Working with Panel Data is a common challenge for business analysts. We often have multiple time series (called Time Series Groups) that have overlapping timestamps (panels). These time series may depend on each other and should be modeled together using cross-sectional modeling strategies to take advantage of relationships between correlated time series. The issue becomes, how to evaluate the cross-sectional model over time so we can select the most robust model.
The challenge when working with Panel Data is judging how cross-sectional models will perform over time. A single cross-section is not sufficient to instill confidence. Rather, we need to resample to assess stability of our models prior to model selection.
Modeltime Resample provides a convienent way for generating resample predictions across time for Panel Data, simplifying the model comparison process.
This is an advanced tutorial. Working with Panel Data requires working with multiple time series groups at the same time, and you need to be comfortable setting up the datasets required to generate training sets and forecast sets. I cover working with Panel Data and Time Series Groups in my High-Performance Time Series Course.
Load the following R packages.
library(tidymodels)
library(modeltime)
library(modeltime.resample)
library(timetk)
library(tidyverse)
library(tidyquant)
We’ll use the walmart_sales_weekly
dataset from
timetk
. This contains 7 time series groups, which
correspond to the revenue over time for seven departments in one Walmart
Store.
%>%
walmart_sales_weekly group_by(id) %>%
plot_time_series(Date, Weekly_Sales, .facet_ncol = 3, .interactive = FALSE)
We’ll create 2 datasets that incorporate a grouping variable:
Training Data Set,
data_prepared_tbl
: Dataset that contains
information on the training region for each time series group
Forecast Data Set, future_tbl
:
Dataset that contains information on the forecast region for each time
series group. We’ll extend each time series group by
"3 months"
based on the business forecast needs.
# Full = Training + Forecast Datasets
<- walmart_sales_weekly %>%
full_data_tbl select(id, Date, Weekly_Sales) %>%
# Apply Group-wise Time Series Manipulations
group_by(id) %>%
future_frame(
.date_var = Date,
.length_out = "3 months",
.bind_data = TRUE
%>%
) ungroup() %>%
# Consolidate IDs
mutate(id = fct_drop(id))
# Training Data
<- full_data_tbl %>%
data_prepared_tbl filter(!is.na(Weekly_Sales))
# Forecast Data
<- full_data_tbl %>%
future_tbl filter(is.na(Weekly_Sales))
The first step is to make a resample strategy. Our
business objective is to forecast 3 months so we’ll use
time_series_cv()
with the following strategy:
This generates 6 resample sets.
<- data_prepared_tbl %>%
walmart_tscv time_series_cv(
date_var = Date,
assess = "3 months",
skip = "3 months",
cumulative = TRUE,
slice_limit = 6
)
walmart_tscv
## # Time Series Cross Validation Plan
## # A tibble: 6 x 2
## splits id
## <list> <chr>
## 1 <split [917/84]> Slice1
## 2 <split [833/84]> Slice2
## 3 <split [749/84]> Slice3
## 4 <split [665/84]> Slice4
## 5 <split [581/84]> Slice5
## 6 <split [497/84]> Slice6
We can visualize the resample sets with
plot_time_series_cv_plan()
. They look a little crazy
because there are multiple time series groups. The important thing is to
make sure the red and blue parts line up as expected in relation to our
sampling strategy.
%>%
walmart_tscv tk_time_series_cv_plan() %>%
plot_time_series_cv_plan(Date, Weekly_Sales,
.facet_ncol = 2, .interactive = F)
We’ll create:
1 Recipe: This applies engineered features from calendar variables
3 Fitted Models: Prophet, XGBoost, and Prophet
Boost fitted on the data_prepared_tbl
dataset
We’ll create a recipe that leverages
step_timeseries_signature()
to generate calendar
features.
<- recipe(Weekly_Sales ~ .,
recipe_spec data = training(walmart_tscv$splits[[1]])) %>%
step_timeseries_signature(Date) %>%
step_rm(matches("(.iso$)|(.xts$)|(day)|(hour)|(minute)|(second)|(am.pm)")) %>%
step_mutate(Date_week = factor(Date_week, ordered = TRUE)) %>%
step_dummy(all_nominal(), one_hot = TRUE)
Let’s generate 3 Models: Prophet, XGBoost, and Prophet Boost.
<- workflow() %>%
wflw_fit_prophet add_model(
prophet_reg() %>% set_engine("prophet")
%>%
) add_recipe(recipe_spec) %>%
fit(training(walmart_tscv$splits[[1]]))
<- workflow() %>%
wflw_fit_xgboost add_model(
boost_tree("regression") %>% set_engine("xgboost")
%>%
) add_recipe(recipe_spec %>% step_rm(Date)) %>%
fit(training(walmart_tscv$splits[[1]]))
<- workflow() %>%
wflw_fit_prophet_boost add_model(
prophet_boost(
seasonality_daily = FALSE,
seasonality_weekly = FALSE,
seasonality_yearly = FALSE
%>%
) set_engine("prophet_xgboost")
%>%
) add_recipe(recipe_spec) %>%
fit(training(walmart_tscv$splits[[1]]))
Add the 3 fitted models to a Modeltime Table with
modeltime_table()
.
<- modeltime_table(
model_tbl
wflw_fit_prophet,
wflw_fit_xgboost,
wflw_fit_prophet_boost
)
model_tbl
## # Modeltime Table
## # A tibble: 3 x 3
## .model_id .model .model_desc
## <int> <list> <chr>
## 1 1 <workflow> PROPHET W/ REGRESSORS
## 2 2 <workflow> XGBOOST
## 3 3 <workflow> PROPHET W/ XGBOOST ERRORS
We can make a Panel Data forecast, which forecasts all of the time series groups at once. This method is much more efficient than iteratively performing predictions. However, not all time series models respond well to this approach.
# Calibrate using the Test Sample
<- model_tbl %>%
calibration_tbl modeltime_calibrate(testing(walmart_tscv$splits[[1]]))
# Forecast the Test Sample
<- calibration_tbl %>%
forecast_panel_tbl modeltime_forecast(
new_data = testing(walmart_tscv$splits[[1]]),
actual_data = data_prepared_tbl,
# Keep data allows us keep the ID feature for the time series groups
keep_data = TRUE
)
We can visualize the Panel Data forecasts on a single split. It’s a bit difficult to tell how each model is performing.
%>%
forecast_panel_tbl group_by(id) %>%
plot_modeltime_forecast(
.facet_ncol = 3,
.y_intercept = 0,
.interactive = FALSE,
.title = "Panel Forecasting | 7 Time Series Groups"
)
We’ve made predictions, but this doesn’t tell us how the models will do over time. We need to quantify prediction error. To do so, we’ll evaluate our models using time series resamples. This technique involves making resamples across time series windows and refitting our models to the resample data sets, producing predictions, and quantifying the error from the predictions.
With model_tbl
(models) and walmart_tscv
(resamples) in had, we are ready to iteratively fit and predict each of
the models on each of the resampling plan sets, producing resample
predictions.
<- model_tbl %>%
resample_results modeltime_fit_resamples(
resamples = walmart_tscv,
control = control_resamples(verbose = FALSE)
)
A new column (“.resample_results”) containing the resample
predictions has been added to the original model_tbl
.
resample_results
## # Modeltime Table
## # A tibble: 3 x 4
## .model_id .model .model_desc .resample_results
## <int> <list> <chr> <list>
## 1 1 <workflow> PROPHET W/ REGRESSORS <rsmp[+]>
## 2 2 <workflow> XGBOOST <rsmp[+]>
## 3 3 <workflow> PROPHET W/ XGBOOST ERRORS <rsmp[+]>
With resampled predictions, we can now assess the robustness of our models over time. This will increase our confidence in the stability of our models, enabling us select the best model(s) for the time-varying data.
We can visually evaluate the accuracy with
plot_modeltime_resamples()
. We can see that Prophet Boost
and XGBoost Models have the much lower average RMSE compared to the
other Prophet w/ Regressors Model.
%>%
resample_results plot_modeltime_resamples(
.summary_fn = mean,
.point_size = 3,
.interactive = FALSE
)
We can get an interactive or static table using
modeltime_resample_accuracy()
. I’m interested not only in
the average metric value but also in the variability (standard
deviation). I can get both of these by adding multiple summary functions
using a list()
.
%>%
resample_results modeltime_resample_accuracy(summary_fns = list(mean = mean, sd = sd)) %>%
table_modeltime_accuracy(.interactive = FALSE)
Accuracy Table | |||||||||||||||
.model_id | .model_desc | .type | n | mae_mean | mae_sd | mape_mean | mape_sd | mase_mean | mase_sd | smape_mean | smape_sd | rmse_mean | rmse_sd | rsq_mean | rsq_sd |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | PROPHET W/ REGRESSORS | Resamples | 6 | 6328.68 | 1168.42 | 17.26 | 4.09 | 0.19 | 0.04 | 16.49 | 3.68 | 8792.52 | 1322.96 | 0.95 | 0.02 |
2 | XGBOOST | Resamples | 6 | 4854.34 | 519.97 | 12.58 | 3.50 | 0.14 | 0.02 | 11.87 | 2.44 | 6969.84 | 970.69 | 0.97 | 0.02 |
3 | PROPHET W/ XGBOOST ERRORS | Resamples | 6 | 4677.63 | 747.38 | 13.73 | 4.96 | 0.14 | 0.02 | 12.12 | 3.14 | 6724.22 | 1221.71 | 0.97 | 0.02 |
If we are interested in a single model, we should select either the XGBoost Model or the Prophet Boost Model, which have lower RMSE than the Prophet w/ Regressors Model.
Working with Panel Data can be challenging due to managing multiple models, overlapping time series groups, and multiple resample sets.
Modeltime Resample makes working with Panel Data much easier. We saw how we can evaluate multiple models on varying time series windows. This increased our confidence that selecting either the XGBoost or Prophet Boost models were best for this data.
This is a quick overview of working with Panel Data. To learn how to evaluate Panel Data in-depth, take my High-Performance Time Series Course.