##       <int> <list>     <chr>                  
## 1         1 <fit[+]>   ARIMA(0,0,1)(0,1,0)[52]
## 2         2 <workflow> GLMNET                 
## 3         3 <workflow> XGBOOST                
## 4         4 <workflow> NNAR(4,1,10)[52]       
## 5         5 <workflow> PROPHET W/ REGRESSORS

We can get the accuracy on the hold-out set using modeltime_accuracy() and table_modeltime_accuracy(). The best model is the Prophet with Regressors with a MAE of 1031.

submodels_tbl %>% 
    modeltime_accuracy(testing(splits)) %>%
    table_modeltime_accuracy(.interactive = FALSE)
Accuracy Table
.model_id .model_desc .type mae mape mase smape rmse rsq
1 ARIMA(0,0,1)(0,1,0)[52] Test 1359.99 6.77 1.02 6.93 1721.47 0.95
2 GLMNET Test 1222.38 6.47 0.91 6.73 1349.88 0.98
3 XGBOOST Test 1089.56 5.22 0.82 5.20 1266.62 0.96
4 NNAR(4,1,10)[52] Test 2529.92 11.68 1.89 10.73 3507.55 0.93
5 PROPHET W/ REGRESSORS Test 1031.53 5.13 0.77 5.22 1226.80 0.98

And, we can visualize the forecasts with modeltime_forecast() and plot_modeltime_forecast().

submodels_tbl %>%
    modeltime_forecast(
        new_data    = testing(splits),
        actual_data = store_1_1_tbl
    ) %>%
    plot_modeltime_forecast(.interactive = FALSE)

plot of chunk unnamed-chunk-17

Build Modeltime Ensembles
This is exciting.

We’ll make Average, Median, and Weighted Ensembles. If you are interested in making Super Learners (Meta-Learner Models that leverage sub-model predictions), I teach this in my new High-Performance Time Series course.

I’ve made it super simple to build an ensemble from a Modeltime Tables. Here’s how to use ensemble_average().

  • Start with your Modeltime Table of Sub-Models
  • Pipe into ensemble_average(type = "mean")

You now have a fitted average ensemble.

# Simple Average Ensemble
ensemble_fit_avg <- submodels_tbl %>%
    ensemble_average(type = "mean")

ensemble_fit_avg
## ── Modeltime Ensemble ───────────────────────────────────────────
## Ensemble of 5 Models (MEAN)
## 
## # Modeltime Table
## # A tibble: 5 x 3
##   .model_id .model     .model_desc            
##       <int> <list>     <chr>                  
## 1         1 <fit[+]>   ARIMA(0,0,1)(0,1,0)[52]
## 2         2 <workflow> GLMNET                 
## 3         3 <workflow> XGBOOST                
## 4         4 <workflow> NNAR(4,1,10)[52]       
## 5         5 <workflow> PROPHET W/ REGRESSORS

We can make median and weighted ensembles just as easily. Note - For the weighted ensemble I’m loading the better performing models higher.

# Simple Median Ensemble
ensemble_fit_med <- submodels_tbl %>%
    ensemble_average("median")

# Higher Loading on Better Models (Test RMSE)
ensemble_fit_wt <- submodels_tbl %>%
    ensemble_weighted(loadings = c(2, 4, 6, 1, 6))

Ensemble Evaluation
Let’s see how we did

We need to have Modeltime Tables that organize our ensembles before we can assess performance. Just use modeltime_table() to organize ensembles just like we did for the Sub-Models.

ensemble_models_tbl <- modeltime_table(
    ensemble_fit_avg,
    ensemble_fit_med,
    ensemble_fit_wt
)

ensemble_models_tbl
## # Modeltime Table
## # A tibble: 3 x 3
##   .model_id .model         .model_desc                  
##       <int> <list>         <chr>                        
## 1         1 <ensemble [5]> ENSEMBLE (MEAN): 5 MODELS    
## 2         2 <ensemble [5]> ENSEMBLE (MEDIAN): 5 MODELS  
## 3         3 <ensemble [5]> ENSEMBLE (WEIGHTED): 5 MODELS

Let’s check out the Accuracy Table using modeltime_accuracy() and table_modeltime_accuracy().

  • From MAE, Ensemble Model ID 1 has 1000 MAE, a 3% improvement over our best submodel (MAE 1031).
  • From RMSE, Ensemble Model ID 3 has 1228, which is on par with our best submodel.
ensemble_models_tbl %>%
    modeltime_accuracy(testing(splits)) %>%
    table_modeltime_accuracy(.interactive = FALSE)
Accuracy Table
.model_id .model_desc .type mae mape mase smape rmse rsq
1 ENSEMBLE (MEAN): 5 MODELS Test 1000.01 4.63 0.75 4.58 1408.68 0.97
2 ENSEMBLE (MEDIAN): 5 MODELS Test 1146.60 5.68 0.86 5.77 1310.30 0.98
3 ENSEMBLE (WEIGHTED): 5 MODELS Test 1056.59 5.15 0.79 5.20 1228.45 0.98

And finally we can visualize the performance of the ensembles.

ensemble_models_tbl %>%
    modeltime_forecast(
        new_data    = testing(splits),
        actual_data = store_1_1_tbl
    ) %>%
    plot_modeltime_forecast(.interactive = FALSE)

plot of chunk unnamed-chunk-22

It gets better
You’ve just scratched the surface, here’s what’s coming…

The modeltime.ensemble package functionality is much more feature-rich than what we’ve covered here (I couldn’t possibly cover everything in this post). 😀

Here’s what I didn’t cover:

  • Super-Learners: We can make use resample predictions from our sub-models as inputs to a meta-learner. This can result is significantly better accuracy (5% improvement is what we achieve in my Time Series Course).

  • Multi-Level Modeling: This is the strategy that won the Grupo Bimbo Inventory Demand Forecasting Challenge where multiple layers of esembles are used.

  • Refitting Sub-Models and Meta-Learners: Refitting is special task that is needed prior to forecasting future data. Refitting requires careful attention to control the sub-model and meta-learner retraining process.

I teach each of these techniques and strategies so you become the time series expert for your organization. Here’s how. 👇

Advanced Time Series Course
Become the times series domain expert in your organization.

Make sure you’re notified when my new Advanced Time Series Forecasting in R course comes out. You’ll learn timetk and modeltime plus the most powerful time series forecasting techiniques available. Become the times series domain expert in your organization.

👉 Advanced Time Series Course.


You will learn:

  • Time Series Preprocessing, Noise Reduction, & Anomaly Detection
  • Feature Engineering using lagged variables & external regressors
  • Hyperparameter Tuning
  • Time Series Cross-Validation
  • Ensembling Multiple Machine Learning & Univariate Modeling Techniques (Competition Winner)
  • NEW - Deep Learning with GluonTS (Competition Winner)
  • and more.

Unlock the High-Performance Time Series Course


Documentation

More information on the modeltime ecosystem can be found in the software documentation Modeltime, Modeltime Ensemble, and Timetk.

Have questions about Modeltime Ensemble?

Make a comment in the chat below. 👇

And, if you plan on using modeltime.ensemble for your business, it’s a no-brainer - Take my Time Series Course.