Research, strategy, and turning insight into action.

Modelled repeated interaction between Beiersdorf and Unilever in Vietnam’s skincare market, built the payoff matrices from case data, and recommended an omnichannel position that balanced digital growth against channel conflict.

• Built the payoffs from the case tables rather than assumed numbers.
• Handled uncertainty as part of the model: signalling and incomplete information, not one scenario.

Modelled the 2025 US–China semiconductor dispute as a simultaneous, repeated, non-cooperative game between two governments, each choosing to cooperate or defect on export controls.

• Built the payoffs from real value-chain data, not assumed numbers: the share of industry value each side captures in the segments it dominates — US in IP, design and equipment, China in materials, wafer fabrication and packaging — set against what each would spend on self-sufficiency.
• The equilibrium was mutual cooperation. Both governments do the opposite in reality, so the real work was explaining the gap. The answer was strategic mistrust: neither side trusts the other to keep cooperating, so the safe move beats the optimal one.
• Extended the matrix with a binding-commitment option to test whether enforceable terms could move the equilibrium. They could.
• Then broke it on purpose. Giving one side a small non-monetary payoff for acting against its own material interest flipped the equilibrium completely, and the recommendation with it.

Evaluated capital investment projects for a real company: valuation methods, NPV and IRR, then the business and financial risks attached to the preferred project and how to mitigate them. Half the marks were on ESG factors and whether the investment supported sustainable growth.

Valued a real company end to end — cash flows, risk and return, cost of capital, payout policy and market efficiency — and recommended to a board whether to invest. Marked partly on how the recommendation aligned with specific UN Sustainable Development Goals.

Built a long-horizon savings and investment plan in Excel, calculated the return required to fund it, and analysed how a personal risk profile and political and economic events change the outcome.

• Two of the three assessments were marked partly on ESG factors and SDG alignment inside a financial decision, not alongside it.
• All calculations built in Excel and submitted as working sheets.

Wrote a board-level report on how a component manufacturer could move from socially responsible practice to being a genuinely sustainable business, set against its local and industry context.

Examined Coca-Cola’s sustainability claims against independent reporting on its water replenishment record, and asked which stakeholders carried the cost of the gap.

• First time I had to separate what a company reports from what it can be shown to have done.
• Directly behind the responsible-consumption angle in my published research.

Analysed World Bank governance indicators in R: descriptive statistics, 95% and 99% confidence intervals, and a hypothesis test on control-of-corruption scores — written for a non-technical reader.

Built and compared five regression models on real cross-sectional data, ran residual and normality checks, and justified the preferred specification.

Team report on model specification and goodness of fit, covering multicollinearity, heteroskedasticity and endogeneity, ending in policy recommendations.

• Where I learned to test a model before trusting it — the habit behind the limits stated on every project page here.

The team proposed a machine learning solution to a real business problem — sorting 6,397 Prime Video films into four return tiers — and presented it live to lecturers and other teams: 20 minutes of presenting, then 10 minutes of questions. The work was split three ways: framing the business problem as a learning task, the data and ethics review, and model selection and evaluation. I worked on the third part, with one teammate.

• Split the films by release date rather than at random: 70% train, 15% test, 15% out-of-time. Tuning happened inside the training set only, on a forward-chaining time-series split, so no later film could inform an earlier prediction.
• Compared five models: LightGBM, XGBoost, CatBoost, random forest and logistic regression. LightGBM came first on accuracy, F1, precision and recall. XGBoost had slightly better probabilities, so we kept it as a benchmark rather than dropping it.
• Tested Platt scaling to make the predicted probabilities more reliable. It made them worse: log loss rose on every model. We reported that result and used the uncalibrated models. Writing up the check that failed was the part I learned most from.
• Ran SHAP on the winning model to see what it leaned on. Budget dominated every tier, genre came next, then release timing. SHAP shows magnitude rather than direction, so it tells you what the model uses, not what actually moves a return.
• Wrote the closing read. Accuracy held from test to out-of-time (0.627 to 0.615) but log loss rose, and the train–test gap showed the model was still overfitting. So the recommendation was triage and risk scoring, with human review on borderline films and periodic retraining. We said plainly that no causal claim was available from this design.