Modeling International Expansion Under Uncertainty.

The project began with a simple question: how could a country-screening framework become a financially coherent expansion model without turning every strategic score into a fake valuation coefficient? Answering that question required more than adding randomness to a spreadsheet. It required a governed architecture for what was empirical, what was structural, how variables moved together, and how operating results could change the rollout itself.

From simulation idea to decision system

The easy version would have sampled independent variables from convenient distributions. The final project instead had to answer a harder question: how do you construct a coherent joint country-state system when factor distributions differ by income group, relationships are dependent, and the assembled correlation structure is not automatically valid?

Scale of the final model

• 15-year horizon, six-warehouse expansion framework
• Four World Bank income groups × three rollout policies = 12 scenarios
• 10,000 Monte Carlo paths per scenario
• 21 country-screening factors + two FX states
• Separate 21×21 AHP governance matrix and 23×23 stochastic dependence matrix

Strategic importance is not financial sensitivity

The most distinctive part of the project was the integration. Country screening, AHP governance, empirical distributions, financial transmission, rollout strategy, financing constraints, and valuation all had to work as one system. A country factor only mattered financially through a defined economic channel; the AHP weights never became NPV coefficients just because they were numeric.

Building realistic marginal distributions

I calibrated income-group-specific empirical marginals using adaptive/asymmetric kernel estimation where appropriate, cross-validated partial pooling, soft tail saturation, 6,000-point numerical grids, and 100,000 diagnostic draws per factor/group.

Making the variables move together coherently

I estimated pairwise Spearman dependence, converted it through the standard Gaussian-copula relationship, and then dealt with a critical numerical issue: the assembled dependence matrix was not positive semidefinite.

I repaired that matrix using Higham’s nearest-correlation procedure and validated the transformed joint system with 50,000 draws per income group after inverse-CDF transformation.

Separating company calibration from destination-country effects

A separate modeling issue was that Costco-derived operating relationships already embedded a U.S. country state. I built a bi-directionally validated True-Zero bridge to remove that embedded state before applying destination-country conditions through explicit revenue, cost, working-capital, timing, FX, and WACC channels.

Making strategy conditional on financing capacity

The staged/moderate policy does not assume every warehouse opens on schedule. Openings can defer if cash plus incremental borrowing headroom is insufficient, with researcher-defined continuation checks around leverage, interest coverage, and maintenance cash flow/debt.

Model governance: investigate before you “fix”

Several of the strongest lessons came from moments when the model or documentation looked suspicious. I learned not to defend the first version just because it ran. I traced source lineage, separated empirical inputs from structural priors, tested calibration and financing invariants, and reconciled the executable model against the written report.

That distinction mattered during the warehouse-age ramp review: investigation showed that the executive model was correct and the documentation was wrong. In other cases, such as inflation/network-support behavior, the economic mechanics themselves needed revision. Being able to tell a documentation error from a model error became part of the analytical work.

Where I deliberately refused false precision

Some uncertainty could be estimated; some could not. Ten annual company observations were not enough to justify fitting a six-dimensional company copula, so I did not pretend they were. The model used a simpler central treatment plus a robustness comparison instead. More generally, I tried to make priors visible and stress-test them rather than disguise them as estimated facts.

What the model found

The strategy comparison was not a simple ranking. Aggressive bought rollout certainty with external capital and wider downside. Moderate reduced sponsor exposure and achieved the highest strict-success frequency in every income group, but generally took longer. Organic preserved outside capital most aggressively, yet often failed to complete the intended six-warehouse network.

Validation and reproducibility

What changed in how I think

The biggest lesson was that a sophisticated model is mostly an exercise in defining and validating assumptions. The simulation engine was not the hard part; building a defensible joint state, enforcing coherent financing behavior, and proving the implementation reproduced its own rules was.