Climatic conditions are weak predictors of asylum migration

The study investigates how well climatic conditions predict near-future asylum migration to the European Union compared with economic, political, and contextual factors. While climate variability and extreme events can affect migration directly and indirectly, international migration decisions are shaped by multiple intertwined drivers. Prior work has found statistical links between climate anomalies and asylum applications and even projected large increases under high warming, but such extrapolations may lack real-world relevance. The authors aim to place climate effects in context by testing the relative predictive performance of climatic, economic, and violence-related indicators for asylum migration, addressing concerns about overfitting and the complex, nonlinear interactions among determinants.

Previous research provides more evidence for climatic effects on short-distance displacement than on international migration, though several studies link adverse climate conditions in origin countries to refugee and asylum flows. One influential analysis suggested EU asylum applications could triple under high warming, assuming other drivers remain constant. However, migration is not monocausal; socioeconomic and political drivers are widely considered primary, and interactions among factors complicate inference. Scholars have called for nuanced approaches beyond the notion of “climate refugees,” warning against securitization. The EU’s recent asylum surge illustrates broader forced migration trends, traditionally driven by conflict and repression, though climate stressors may add to perceived threats or indirectly affect conflict—pathways that remain debated. Policy discourse has begun to consider climate-related claims for protection, but empirical assessments of relative importance across drivers remain limited.

The authors compile a country-year dataset for 175 independent countries from 1999–2018, measuring per-capita first-time asylum applications to any EU-28 state as the outcome (log-transformed). Predictors are organized into thematic components and baseline contextual indicators, all lagged one year. Climate component: population-weighted mean annual temperature; growing-season-weighted SPEI-3 anomalies separated into positive (wet) and negative (dry) deviations; and 3-year moving averages of these anomalies for cropland areas (derived from CRU TS data, MIRCA2000 growing seasons, GPW v4 population weights). Economy component: GDP per capita (PPP, log), year-on-year GDP per capita growth, interpersonal globalization index (KOF), share of 20–29-year-olds with post-secondary education (Wittgenstein Centre), and infant mortality rate (WDI). Violence component: battle-related deaths (UCDP/PRIO, log), population within 20 km of conflict events (UCDP GED with buffers, GPW population; log), physical integrity rights (V-Dem), freedom of movement (V-Dem), and homicide rate (GBD; log). Baseline indicators included in all models: highest democracy score among contiguous/nearby neighbors (V-Dem, COW contiguity), country area (log), population size (log), urbanization rate, and geodesic distance to nearest EU-28 member. Missing data are addressed via multiple imputation (Amelia II), creating ten imputed datasets. Prediction uses random forests with leave-future-out cross-validation: models are trained on sliding 4-year windows (1999–2002, …, 2014–2017) and predict the subsequent year (2003–2018), repeated across imputations (160 simulations). Performance is assessed by mean absolute error on log per-capita applications. Accumulated Local Effects (ALE) are used to interpret marginal predictive contributions. Robustness checks include removing baseline indicators, alternative lags (contemporaneous, 2-year), varying training window lengths and forecast horizons (up to 12-year windows, up to 8 years ahead), different transformations, and alternative outcome definitions (stocks, global). Conventional in-sample regressions are also estimated for comparison.

• Out-of-sample prediction shows the violence component model is the most accurate predictor of asylum flows to the EU, outperforming the economy model and even the full model; the climate component performs worst.
• Temporal dynamics of the 2012–2016 surge in EU asylum applications are best captured by violence indicators, aligning with increases in armed conflict and repression; the post-2016 decline corresponds with reduced conflict severity in major sending regions.
• ALE analyses indicate strong, robust associations for violence indicators: asylum applications rise approximately log-linearly beyond about 500 annual battle-related deaths; deterioration in civil liberties (physical integrity rights and freedom of movement) increases predicted asylum migration; homicide has weaker effects.
• Economic indicators shape latent migration potential: GDP per capita is a strong predictor with a breakpoint around USD 10,000 per capita, above which predicted asylum migration declines; growth rates have weaker predictive power than levels; education and infant mortality change slowly and primarily capture cross-sectional differences.
• Baseline context matters: predicted per-capita asylum migration declines sharply with greater distance to the EU; it peaks around an origin-country population of roughly three million, suggesting structural constraints and substitution with internal displacement in larger countries.
• Climate indicators add little predictive power: temperature shows a weak, mostly negative association with asylum migration and has limited temporal variation; SPEI-based drought/wetness anomalies (including multi-year anomalies) have virtually flat ALEs, indicating minimal marginal predictive contribution.
• In-sample regressions reproduce a U-shaped temperature association, but this does not translate into out-of-sample predictive performance, highlighting that statistically significant regressors can be poor predictors on new data.
• Robustness checks across alternative specifications and horizons confirm the comparatively weak predictive performance of climatic conditions relative to violence and economic factors.

The findings address the core question by demonstrating that, when evaluated through realistic out-of-sample prediction, climatic conditions are weak predictors of EU-bound asylum migration compared with political violence and repression, and to a lesser extent economic conditions and structural context. This suggests that near-term asylum flows are more responsive to political and security changes in origin countries than to climate anomalies. The results have practical implications for early warning and policy, emphasizing monitoring of conflict intensity and civil liberties. The study also notes that origin-country conditions are only part of the picture: destination policies, transit route dynamics, migrant networks, and access to resources shape realized flows. The 2016 EU–Turkey agreement and COVID-19 border closures illustrate how policy shifts can affect arrivals. Moreover, asylum migration is a specific and relatively rare form of mobility compared to internal displacement or regional cross-border movements, which may be more sensitive to environmental shocks. The analysis does not refute potential long-term, indirect climate effects or future impacts once adaptation limits are exceeded; rather, it contextualizes their limited short-term predictive role for EU asylum applications.

The study contributes by systematically benchmarking the predictive value of climatic, economic, political, and contextual factors for EU asylum migration using machine-learning with leave-future-out validation. It shows that climate-based predictors (temperature and drought anomalies) add little to predicting near-term asylum flows, while political violence and repression are strong and policy-relevant early warning signals; economic and structural factors shape latent migration potential. Policy implications point toward investments in conflict resolution, peacebuilding, and governance improvements in fragile or repressive states to mitigate future asylum pressures. Future research should assess the relative roles of climate and non-climate drivers for internal displacement and regional migration, explore longer-term indirect climate effects, and integrate destination-side policy dynamics and network effects into predictive frameworks.

• The prediction framework does not estimate causal effects; strong predictive performance does not imply causality, and weak predictive performance does not preclude context-specific causal pathways.
• Focus on EU-28 asylum applications may not generalize to other destinations or to internal displacement; data limitations precluded a comparable global analysis of internal movements.
• Destination-side policies, enforcement, and route availability (e.g., EU–Turkey agreement, COVID-19 border closures) can strongly affect arrivals but are not fully captured by origin-country predictors.
• Some predictors change slowly over time, emphasizing cross-sectional differences rather than short-term fluctuations; climate variables may have longer-term impacts not captured in 1-year-lagged, near-term predictions.
• Missing data were handled via multiple imputation, introducing some uncertainty; event data limitations (e.g., Syria GED coverage) required substitutions.
• The analysis uses per-capita, log-transformed applications and specific modeling choices (RF, window sizes), though robustness checks suggest results are not driven by these decisions.