Extreme weather impacts do not improve conflict predictions in Africa

Most prior quantitative climate–conflict research focuses on temperature and precipitation, leaving the role of biophysical extreme weather impacts (droughts, floods, crop production shocks, tropical cyclones) less explored, despite their expected intensification under climate change and rising exposure and vulnerability. Existing work highlights complex and context-specific links between disasters and conflict dynamics; extreme events may catalyze peace in some settings and exacerbate violence in others, and in many cases have no discernible impact. Forecasting need not assume universal effects—extreme weather could still improve predictions if their impacts matter under identifiable conditions. This study asks whether information on extreme weather impacts can improve conflict predictions beyond well-known predictors (socioeconomic indicators, governance, and conflict history). Using high-resolution impact data from state-of-the-art climate impact models, the study explicitly tests with a non-parametric Generalized Random Forest (GRF) whether adding extreme weather impacts improves forecasting accuracy. Prediction-focused analysis allows assessing the total additional predictive information from extreme weather impacts as a group, accounting for non-linear and heterogeneous relationships, and avoids overemphasizing variables with causal but weak predictive value. Prior projection-focused work has been limited in scope of variables and often relies on single impact indicators; this study instead evaluates whether a broad set of extreme weather impacts adds predictive value to a fully specified conflict prediction model across Africa. The authors contribute to climate–conflict and forecasting literatures by integrating ISIMIP2a-based extreme weather impact data into a GRF framework and testing improvements over established predictors, examining both conflict incidence and the potential for predicting onsets versus persistence.

The literature has largely emphasized temperature and precipitation as climate–conflict drivers (e.g., Burke et al. 2015; Hsiang & Burke 2014), with fewer studies considering biophysical impacts such as droughts, floods, and crop shocks, typically one at a time (e.g., Von Uexkull et al. 2016; Ide et al. 2021; Vesco et al. 2021). Empirical findings on disasters and conflict dynamics are mixed and context-dependent (Ide 2023; Gaillard et al. 2008). Long-term projection studies incorporating climate-sensitive variables (e.g., Hoch et al. 2021; Witmer et al. 2017) often include limited socioeconomic and governance covariates due to scenario constraints, limiting tests of added predictive value. Other predictive efforts report modest or null gains from adding climatic or disaster data (Schleussner et al. 2016; Perry 2013; Linke et al. 2022; Bazzi et al. 2022). The field agrees that climate effects on conflict are mostly indirect and heterogeneous, conditional on socioeconomic and political context (Mach et al. 2019; von Uexkull & Buhaug 2021). Against this backdrop, the present study tests whether a comprehensive set of extreme weather impact indicators provides unique predictive information beyond strong baseline predictors (conflict history, governance, socioeconomic conditions).

Study area and period: Mainland Africa. Outcomes: annual conflict incidence (binary: any event vs. none) at varying administrative levels. Conflict data from UCDP (battles linked to conflicts with >25 deaths/year; 1994–2012 analysis window) and ACLED (battles; riots/protests; other events; 2002–2012 due to coverage). To ensure comparable lag structures, ACLED-based models start in 2002 (five-year lags available). Spatial units: GADM administrative levels—national (level 0), first subnational (level 1), and second subnational (level 2; 6269 polygons), to mitigate aggregation bias and capture different spatial patterns. Predictors grouped into four sets: (1) Conflict history: for each of the prior 5 years, counts of events and deaths by conflict type (UCDP battles; ACLED battles; ACLED riots/protests; ACLED other). For UCDP also battle and civilian deaths. National-level displacement: total displaced and newly displaced by conflict in prior year. (2) Socioeconomic: five-year moving averages and 15-year growth where applicable for crop production (soy, maize, wheat, rice yields), infant mortality (static, 2000), land use (cropland share in 2000; LUH2 averages and trends for pasture, irrigated/rainfed/total cropland), population size/density/urbanization (HYDE), area, GDP and GDP per capita (levels and growth; Kummu et al. 2018). (3) Governance: country-level World Governance Indicators (government effectiveness, rule of law, control of corruption, political stability, regulatory quality, voice & accountability) and V-Dem liberal democracy index; subnational: count of ethnic homelands per administrative unit. (4) Extreme weather impacts (from ISIMIP2a and related datasets, harmonized): - Floods: inundation areas from CaMa-Flood driven by WaterGAP2 with PGFv2 forcing; exposure aggregated to admin units by land area and by population; includes standard and 100-year (or FLOPROS-standard) events. - Droughts: monthly soil moisture below 10th percentile; aggregated by population and populated area; five-year lags. - Tropical cyclones: exposure to 34/64/96-knot winds (TCE-DAT), aggregated by population and populated area over prior five years. - Crop production shocks: count in past five years of years with yield deviations beyond ±20% from long-term mean, separately for positive/negative shocks. - Disaster-induced displacement: nationally reported newly displaced by “natural” disasters (includes some non-weather events). Forecasting model: Generalized Random Forest (GRF) with 2000 trees, using “honesty” (separate subsamples for split selection and leaf estimation) and internal cross-validation to reduce overfitting and provide unbiased out-of-sample predictions. Forecast horizon: one-year-ahead conflict incidence. Evaluation metrics: area under the ROC curve (AUC-ROC) and area under the Precision-Recall curve (AUC-PR). ROC facilitates comparison across outcomes (random baseline AUC=0.5); PR is more informative under class imbalance (about 5% conflict incidence at level 2) and emphasizes correct positive predictions. Analytical comparisons: - Full model vs. leave-one-group-out to assess unique predictive contribution of each group. - Single-group models to assess standalone predictive strength. - Robustness across administrative levels and conflict datasets (UCDP vs ACLED). - Subsamples by recent conflict status (places with and without conflict in prior 5 years) to probe onset/cessation predictive value. Correlation and variance decomposition: regress leading socioeconomic predictors on extreme weather impacts to assess shared variance (R^2), with and without place fixed effects to separate spatial from temporal co-variation. Decompose model prediction variance into spatial (place fixed effects) versus temporal (year fixed effects) components to assess whether predictive information is predominantly spatial or temporal.

- Extreme weather impacts (EWI) do not add unique predictive information to fully specified models. Adding EWI to models with conflict history, governance, and socioeconomic predictors does not improve AUC-ROC or AUC-PR across national and subnational levels and across conflict datasets. - Nonetheless, EWI alone predict conflict better than chance. Single-group models show EWI have modest standalone predictive power but are weaker than conflict history, governance, or socioeconomic groups. - Conflict history is the strongest and most uniquely informative predictor. Removing conflict history from the full model substantially degrades performance; prior conflict robustly predicts future conflict incidence. - Governance indicators both perform well alone and add unique information beyond other groups. - Socioeconomic indicators perform comparably to conflict history in standalone models but add no unique information when other groups are included. - Results hold for both conflict persistence and onset contexts. Splitting samples into places with vs. without conflict in the prior five years yields similar conclusions: EWI add no predictive power in either subgroup. - Correlation analysis indicates limited linkage between EWI and key socioeconomic predictors, and what correlation exists is spatial. For leading socioeconomic predictors, regressions on EWI show low R^2 (e.g., GDP per capita R^2≈0.15; infant mortality R^2≈0.15), which largely vanish after controlling for place fixed effects (e.g., GDP per capita R^2≈0.03), implying overlap in spatial patterns rather than temporal co-variation. - Spatial versus temporal variation: model-predicted conflict risk varies predominantly across space, not time. Year fixed effects explain almost none of the prediction variance, whereas place fixed effects explain more than half, and up to about 90% in some cases. For socioeconomic predictors, over 90% of predictive information is spatial; even for EWI, spatial variation dominates temporal variation. - Interpretation: EWI spatial patterns overlap with socioeconomic and governance spatial patterns, so the predictive information in EWI is largely redundant with established predictors, explaining the absence of gains when adding EWI to full models.

The study directly tests whether biophysical extreme weather impacts from harmonized climate impact models improve annual conflict incidence forecasts when added to comprehensive models using established predictors. The findings answer the research question in the negative: EWI do not improve forecasts beyond conflict history, governance, and socioeconomic indicators. This is explained by substantial redundancy in spatial information—EWI exposure tends to co-occur with particular socioeconomic and governance conditions, so EWI contribute little unique predictive content once these are included. Although EWI have modest standalone predictive power, their added value disappears in fully specified models. These results align with work finding limited predictive benefit from climatic variables and contrast with studies that identify conditional associations (e.g., conflict onsets following disasters in specific contexts). Differences in temporal resolution (annual vs. monthly), outcome focus (incidence vs. onset), and spatial scale (subnational vs. national) may account for divergences. Importantly, the dominance of spatial over temporal predictive variation suggests that forecasting where conflict risk is high is easier than predicting when it changes. For early warning systems targeting annual incidence, incorporating EWI is not recommended; investing in high-quality conflict history and governance data remains more impactful. The study also sets a methodological benchmark: before integrating climate indicators into operational forecasts, their incremental predictive value should be explicitly tested within non-parametric frameworks that capture conditional, non-linear relationships.

This paper integrates ISIMIP2a-based extreme weather impact data (floods, droughts, tropical cyclones, crop shocks, and disaster-induced displacement) into GRF conflict forecasting models for Africa and rigorously tests their incremental predictive value. The main contribution is a clear negative result: EWI do not improve one-year-ahead conflict incidence forecasts once conflict history, governance, and socioeconomic predictors are included, despite EWI’s modest standalone predictive ability. The analysis explains this by overlapping spatial information between EWI and socioeconomic/governance conditions and by the predominance of spatial over temporal predictive variation. Policy and modeling implication: operational early warning models for annual conflict incidence should prioritize conflict history and governance data over EWI. Future research should test other regions, longer and more recent time periods as data extend beyond 2012, alternative climate impact models and event categories, different conflict outcomes (e.g., onsets at higher temporal resolution), and modeling strategies emphasizing rare-event prediction (e.g., class weighting), and continue to require explicit tests of incremental predictive value before incorporating climate indicators into forecasts.

- Temporal coverage: ISIMIP2a impact data limit the study to 2012; ACLED coverage restricts analysis to 2002–2012 for ACLED-based outcomes. - Spatial and outcome scope: Focus on mainland Africa and annual incidence; results may differ in other regions, at higher temporal resolution (e.g., monthly), or for conflict onsets/cessations specifically. - Data granularity: Governance indicators are predominantly national-level; subnational governance data are sparse, potentially omitting relevant local political dynamics. - National-level sample size: With 11 annual observations for 49 countries, national-level GRF models may have limited flexibility relative to subnational models. - Extreme event representation: Disaster-induced displacement includes some non-weather disasters (e.g., earthquakes); crop and hazard simulations are constrained to the ISIMIP2a setup and may omit other impact channels. - Predictive target imbalance: Low incidence of conflict at fine scales poses class imbalance challenges; while PR curves are used, alternative weighting or rare-event methods might yield different insights. - Generalizability: Correlative predictive findings do not preclude EWI improving forecasts in other contexts (regions, periods, models) or for different conflict outcomes; the study explicitly cautions that results may vary.