Introducing Modeltime: Tidy Time Series Forecasting using Tidymodels

Written by Matt Dancho on June 29, 2020



I’m beyond excited to introduce modeltime, a new time series forecasting package designed to speed up model evaluation, selection, and forecasting. modeltime does this by integrating the tidymodels machine learning ecosystem of packages into a streamlined workflow for tidyverse forecasting. Follow the updated modeltime article to get started with modeltime. If you like what you see, I have an Advanced Time Series Course coming soon (join the waitlist) where you will become a time-series expert for your organization by learning modeltime and timetk.

Modeltime
The forecasting framework for the tidymodels ecosystem

modeltime is a new package designed for rapidly developing and testing time series models using machine learning models, classical models, and automated models. There are three key benefits:

  1. Systematic Workflow for Forecasting. Learn a few key functions like modeltime_table(), modeltime_calibrate(), and modeltime_refit() to develop and train time series models.

  2. Unlocks Tidymodels for Forecasting. Gain the benefit of all or the parsnip models including boost_tree() (XGBoost, C5.0), linear_reg() (GLMnet, Stan, Linear Regression), rand_forest() (Random Forest), and more

  3. New Time Series Boosted Models including Boosted ARIMA (arima_boost()) and Boosted Prophet (prophet_boost()) that can improve accuracy by applying XGBoost model to the errors

Getting Started
Let’s kick the tires on modeltime

Install modeltime.

install.packages("modeltime")

Load the following libraries.

library(tidymodels)
library(modeltime)
library(timetk)   
library(lubridate)
library(tidyverse)

Get Your Data
Forecasting daily bike transactions

We’ll start with a bike_sharing_daily time series data set that includes bike transactions. We’ll simplify the data set to a univariate time series with columns, “date” and “value”.

bike_transactions_tbl <- bike_sharing_daily %>%
  select(dteday, cnt) %>%
  set_names(c("date", "value")) 

bike_transactions_tbl
## # A tibble: 731 x 2
##    date       value
##    <date>     <dbl>
##  1 2011-01-01   985
##  2 2011-01-02   801
##  3 2011-01-03  1349
##  4 2011-01-04  1562
##  5 2011-01-05  1600
##  6 2011-01-06  1606
##  7 2011-01-07  1510
##  8 2011-01-08   959
##  9 2011-01-09   822
## 10 2011-01-10  1321
## # … with 721 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.

bike_transactions_tbl %>%
  plot_time_series(date, value, .interactive = FALSE)

plot of chunk unnamed-chunk-4

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

Next, use time_series_split() to make a train/test set.

  • Setting assess = "3 months" tells the function to use the last 3-months of data as the testing set.
  • Setting cumulative = TRUE tells the sampling to use all of the prior data as the training set.
splits <- bike_transactions_tbl %>%
  time_series_split(assess = "3 months", 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, value, .interactive = FALSE)

plot of chunk unnamed-chunk-6

Modeling
This is exciting.

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

Automatic Models

Automatic models are generally modeling approaches that have been automated. This includes “Auto ARIMA” and “Auto ETS” functions from forecast and the “Prophet” algorithm from prophet. These algorithms have been integrated into modeltime. The process is simple to set up:

  • Model Spec: Use a specification function (e.g. arima_reg(), prophet_reg()) to initialize the algorithm and key parameters
  • Engine: Set an engine using one of the engines available for the Model Spec.
  • Fit Model: Fit the model to the training data

Let’s make several models to see this process in action.

Auto ARIMA

Here’s the basic Auto Arima Model fitting process.

  • 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(value ~ date, training(splits)) <– All modeltime models require a date column to be a regressor.
model_fit_arima <- arima_reg() %>%
  set_engine("auto_arima") %>%
  fit(value ~ date, training(splits))
## frequency = 7 observations per 1 week
model_fit_arima
## parsnip model object
## 
## Fit time:  313ms 
## Series: outcome 
## ARIMA(0,1,3) with drift 
## 
## Coefficients:
##           ma1      ma2      ma3   drift
##       -0.6106  -0.1868  -0.0673  9.3169
## s.e.   0.0396   0.0466   0.0398  4.6225
## 
## sigma^2 estimated as 730568:  log likelihood=-5227.22
## AIC=10464.44   AICc=10464.53   BIC=10486.74

Prophet

Prophet is specified just like Auto ARIMA. Note that I’ve changed to prophet_reg(), and I’m passing an engine-specific parameter yearly.seasonality = TRUE using set_engine().

model_fit_prophet <- prophet_reg() %>%
  set_engine("prophet", yearly.seasonality = TRUE) %>%
  fit(value ~ date, training(splits))

model_fit_prophet
## parsnip model object
## 
## Fit time:  145ms 
## PROPHET Model
## - growth: 'linear'
## - n.changepoints: 25
## - seasonality.mode: 'additive'
## - extra_regressors: 0

Machine Learning Models

Machine learning models are more complex than the automated models. This complexity typically requires a workflow (sometimes called a pipeline in other languages). The general process goes like this:

  • Create Preprocessing Recipe
  • Create Model Specifications
  • Use Workflow to combine Model Spec and Preprocessing, and Fit Model

Preprocessing Recipe

First, I’ll create a preprocessing recipe using recipe() and adding time series steps. The process uses the “date” column to create 45 new features that I’d like to model. These include time-series signature features and fourier series.

recipe_spec <- recipe(value ~ date, training(splits)) %>%
  step_timeseries_signature(date) %>%
  step_rm(contains("am.pm"), contains("hour"), contains("minute"),
          contains("second"), contains("xts")) %>%
  step_fourier(date, period = 365, K = 5) %>%
  step_dummy(all_nominal())

recipe_spec %>% prep() %>% juice()
## # A tibble: 641 x 47
##    date       value date_index.num date_year date_year.iso date_half
##    <date>     <dbl>          <int>     <int>         <int>     <int>
##  1 2011-01-01   985     1293840000      2011          2010         1
##  2 2011-01-02   801     1293926400      2011          2010         1
##  3 2011-01-03  1349     1294012800      2011          2011         1
##  4 2011-01-04  1562     1294099200      2011          2011         1
##  5 2011-01-05  1600     1294185600      2011          2011         1
##  6 2011-01-06  1606     1294272000      2011          2011         1
##  7 2011-01-07  1510     1294358400      2011          2011         1
##  8 2011-01-08   959     1294444800      2011          2011         1
##  9 2011-01-09   822     1294531200      2011          2011         1
## 10 2011-01-10  1321     1294617600      2011          2011         1
## # … with 631 more rows, and 41 more variables: date_quarter <int>,
## #   date_month <int>, date_day <int>, date_wday <int>, date_mday <int>,
## #   date_qday <int>, date_yday <int>, date_mweek <int>, date_week <int>,
## #   date_week.iso <int>, date_week2 <int>, date_week3 <int>, date_week4 <int>,
## #   date_mday7 <int>, date_sin365_K1 <dbl>, date_cos365_K1 <dbl>,
## #   date_sin365_K2 <dbl>, date_cos365_K2 <dbl>, date_sin365_K3 <dbl>,
## #   date_cos365_K3 <dbl>, date_sin365_K4 <dbl>, date_cos365_K4 <dbl>,
## #   date_sin365_K5 <dbl>, date_cos365_K5 <dbl>, 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_wday.lbl_1 <dbl>, date_wday.lbl_2 <dbl>,
## #   date_wday.lbl_3 <dbl>, date_wday.lbl_4 <dbl>, date_wday.lbl_5 <dbl>,
## #   date_wday.lbl_6 <dbl>

With a recipe in-hand, we can set up our machine learning pipelines.

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))
workflow_fit_glmnet <- workflow() %>%
  add_model(model_spec_glmnet) %>%
  add_recipe(recipe_spec %>% step_rm(date)) %>%
  fit(training(splits))

Random Forest

We can fit a Random Forest using a similar process as the Elastic Net.

model_spec_rf <- rand_forest(trees = 500, min_n = 50) %>%
  set_engine("randomForest")

workflow_fit_rf <- workflow() %>%
  add_model(model_spec_rf) %>%
  add_recipe(recipe_spec %>% step_rm(date)) %>%
  fit(training(splits))

New Hybrid Models

I’ve included several hybrid models (e.g. arima_boost() and prophet_boost()) that combine both automated algorithms with machine learning. I’ll showcase prophet_boost() next!

Prophet Boost

The Prophet Boost algorithm combines Prophet with XGBoost to get the best of both worlds (i.e. Prophet Automation + Machine Learning). The algorithm works by:

  1. First modeling the univariate series using Prophet
  2. Using regressors supplied via the preprocessing recipe (remember our recipe generated 45 new features), and regressing the Prophet Residuals with the XGBoost model

We can set the model up using a workflow just like with the machine learning algorithms.

model_spec_prophet_boost <- prophet_boost() %>%
  set_engine("prophet_xgboost", yearly.seasonality = TRUE) 

workflow_fit_prophet_boost <- workflow() %>%
  add_model(model_spec_prophet_boost) %>%
  add_recipe(recipe_spec) %>%
  fit(training(splits))

workflow_fit_prophet_boost
## ══ Workflow [trained] ═════════════════════════════════════════════════════════════════════════════════
## Preprocessor: Recipe
## Model: prophet_boost()
## 
## ── Preprocessor ───────────────────────────────────────────────────────────────────────────────────────
## 4 Recipe Steps
## 
## ● step_timeseries_signature()
## ● step_rm()
## ● step_fourier()
## ● step_dummy()
## 
## ── Model ──────────────────────────────────────────────────────────────────────────────────────────────
## PROPHET w/ XGBoost Errors
## ---
## Model 1: PROPHET
## - growth: 'linear'
## - n.changepoints: 25
## - seasonality.mode: 'additive'
## 
## ---
## Model 2: XGBoost Errors
## 
## xgboost::xgb.train(params = list(eta = 0.3, max_depth = 6, gamma = 0, 
##     colsample_bytree = 1, min_child_weight = 1, subsample = 1), 
##     data = x, nrounds = 15, watchlist = wlist, verbose = 0, objective = "reg:squarederror", 
##     nthread = 1)

The Modeltime Workflow
Speed up model evaluation and selection with modeltime

Modeltime Workflow

The modeltime workflow is designed to speed up model evaluation and selection. Now that we have several time series models, let’s analyze them and forecast the future with the modeltime workflow.

Modeltime Table

The Modeltime Table organizes the models with IDs and creates generic descriptions to help us keep track of our models. Let’s add the models to a modeltime_table().

model_table <- modeltime_table(
  model_fit_arima, 
  model_fit_prophet,
  workflow_fit_glmnet,
  workflow_fit_rf,
  workflow_fit_prophet_boost
) 

model_table
## # Modeltime Table
## # A tibble: 5 x 3
##   .model_id .model     .model_desc              
##       <int> <list>     <chr>                    
## 1         1 <fit[+]>   ARIMA(0,1,3) WITH DRIFT  
## 2         2 <fit[+]>   PROPHET                  
## 3         3 <workflow> GLMNET                   
## 4         4 <workflow> RANDOMFOREST             
## 5         5 <workflow> PROPHET W/ XGBOOST ERRORS

Calibration

Model Calibration is used to quantify error and estimate confidence intervals. We’ll perform model calibration on the out-of-sample data (aka. the Testing Set) with the modeltime_calibrate() function. Two new columns are generated (“.type” and “.calibration_data”), the most important of which is the “.calibration_data”. This includes the actual values, fitted values, and residuals for the testing set.

calibration_table <- model_table %>%
  modeltime_calibrate(testing(splits))

calibration_table
## # Modeltime Table
## # A tibble: 5 x 5
##   .model_id .model     .model_desc               .type .calibration_data
##       <int> <list>     <chr>                     <chr> <list>           
## 1         1 <fit[+]>   ARIMA(0,1,3) WITH DRIFT   Test  <tibble [90 × 4]>
## 2         2 <fit[+]>   PROPHET                   Test  <tibble [90 × 4]>
## 3         3 <workflow> GLMNET                    Test  <tibble [90 × 4]>
## 4         4 <workflow> RANDOMFOREST              Test  <tibble [90 × 4]>
## 5         5 <workflow> PROPHET W/ XGBOOST ERRORS Test  <tibble [90 × 4]>

Forecast (Testing Set)

With calibrated data, we can visualize the testing predictions (forecast).

  • Use modeltime_forecast() to generate the forecast data for the testing set as a tibble.
  • Use plot_modeltime_forecast() to visualize the results in interactive and static plot formats.
calibration_table %>%
  modeltime_forecast(actual_data = bike_transactions_tbl) %>%
  plot_modeltime_forecast(.interactive = FALSE)

plot of chunk unnamed-chunk-16

Accuracy (Testing Set)

Next, calculate the testing accuracy to compare the models.

  • Use modeltime_accuracy() to generate the out-of-sample accuracy metrics as a tibble.
  • Use table_modeltime_accuracy() to generate interactive and static
calibration_table %>%
  modeltime_accuracy() %>%
  table_modeltime_accuracy(.interactive = FALSE)
Accuracy Table
.model_id .model_desc .type mae mape mase smape rmse rsq
1 ARIMA(0,1,3) WITH DRIFT Test 2540.11 474.89 2.74 46.00 3188.09 0.39
2 PROPHET Test 1221.18 365.13 1.32 28.68 1764.93 0.44
3 GLMNET Test 1197.06 340.57 1.29 28.44 1650.87 0.49
4 RANDOMFOREST Test 1309.79 327.88 1.42 30.24 1809.05 0.47
5 PROPHET W/ XGBOOST ERRORS Test 1189.28 332.44 1.28 28.48 1644.25 0.55

Analyze Results

From the accuracy measures and forecast results, we see that:

  • Auto ARIMA model is not a good fit for this data.
  • The best model is Prophet + XGBoost

Let’s exclude the Auto ARIMA from our final model, then make future forecasts with the remaining models.

Refit and Forecast Forward

Refitting is a best-practice before forecasting the future.

  • modeltime_refit(): We re-train on full data (bike_transactions_tbl)
  • modeltime_forecast(): For models that only depend on the “date” feature, we can use h (horizon) to forecast forward. Setting h = "12 months" forecasts then next 12-months of data.
calibration_table %>%
  # Remove ARIMA model with low accuracy
  filter(.model_id != 1) %>%
  
  # Refit and Forecast Forward
  modeltime_refit(bike_transactions_tbl) %>%
  modeltime_forecast(h = "12 months", actual_data = bike_transactions_tbl) %>%
  plot_modeltime_forecast(.interactive = FALSE)

plot of chunk unnamed-chunk-18

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

The modeltime 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:

  • Feature engineering: The art of time series analysis is feature engineering. Modeltime works with cutting-edge time-series preprocessing tools including those in recipes and timetk packages.

  • Hyper parameter tuning: ARIMA models and Machine Learning models can be tuned. There’s a right and a wrong way (and it’s not the same for both types).

  • Scalability: Training multiple time series groups and automation is a huge need area in organizations. You need to know how to scale your analyses to thousands of time series.

  • Strengths and weaknesses: Did you know certain machine learning models are better for trend, seasonality, but not both? Why is ARIMA way better for certain datasets? When will Random Forest and XGBoost fail?

  • Advanced machine learning and deep learning: Recurrent Neural Networks (RRNs) have been crushing time series competitions. Will they work for business data? How can you implement them?

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.

👉 Get notified here: 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 RNNs (Competition Winner)
  • and more.

Signup for the Time Series Course waitlist


Future Work

I’m just getting started with modeltime. The main functionality should not change so you can begin using. Let me know of any issues via GitHub. Regarding future work, here’s a short list of what’s coming over the next few months.

Ensembles and Model Stacking

A top priority on the software roadmap is to include model ensembling, various techniques for combining models to improve forecast results. The plan is to collaborate with the tidymodels team to develop ensembling tools.

More Time Series Algorithms

It’s critical to have a diverse set of algorithms included in modeltime or as extensions to modeltime because this improves the speed of experimentation, model selections, and moving into production. To support extensibility:

Comment on GitHub Issue #5 to let me know what you would like to see or if you have plans to extend modeltime.

Modeling time Resources
With modeltime

Have questions about modeltime?

Make a comment in the chat below. 👇

And, if you plan on using modeltime for your business, it’s a no-brainer - Join my Time Series Course Waitlist (It’s coming, it’s really insane).