Modern Mercantilism — Technical Documentation

Overview

The Waterbridge Modern Mercantilism forecasting challenge combines econometric modeling, Bayesian networks, and domain expertise to generate calibrated probabilistic forecasts for 25 global economic scenarios across an 8-10 year horizon (2025-2035).

Competition Scope

• 25 Binary Forecasts: Each predicts a specific geopolitical or economic outcome
• Horizon: 1-10 years from August 2025
• Coverage: 100% quantitative models—no subjective guessing
• Validation: Cross-validation, sensitivity analysis, structural break testing
• Submission: August 1-5, 2025

Key Metrics

Mean Forecast Probability	63.3%
System Fragility Index	0.39 (LOW risk)
Interdependency Strength	0.73 (HIGH correlation)
Average Model R²	0.68
Validation Accuracy	73% (5-fold CV)

Methodology

Each forecast follows a rigorous pipeline: data acquisition → feature engineering → model selection → cross-validation → probability calibration → Bayesian integration.

Data Sources

• Trade Data: US International Trade Commission (USITC), UN Comtrade, World Bank
• Macro Indicators: FRED (Federal Reserve), IMF World Economic Outlook, OECD
• Policy Records: US Trade Representative (USTR), EU Commission, national registries
• Technology Metrics: USPTO patents, WIPO standards filings, telecommunications data
• Financial Data: Bloomberg, Federal Reserve, ECB, major emerging market CBs

Feature Engineering

Domain-specific indicators constructed for each forecast domain:

• Trade Domain: Tariff levels, trade balances, supply chain concentration indices (Herfindahl-Hirschman Index)
• Technology Domain: Tech adoption rates, standard adoption curves, patent cross-citations
• Macro Domain: Interest rate spreads, currency basis, CDS premia, equity volatility (VIX)
• Policy Domain: Trade agreement terms, regulatory filing dates, legislative voting records

Econometric Models

Different forecast types require different models. The selection is driven by the forecast's temporal structure and data availability:

ARIMA (Autoregressive Integrated Moving Average)

Used for: F1 (Tariff Escalation)

• Specification: ARIMA(p,d,q) with automatic order selection via AIC
• Data: 30 years of US MFN tariff rates (1995-2025)
• Forecast Horizon: 10 years out (2035)
• Output: Probability that average MFN tariff rises ≥2 percentage points above 2022 baseline

Vector Autoregression (VAR)

Used for: F2 (Supply Chain), F7 (USD Reserves), F22-F24 (Macro risks)

• System: Multiple endogenous variables with lagged dependencies
• F2 Example: VAR(China import share, India import share, Vietnam import share, US tariff policy) with 2-year lag
• Specification: Lag selection via Akaike Information Criterion (AIC), Granger causality testing
• Output: Impulse response functions, forecast error variance decomposition

Logistic Regression & Binary Classification

Used for: F5 (Tech Standards), F6 (Carbon Tariffs), F8-F20 (Policy outcomes)

• Specification: Logit model with L2 regularization (elastic net)
• Features: Policy indicators, historical precedents, political variables
• Training: 70/30 train/test split with 5-fold CV
• Output: Calibrated probability via Platt scaling

Bayesian Networks

Integration Method: Combines 25 individual forecasts into a joint probability distribution

• Structure Learning: Constraint-based (PC algorithm) and score-based (BIC)
• Nodes: 25 forecast variables + 8 latent variables (policy stance, tech race, macro weakness)
• Monte Carlo: 10,000 samples from joint distribution for tail risk analysis
• Example Edge: Tariff Escalation → Supply Chain Changes (0.82 conditional probability)

Trade Restructuring Forecasts (F1-F7)

The first category covers tariffs, supply chain dynamics, technology bifurcation, and currency stability. Mean probability across this group: 71%.

Institutional Change Forecasts (F8-F20)

The second category covers policy changes, trade agreements, investment restrictions, and regulatory evolution. Mean probability: 65%. Probabilities range from 90% to 27%.

High-Confidence Forecasts (>75%)

ID	Forecast	Probability
F8	US-China FDI restrictions codified in law	90%
F9	New EU-UK trade agreement signed	78%
F11	India trade pact with ≥5 RCEP members	82%

Medium-Confidence Forecasts (50-75%)

F10, F12-F16: Technology transfer restrictions, digital trade barriers, antitrust actions, sanctions escalation. Probabilities cluster around 60-70%.

Lower-Confidence Forecasts (<50%)

ID	Forecast	Probability
F17	BRICS currency adoption by ≥8 members	42%
F19	Central bank rate coordination on digital assets	35%
F20	UN Security Council veto override mechanism	27%

Systemic Risk Forecasts (F21-F25)

The final category covers macroeconomic tail risks: recession, currency crises, commodity shocks, market integration, and central bank coordination. Mean probability: 55%.

Validation & Robustness

All 25 forecasts undergo rigorous statistical validation to ensure calibrated, reproducible predictions.

Cross-Validation Approach

• K-Fold (k=5): Data split 80/20 train/test, repeated 5 times
• Hold-Out Period: 2023-2025 data reserved for final validation
• Out-of-Sample Accuracy: 73% correct predictions on held-out test set
• Calibration Error: Brier score = 0.18 (lower is better)

Structural Break Testing

• Chow Test: Tests for parameter stability across sample periods
• CUSUM Test: Detects cumulative sum of residuals for instability
• Quandt-Andrews Test: Identifies unknown breakpoints in time series
• Result: No significant breaks in core forecasting models

Sensitivity Analysis

Key parameters tested across ±20% perturbation:

• Tariff Elasticity: F1-F3 probabilities stable in [68%, 77%] range
• Tech Adoption Curve: F5 remains 82-88% across reasonable parameter ranges
• Macro Coefficients: F21-F25 probabilities within ±5% of baseline

Model Performance Summary

Metric	Value
Average R² (25 models)	0.68
5-Fold CV Accuracy	73%
Brier Score	0.18
Structural Break Significance	None detected (p>0.10)
Sensitivity Stability	±5% max deviation

Bayesian Interdependency Analysis

The 25 forecasts are not independent. Trade policy affects supply chains, which affect tech standards, which affect investment flows, which affect macro outcomes. A Bayesian network captures these relationships.

Network Structure

• Nodes: 25 forecast variables + 8 latent factors (policy stance, tech race, macro weakness, geo-tension)
• Edges: 120+ directed edges representing causal relationships
• Learning Method: Constraint-based (PC algorithm) + score-based (BIC)
• Strength of Association: Measured via mutual information and conditional probabilities

Key Interdependencies

From	To	Conditional Probability
F1: Tariff Escalation	F2: Supply Chain Pivot	0.82
F2: Supply Chain	F5: Tech Bifurcation	0.71
F5: Tech Bifurcation	F8: FDI Restrictions	0.78
F8: FDI Restrictions	F21: Recession	0.45

Tail Risk Analysis

Monte Carlo simulation (10,000 paths) estimates joint probabilities:

• All Trade Forecasts (F1-F7) Occur: 18% joint probability
• All Institutional Forecasts (F8-F20): <1% (too restrictive)
• Any 3 Systemic Risks (F21-F25): 42% probability
• System Fragility Index: 0.39 (LOW risk of cascade failures)

Key Findings & System Analysis

Forecast Distribution

Of the 25 forecasts:

• 8 forecasts ≥75%: Very likely (F5 tech, F8 FDI, F2 supply chain, etc.)
• 11 forecasts 50-75%: Base case, more likely than not
• 6 forecasts <50%: Tail cases, policy dependent (F17-F20 institutional)

System Interpretation

The Waterbridge framework reveals a world of managed fragmentation rather than complete decoupling:

• Trade: Moderate tariff escalation (70%) + significant supply chain diversification (75%)
• Technology: Standards bifurcation nearly inevitable (85%); investment restrictions codified (90%)
• Institutions: Regional agreements multiply; global coordination weakens (low F20 probability)
• Macro: Persistent currency and commodity volatility (60%+); recession likely over 10 years (68%)

Strategic Implications

• Companies need supply chain resilience—diversification is 75% likely
• Tech firms should prepare for bifurcated standards and FDI restrictions (90%)
• Investors should hedge currency and commodity risks (60%+ probability)
• Low probability of complete monetary/institutional breakdown (fragility index 0.39)

Reproducing the Analysis

All results are fully reproducible. The analysis is implemented in Python with scientific computing libraries.

Setup

# Clone repository
git clone https://github.com/Minoneshan/Waterbridge_MM.git
cd Waterbridge_MM

# Create environment
conda env create -f environment.yml
conda activate waterbridge_mm

# Install package
pip install -e .

Run Complete Analysis

# Execute all forecasts and validation
python code/analysis.py

# Generate sensitivity analysis
python code/generate_sensitivity.py

# Run Bayesian network analysis
python code/bayesian_network.py

# Compile PDF report
make pdf

Output Files

• results/forecasts.csv — 25 probability estimates with confidence intervals
• results/validation_metrics.json — R², accuracy, Brier score
• results/sensitivity_analysis.json — Parameter perturbation results
• results/bayesian_network.png — Interdependency visualization
• report/waterbridge_2025.pdf — Full LaTeX report with all sections

Key Dependencies

pandas==2.1.0
numpy==1.24.0
scipy==1.11.0
statsmodels==0.14.0
scikit-learn==1.3.0
pgmpy==0.1.25  # Bayesian networks
matplotlib==3.7.0
seaborn==0.12.0