Time Series Regression with Intervention Analysis
Loading...
Date
Authors
Musci Hood, Jennifer Kay
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This analysis is focused on time series modeling utilizing the ARIMA model, the ARIMA model with
Intervention, and an extension of Dense Neural Networks for Intervention. The theory behind each
method is explored in detail throughout the following chapters. In addition, the methods are applied to a
constructed dataset reflecting monthly sales of fifty automobiles from January 2007 until February 2021.
This analysis uses mean absolute percentage error, root mean square error, and mean square relative error
to compare model results between the three methods.
Overall, it was determined that the ARIMA model with Intervention performed better than the traditional
ARIMA model when a pulse-response or step-response intervention is present. In addition, the ARIMA
with Intervention Model and Dense Neural Network Model Extended for Intervention performed similarly
for models where a pulse-response or step-response intervention is present. When determining which
model to use, the benefits and drawbacks of dense neural networks are important to note. The benefits
are that dense neural networks can model non-linear functions, the predictor data does not need to be
stationary, and multiple predictor variables can be utilized within the model. On the contrary, dense neural
networks are prone to over-fitting, underlying trends are not modeled as well as they are in traditional
ARIMA models, and predictive capability for out-of-sample observations can be poor, due to nonlinear
forecasting. The analysis in this research was completed with the combination of R version 4.0.3 and
RStudio.