07-29, 13:00–13:10 (UTC), Red Track
Time series models with time-varying parameters have become increasingly popular over the years due to their advantages in capturing the dynamics of series of interest. In this context, score-driven models represent a recently developed and powerful framework for modeling time series considering non-Gaussian predictive distributions. In this talk, we present ScoreDrivenModels.jl, a Julia package for modeling, forecasting, and simulating data using score-driven models.
In this talk, we will first provide a brief overview of score-driven models, also known as generalized autoregressive models, based on the paper “Generalized Autoregressive Score Models With Applications” by D. Creal et al. This class of models represents a powerful and flexible tool for dealing with time-series data under non-Gaussian distributions and different autoregressive structures.
After a high-level introduction to the theory, we will present the ScoreDrivenModels.jl package, going through all steps that a user would possibly need to do: model specification, estimation, forecasting, and simulation. We will show how the package is built on top of Distributions.jl, thus allowing easy integration of any distribution, including more obscure ones, for example, a non-zero mean Student’s t which employs the LocationScale abstraction. The package also has a manual for integrating new distributions with small effort.
Finally, we will motivate the use of this class of models by providing examples of applications in different fields, e.g., simulating scenarios for renewable generation and analyzing financial time-series data, while showcasing the use of the package.
For more information on score-driven models, we refer the interested reader to http://www.gasmodel.com.
I am a Master's student at PUC-Rio in Rio de Janeiro, Brazil. I have contributed to some packages from the JuliaOpt and participated in GSOC'19 with the project Dualization.jl. I am also one of the developers of StateSpaceModels.jl.