Explaining Time-Series Forecasts with Sensitivity Analysis (ahead::dynrmf and external regressors)

Following the post on exact Shapley values for time series explainability, this post illustrates an example of how to use sensitivity analysis to explain time-series forecasts, based on the ahead::dynrmf model and external regressors. What is sensitivity analysis in this context? It’s about evaluating the impact of changes in the external regressors on the time-series forecast.

The post uses the ahead::dynrmf_sensi function to compute the sensitivities, and the ahead::plot_dynrmf_sensitivity function to plot the results.

First, install the package:

devtools::install_github("Techtonique/ahead")

Then, run the following code:

# devtools::install_github("Techtonique/ahead")# install.packages(c("fpp2", "e1071", "patchwork"))library(ahead)library(fpp2)library(patchwork)library(e1071)#' # Example 1: US Consumption vs Incomesensitivity_results_auto <- ahead::dynrmf_sensi(y = fpp2::uschange[, "Consumption"],xreg = fpp2::uschange[, "Income"],h = 10)plot1 <- ahead::plot_dynrmf_sensitivity(sensitivity_results_auto,                            title = "Sensitivity of Consumption to Income (Ridge)",                           y_label = "Effect (ΔConsumption / ΔIncome)")#' # Example 1: US Consumption vs Incomesensitivity_results_auto_svm <- ahead::dynrmf_sensi(  y = fpp2::uschange[, "Consumption"],  xreg = fpp2::uschange[, "Income"],  h = 10,   fit_func = e1071::svm # additional parameter passed to ahead::dynrmf)plot2 <- ahead::plot_dynrmf_sensitivity(sensitivity_results_auto_svm,                                         title = "Sensitivity of Consumption to Income (SVM)",                                        y_label = "Effect (ΔConsumption / ΔIncome)") # Example 2: TV Advertising vs Insurance Quotessensitivity_results_tv <- ahead::dynrmf_sensi( y = fpp2::insurance[, "Quotes"],   xreg = fpp2::insurance[, "TV.advert"],   h = 8 )plot3 <- ahead::plot_dynrmf_sensitivity(sensitivity_results_tv,                           title = "Sensitivity of Insurance Quotes to TV Advertising (Ridge)",                           y_label = "Effect (ΔQuotes / ΔTV.advert)")sensitivity_results_tv_svm <- ahead::dynrmf_sensi(  y = fpp2::insurance[, "Quotes"],  xreg = fpp2::insurance[, "TV.advert"],  h = 8,   fit_func = e1071::svm # additional parameter passed to ahead::dynrmf)plot4 <- ahead::plot_dynrmf_sensitivity(sensitivity_results_tv_svm,                                        title = "Sensitivity of Insurance Quotes to TV Advertising (SVM)",                                        y_label = "Effect (ΔQuotes / ΔTV.advert)")(plot1+plot2)(plot3+plot4)