Replication Data for: "An Ever-changing World? A Forecasting Approach to Assessing Temporal Heterogeneity in the Causal Process of Conflict"
Author: Luuk Boekestein (luuk.boekestein@gmail.com)
Master's thesis, Department of Peace and Conflict Research
Uppsala University, 2026
This repository contains all replication data, code and documentation necessary to reproduce the results of the master's thesis.
├── data/
│ ├── auxiliary # Mapping of country_ids to country names
│ ├── views_data/ # VIEWS data (querysets as parquet files)
├── docs/
│ ├── codebook.xlsx # Codebook describing the data and code in more detail
├── figs/ # All figures seen in the main thesis and in the appendices
├── tables/ # All tables seen in the main thesis and in the appendices
├── src/
│ ├── dynamic.py # Main code for the forecasting approach and dynamic model manager
│ └── hurdleregression.py # Implementation of HurdleRegression model
├── README.md # This file
├── replication.ipynb # Main replication file, which runs all models and produces all figures
└── requirements.txt # pip requirements
To replicate the results of the thesis, Python 3.11 or higher is required, as well as the packages listed in the requirements.txt file.
To set up the environment, run the following command in the terminal:
pip install -r requirements.txtThen, to run the replication, simply run the replication.ipynb file in a Jupyter notebook environment. This will produce all figures seen in the main thesis and in the appendices, which will be saved in the figs/ folder, as well as the table with descriptive statistics as seen in the main thesis, which will be saved in the tables/ folder. The expected runtime of the replication is around 5-10 minutes for the main results (on Apple M3) and around 20-30 minutes for the full notebook including the robustness checks.
The five models used in this thesis are replicated constituent models from the VIEWS fatalities002 model, which can be found on the PRIO-data GitHub repository here. Specific details on constituent models can be found in the model definitions file here, as well as information on their querysets and features.
More information on the specific replication details of the models is summarized in the table below:
| Model name | VIEWS model name | Algorithm | Queryset | Features |
|---|---|---|---|---|
| Narrow | fatalities002_joint_narrow_xgb |
XGBRFRegressor(n_estimators=250) |
link | link |
| Broad | fatalities002_joint_broad_rf |
XGBRFRegressor(n_estimators=250) |
link | link |
| Conflict History | fatalities002_conflict_history_rf |
XGBRFRegressor(n_estimators=250) |
link | link |
| WDI | fatalities002_wdi_rf |
XGBRFRegressor(n_estimators=250) |
link | link |
| VDEM | fatalities002_vdem_hurdle_xgb |
HurdleRegression() |
link | link |
The HurdleRegression model is implemented in the src/hurdleregression.py file, and is an adaptation of the class defined by VIEWS here, which is in turn a derivative of code provided by Geoff Ruddock here.
Using these model specifications, the models are then implemented in the replication.ipynb notebook. More information of the underlying features of each models can also be found in the docs/codebook.xlsx file, which contains a codebook describing the data and the transformations of each feature in more detail. The data for each model is loaded from the data/views_data/ folder, which contains the querysets as parquet files.
The core part of the analysis is the forecasting approach and is implemented in the src/dynamic.py file, which contains the DynamicModelManager class. This class takes care of all model training, forecasting and evaluation over different test sets, and is used in the replication.ipynb notebook to run all models. This DynamicModelManager class manages multiple DynamicModel instances, which are then trained on the sliding sets of training data, and evaluated on blocks of testing data.
The evaluation of all models is done using the get_disaggregated_metrics method of the DynamicModelManager class, which calculates the Mean Squared Log Error for each model-test set combination. The skill score is then calculated using the calculate_skill_score function, which is defined in the replication notebook.
-
VIEWS. 2022. Forecasting Fatalities. Department of Peace and Conflict Research, PRIO. https://uu.diva-portal.org/smash/get/diva2:1667048/FULLTEXT01.pdf.
-
Hegre, Håvard, Sofia Nordenving, Michael Colaresi, Mihai Croicu, and James Dale. 2023. Ensembling and Calibration in VIEWS, Version Fatalities002. Model Documentation Series. VIEWS. https://viewsforecasting.org/wp-content/uploads/VIEWS_documentation_Ensembling_Fatalities002.pdf.
-
Hegre, Håvard, Sofia Nordenving, and James Dale. 2024. Models in VIEWS, Version Fatalities002. Model Documentation Series. VIEWS. https://viewsforecasting.org/wp-content/uploads/VIEWS_documentation_models_Fatalities002.pdf.