Forecasting Regime Breakdown

American Political Science Association (APSA) Annual Meeting
Vancouver, 10 – 14 September 2025

David Randahl1, Vilde Lunnan Djuve2, Carl-Henrik Knutsen2

1Department of Peace and Conflict Researach, Uppsala University
1Department of Political Science, University of Oslo

Background

Background

  • Regime breakdowns are rare but consequential events that can catalyze broader political instability, civil conflict, and transitions to democracy or autocracy.

  • Examples include:

    • Coups
    • Popular uprisings
    • Self-coups
    • Guided liberalizations

  • Predicting these events is of major interest to both researchers and policymakers.

  • Despite their importance, regime breakdowns are hard to anticipate, due to the complexity and heterogeneity of underlying processes.

What Do We Mean by Regime Breakdown?

  • Following Geddes et al. (2014), we define a regime as “the set of formal and informal rules for selecting leaders and keeping them in power.”
  • Regime breakdown = fundamental change in these rules.
  • Includes: military coups, popular uprisings, incumbent-guided reforms, and wars that trigger new political orders.
  • Sometimes leads to democratization or autocratization; sometimes results in equally authoritarian replacements.
  • We build on the Historical Regimes Dataset (HRD), which provides regime-level data from 1789 to the present, including precise dates and modes of breakdown.

Research goals

Our goal in this project is twofold:

  1. Provide (reasonably) accurate forecasts of which countries are most at risk of experience regime breakdown

  2. Enhance out collective understanding of what precipitates, regime breakdown, and evaluate existing theories about regime breakdown from a forecasting perspective

Forecasting Regime Breakdown

Modeling Strategy

  • Unit of analysis: All independent Country-years 1816–2023
  • Forecasting target: Binary indicator of (any) regime breakdown in the following year
  • Models: Random forest classifiers
  • We start with a baseline model that includes key predictors of regime breakdown, and then add thematic features to capture specific mechanisms and evaluate whether they improve predictive performance.

Feature Sets and Thematic Models

Baseline Model

  • GDP, GDP per capita, regime duration, democracy level, population, conflict

Thematic Models

  • Protest
  • Economic volatility
  • Political institutions
  • Political change
  • Internal diffusion
  • External diffusion

Training and Evaluation

  • We evaluate our models on their out-of-sample predictive performance in period 1841–2023
  • Specifically, we use an iteratively re-trained rolling window approach:
  • For each year y in 1841–2023:
    • Train the model on data from years 1816 to t-1
    • Make predictions for year t
    • Move the window one year forward and repeat
  • We also tried fixed 30- and 50-year windows, with slightly worse performance.
  • We evaluate model performance using AUC, AUPR, and Brier Skill Scores, compared against the baseline model.
  • We compare the baseline model against naive benchmarks, one that always predicts no regime breakdown, one that predicts based on the historical frequency of regime breakdowns globally, and one using only duration as predictor.

Results

Regime change in aggreagate

Regime change by category

Feature importance

Predictions for 2024

2024 map

Illustrative cases: China

China

Illustrative cases: Russia

Russia

Takeaways

  • Regime breakdowns remain difficult to predict, but progress is possible
  • Key predictors include protest movements and diffusion mechanisms
  • Model offers a tool for early warning, theory evaluation, and stakeholder planning
  • Next steps:
    • Update data to 2024 and make forecasts for 2025
    • Decide on final direction of paper
    • Explore additional features and modeling approaches

Thank you!