PTB-OEDA / VIEWS2-DensityForecasts

Patrick Brandt VIEWS 2 Density Forecasts

output
pdf_document html_document default default

VIEWS2-DensityForecasts

Patrick T. Brandt

VIEWS 2 Bayesian Density Forecasts

These are the files used to generate the models for the 2023/24 VIEWS Prediction Challenge. These setup and the data and replicate the analysis presented there on Bayesian Density Forecasts. You access the data used directly from here since they are large not included in this repo.

The forthcoming prediction competition article to which this contributes is Hegre, H. et al (Forthcoming), "The 2023/24 VIEWS prediction competition", Journal of Peace Research

The files here have only lightly been edited from what was used originally for the forecasts in October 2023. Additional files are added to cover the post-2021 data and the true future cases into 2024-2025.

1. Main file for the data setup and estimation in R is the Brandt-VIEWS2-Estimation.Rmd. This sets up the data and does the basic model exploration and selection via AIC and some in-sample CRPS computations. This generates the forecast output for 2018.

2. Forecasts for the 2019-2022 periods are handled in separate R scripts:

   1. combined-rolled-2019.R for 2019
   2. combined-rolled-2020.R for 2020
   3. combined-rolled-2021.R for 2021
   4. combined-rolled-2022.R for 2022

3. Organization and reporting of the output for what was submitted to the competition is in reorg.R and forc-org.r scripts (to be called in this order). Final output comes in a file named ForecastDensities_2018-2022.RData.

4. Revised / updated data were provided in May 2024. The earlier analysis was consolidated down into a single script for the extended / updated forecasts. This is what is now in BrandtRevised-2018-2025.R.

The forecasts from these methods are not included here. They are easy to generate and can be done by running the above on a single CPU core in less than 24 hours. So this is not some crazy computational job, nor has it really been optimized (the time could be greatly trimmed with some smart / easy parallelization of the models estimation and forecast generation). But the goal is to create simple, easy to follow replication code here so others can build on it.

Questions to Patrick Brandt at pbrandt_at_utdallas_dot_edu.