Understanding the Scenarios

The Migration Scenario Engine combines machine-learning projections with structural climate displacement models to explore how global bilateral migration could evolve under different futures. Each scenario answers a different "what if" question.

76M The price of inaction
(Adaptation vs. Climate Extreme, 2095)

5 Scenarios tested across
4 SSP pathways to 2100

49M Displacement gap
(ML models underestimate by ~25%)

Five Scenarios, Five Futures

Each scenario modifies the model's input features and displacement assumptions. All figures below use SSP2-4.5 (Middle of the Road) at 2095.

Baseline

"What does the ML model predict from historical patterns alone?"

216M

5-year bilateral flow, SSP2 2095

Pure ensemble prediction (GAM + Random Forest + XGBoost, R² = 0.826) trained on 316,020 observations from 1990-2015. No displacement adjustment. The model captures economic, network, gravity, and moderate climate effects.

Baseline+

"What if we add displacement from uninhabitability thresholds the ML model cannot learn?"

265M

5-year bilateral flow, SSP2 2095

ML prediction plus structural displacement from sea level rise, extreme heat, and drought. Adds 49M to the baseline - the "displacement gap" that statistical models miss because 1990-2015 data contains no precedent for permanent uninhabitability.

Adaptation Success

"What if global cooperation succeeds and adaptation investments are effective?"

250M

5-year bilateral flow, SSP2 2095

Reduced climate sensitivity (×0.8), lower conflict (×0.7), rising incomes (×1.1). Displacement overlay reduced (SLR ×0.8, heat ×0.8, drought ×0.9). The most optimistic scenario with active intervention.

Fragmentation

"What if governance fragments and conflict intensifies?"

285M

5-year bilateral flow, SSP2 2095

Elevated conflict (×1.5), governance breakdown (×1.3), economic stagnation (income ×0.8). Climate sensitivity unchanged (SSP3 already embeds high forcing). Displacement overlay slightly amplified (×1.2 across channels).

Climate Extreme

"What if climate impacts exceed current projections?"

326M

5-year bilateral flow, SSP2 2095

Maximum climate stress test. Climate sensitivity ×1.5, drought ×1.5, flood ×1.4, storm ×1.5. Displacement overlay amplified ×1.5 across all channels. The upper bound of plausible outcomes.

Why This Matters for Policy

The Price of Inaction

76M

The difference between Adaptation Success (250M) and Climate Extreme (326M) at SSP2-4.5 2095. This gap represents the additional bilateral migration that could be prevented through effective climate adaptation, conflict reduction, and governance investment. These are policy-amenable levers.

The Displacement Gap

49M

The difference between Baseline (216M) and Baseline+ (265M) at SSP2-4.5 2095. ML models trained on 1990-2015 data systematically underestimate future migration by approximately 25% because they cannot extrapolate to conditions of permanent uninhabitability - sea level rise, lethal heat, and agricultural collapse - that have no historical precedent.

The Window for Action

2050

All scenarios look similar through 2050 (164-194M range for SSP2). After 2050, they diverge sharply. Policy decisions made in the next 25 years determine which trajectory the world follows. Early intervention compounds; late intervention faces locked-in displacement.

How Scenarios Are Constructed

A hybrid approach: machine learning captures historical patterns, physics-based thresholds capture future regime shifts.

Total Projected Migration = Layer 1 + Layer 2

Layer 1: ML Ensemble

Historical Pattern Prediction

Three-model stacking ensemble trained on 316,020 bilateral observations (1990-2015):

GAM (semiparametric regression): 31.7% weight
Random Forest: 33.3% weight
XGBoost: 35.0% weight

Scenarios modify input features before prediction: 7 channels (climate, conflict, income, drought, flood, storm, governance) are scaled by scenario-specific multipliers. The ensemble's learned nonlinear responses determine the effect.

+

Layer 2: Displacement Overlay

Physical Threshold Displacement

Structural displacement from three channels beyond historical experience:

Sea Level Rise: LECZ populations displaced (687M exposed below 5m globally). Quadratic ramp based on IPCC AR6 SLR projections.
Extreme Heat: Sigmoid displacement at 33°C threshold (Mora et al. 2017). Max 25% of population.
Drought: Agricultural population displaced when precipitation anomaly exceeds collapse thresholds. Max 15% of population.

Stock-to-flow conversion at 10% per 5-year period. Displaced populations distributed across existing migration corridors.

Projection Results: All Scenarios × All SSP Pathways at 2095

Values in millions (M). 5-year bilateral migration flows. Source: scenario_summary_statistics.csv

Scenario	SSP1-2.6 Sustainability	SSP2-4.5 Middle of Road	SSP3-7.0 Regional Rivalry	SSP5-8.5 Fossil-fuelled
Baseline (ML only)	179	216	237	145
Baseline+ (ML + Displacement)	206	265	338	226
Adaptation Success	202	250	305	198
Fragmentation	214	285	378	269
Climate Extreme	232	326	458	342

Displacement Breakdown by Channel (SSP2-4.5, 2095)

Additional displacement in millions, added to ML baseline. Source: displacement_summary.csv

Scenario	Sea Level Rise	Extreme Heat	Drought	Total Displacement
Baseline+	44.9M	3.3M	0.5M	48.7M
Adaptation Success	31.4M	1.7M	0.4M	33.5M
Fragmentation	63.7M	4.4M	0.6M	68.7M
Climate Extreme	98.5M	10.9M	0.7M	110.1M

Displacement composition (Climate Extreme, SSP2 2095):

Sea Level Rise (89.5%) Extreme Heat (9.9%) Drought (0.6%)

Technical Details

For researchers and technical reviewers.

7-Channel Scenario Multiplier Table

Scenarios modify input features before the ML ensemble generates predictions. Seven channels are scaled independently. Multiplier values are grounded in peer-reviewed literature (see methodology document for full citations).

Scenario	Climate Sensitivity	Conflict	Income Elasticity	Drought	Flood	Storm	Governance
Baseline	1.0	1.0	1.0	1.0	1.0	1.0	1.0
Climate Extreme	1.5	1.0	0.9	1.5	1.4	1.5	1.0
Fragmentation	1.0	1.5	0.8	1.0	1.0	1.0	1.3
Adaptation Success	0.8	0.7	1.1	0.9	0.9	1.0	1.0

Literature basis: Conflict ×1.5 from Hegre et al. (2016), Abel et al. (2019); drought/flood/storm from IPCC AR6 WGI Table SPM.1, Hirabayashi et al. (2013); climate sensitivity from Cattaneo & Peri (2016), Hoffmann et al. (2020); governance from Kaufmann & Kraay (2023).

Displacement Formulas and Data Sources

Sea Level Rise
Population below 5m: rate = min(1, (slr_m / 2.5)²)
Population 5-10m: rate = max(0, ((slr_m - 2.0) / 8.0)²)
Data: NASA SEDAC LECZ v3 (234 countries, MERIT-DEM, GHS-POP R2019, snapshots 1990/2000/2015 — MSE uses 2015 as exposure baseline). SLR projections: IPCC AR6 WG1 Table 9.9 (Fox-Kemper et al. 2021).

Extreme Heat
frac = 0.25 / (1 + exp(-0.5 × (effective_temp - 33)))
No displacement below 30°C. Sigmoid onset at 33°C. Maximum 25% of population.
Threshold literature: Mora et al. (2017), Im et al. (2017).

Drought / Desertification
frac = min(0.15, severity × agr_frac × 0.3)
No displacement above -150mm anomaly. Only agricultural population affected (capped at 50%).
Data: CRU TS 4.09 / CMIP6, World Bank WDI (AG.LND.AGRI.ZS).

Stock-to-Flow: flow = stock × 0.10 per 5-year period (10% displacement rate, full relocation over ~50 years).

Corridor distribution: Displaced populations distributed proportional to baseline corridor shares (Massey et al. 1993 network-informed allocation).

Methodology Evolution (V1-V7)

V1-V5 (April 1-10): Attempted to perturb model predictions with ad-hoc coefficients. Failed because multipliers were disconnected from the model's learned sensitivities. Scenarios compressed to baseline or broke the 2020 anchor.

V6 (April 11): Breakthrough - instead of perturbing outputs, modify INPUT FEATURES directly (e.g., tmp_anomaly_orig × 1.5) and re-run the full ensemble. The model's own learned nonlinear responses (GAM splines, RF tree splits, XGBoost boosting residuals) determine the scenario effect.

V7 (April 13, current): V6 + structural displacement overlay. Hybrid ML + physics approach. ML handles "normal" socioeconomic migration; the overlay captures regime shifts (uninhabitability) that the model cannot learn from 1990-2015 data.

Important Limitations

These projections are narrative-conditional stress tests, not unconditional forecasts. They answer "what would happen if..." rather than "what will happen."

Key limitations:

The 10% displacement rate per 5-year period is an assumption, not calibrated to empirical displacement data.
Country-level analysis masks sub-national dynamics and short-term displacement events.
No feedback loop: displaced populations do not reduce the origin population for subsequent periods.
Corridor shares from baseline may not hold under mass displacement conditions.
80 years of projection from 25 years of training data carries inherent extrapolation risk.
Displacement threshold functions (SLR quadratic ramp, heat sigmoid at 33°C) use peer-reviewed data sources but custom rate parameters.