Seven studies connect traceable data and fixed model comparisons to practical decisions. Historical evidence, uncertainty and scenario assumptions stay visible. Summer 2027 internships.
Choose a role to start with two complementary studies. Each study links directly to its findings, methods and implementation.
Showing all seven studies.
300,000 mortgages and 20.15 million monthly records. Chronological SQL snapshots, censored outcomes, fixed models and economic-stress validation. Better held-out discrimination did not justify unchanged probabilities: calibration failures are disclosed.
Methods Chronological SQL cohorts · censored event labels · logistic regression and boosting · paired bootstrap evaluation
Credit risk, model validation, risk analytics
French motor pricing and a separate US auto-reserving study. Fixed pricing comparisons and reserve interval scores expose conditional-calibration and coverage failures. The two populations are not joined.
Methods Poisson–Gamma GLMs · calibrated boosting · Mack prediction error · paired bootstrap interval scores
Actuarial analytics, insurance risk, uncertainty modelling
A 64,000-customer, three-arm randomized experiment. Frozen policies and doubly robust evaluation did not establish that personalized targeting beats the selected simple campaign. Explicit margin and contact costs are scenarios.
Methods Randomized three-arm design · cross-fitting · doubly robust policy value · paired inference
Product analytics, experimentation, causal inference
Interest-rate hedging, published Treasury curves and hypothetical benefit cash flows. Four hedge policies, 56 rate stresses and 108 liquidity-runoff cases. Independent Decimal and formula-workbook reconciliations support the modelling.
Methods Svensson curves · constrained hedge selection · key-rate sensitivities · block bootstrap · cash-flow conservation
Treasury, ALM, financial risk
210 historical M5 item-store series link probabilistic forecasts to replenishment under common capital/capacity constraints. Fifteen final weekly origins and overlapping horizons limit inference. Operating effects are simulations, not observed retailer savings.
Methods Quantile boosting · rolling forecast origins · paired block inference · integer-order heuristic under shared constraints
Business analytics, forecasting, operations and working capital
48 official releases and 144 vintage observations. An 18-release final check retains persistence: RMSE 0.8185 pp versus 1.6901 for fixed AR(1). Release dates and the later base revision remain separate. No current inflation forecast is claimed.
Methods Release-vintage as-of joins · Ridge AR(1) · closed-form coefficient reconciliation · circular-block inference
Economic research, time series, publication-aware data
Nine issuer reports, a traced financial spread and a funded-growth policy model. Final extraction: both methods support 33/36 financial targets correctly, with three justified abstentions. 519 independent valuation checks and a reconciled formula workbook.
Methods Source-cell extraction · accounting identities · incremental investment and ROIC · reverse valuation · independent Decimal mathematics
Corporate finance, FP&A, financial research and finance data