# Migration Scenario Engine

> Open scientific platform projecting global bilateral migration flows
> under climate scenarios for 2020–2100. Stacking ensemble (GAM +
> Random Forest + XGBoost, R² = 0.826) trained on 316,020 dyad-period
> observations across 230 countries, 52,670 bilateral corridors and 6
> historical periods (1990–2015). Independent doctoral research,
> self-funded, no commercial or institutional agenda.

The interactive dashboard at https://migrationengine.org/ is
a single-page Shiny app, so a plain HTTP fetch returns only the JS
shell. The links below point to static HTML mirrors that contain the
full content of every section, plus structured Markdown for AI consumers.

**Author**: Christian Rogalski, M.Sc. — External Doctoral Researcher,
Alexandru Ioan Cuza University of Iași (Faculty of Economics and
Business Administration, Department of Statistics and Cybernetics).
Contact: rogalski.academic@pm.me
ORCID: https://orcid.org/0009-0007-6144-0631
Source code: https://github.com/MrVizzo/migrationengine

**License**: CC BY-NC 4.0 (data and projections); independent research
output, no funding from external sponsors.

**Cite as**: Rogalski, C. (2026). *Migration Scenario Engine: Global
Bilateral Migration Projections under Climate Scenarios, 2020–2100*.
CERIFR Research. https://migrationengine.org

## Project

- [About the project and the author](https://migrationengine.org/about.html): research approach, technical focus, motivation, why this matters, what's next.
- [Methodology](https://migrationengine.org/methodology.html): full pipeline (12 sections) — target variable, stacking ensemble, GAM specification, feature engineering (109 predictors), expanding-window cross-validation, shock models, Migration Pressure Index, Trapped Population Index, scenarios (V7 hybrid: feature multiplication + displacement overlay), diaspora recursion, IPF calibration to UN WPP, conformal prediction intervals.
- [Data catalogue](https://migrationengine.org/data.html): 17 primary data sources (UN DESA IMS 2024, Abel & Cohen 2019, CRU TS 4.09, CMIP6 ScenarioMIP 5 GCMs, World Bank WDI, UCDP v25.1, ND-GAIN, EM-DAT, UN WPP 2024, IIASA SSP Database v3.1, NASA SEDAC LECZ, IPCC AR6, CEPII GeoDist, etc.) with periods, coverage, licenses and download links for the harmonised outputs.
- [Results](https://migrationengine.org/results.html): headline projections across 4 SSP pathways × 5 narrative scenarios, top corridors, hotspot countries, fan-charts of corridor uncertainty.
- [FAQ](https://migrationengine.org/faq.html): 14 questions covering stock-vs-flow disambiguation, "climate refugee" language, SSP × scenario design, projection accuracy, conformal prediction intervals, peer-review status, funding, data sources, citation, disambiguation from database-migration tools, MPI/TPI definitions, ensemble composition, and scope exclusions.
- [Glossary](https://migrationengine.org/glossary.html): 21 terms (MPI, TPI, SSP, ND-GAIN, OOF, SHAP, IPF, CPI, CMIP6, etc.) with concise definitions.

## Key facts at a glance

- **Method**: Stacking ensemble — GAM (weight 31.7%, OOF R² = 0.795) + Random Forest (33.3%, R² = 0.804) + XGBoost (35.0%, R² = 0.813), combined via Ridge meta-learner. Pooled OOF R² = 0.826 across 5-fold expanding-window CV (Fold 2: 0.810 → Fold 5: 0.855, monotonically increasing — validates genuine learning, not overfitting; Fold 1 with single-period training intentionally weak). Reaches 99.9% of the temporal autocorrelation ceiling (r² = 0.827).
- **Target variable**: log(1 + flow rate per 1000 origin population), back-transformed and clipped to [0, 1000] before population-weighted volume estimation.
- **Predictors**: 109 features in 12 groups — climate (54%, including CMIP6 anomalies), economic (GDP, urbanisation), governance (WGI), conflict (UCDP), disasters (EM-DAT), gravity (CEPII), policy (DEMIG VISA), demography (UN WPP), education (Barro-Lee), network centrality (PageRank, betweenness), diaspora dynamics, temporal interactions. v2 features use prior-period data only — no leakage.
- **Calibration**: Iterative Proportional Fitting (Willekens 1999, Abel & Cohen 2019) aligning bilateral flow matrices to UN WPP origin/destination totals with population caps.
- **Uncertainty**: Conformal Prediction Intervals (Vovk 2005, Romano 2019, Barber 2023) — Mondrian-binned by flow magnitude, multiplicative bootstrap (N=500), 50% and 90% coverage, distribution-free guarantees.
- **Scenarios**: 4 IPCC SSPs (SSP1 Sustainability, SSP2 Middle of the Road, SSP3 Rivalry, SSP5 Fossil-fueled) crossed with 5 narrative scenarios (Baseline ML-only, Baseline+, Adaptation Success, Fragmentation, Climate Extreme). Climate Extreme adds ×1.5 displacement overlay from 3 channels: sea level rise (IPCC AR6 + NASA SEDAC LECZ; 687M pop <5m, 1.056B <10m global exposure), extreme heat (Mora et al. sigmoid onset at 33°C), drought/desertification (CRU/CMIP6, –150 mm anomaly threshold).
- **Coverage**: 195 countries, 5-year periods 2020–2100 (and 1990–2015 historical), 4 SSP pathways, 5 scenarios per pathway.
- **Diaspora model**: 5-year survival 0.87 (mortality 0.10, return 0.03); strongest single predictor (r = 0.33); endogenous network feedback in projection.

## Optional

- [Citation File Format (CFF)](https://migrationengine.org/citation.cff) — machine-readable citation metadata.
- [Sitemap](https://migrationengine.org/sitemap.xml) — full URL inventory.
- [Detailed methodology and data documentation as a single long document](https://migrationengine.org/llms-full.txt) — everything inline, no further fetches needed.
- [Privacy policy](https://migrationengine.org/about.html#privacy) — no tracking, no third-party cookies; only one localStorage flag (`mse_welcome_dismissed`) for the dismissal of the welcome dialog.