Marks Guo.

A production-grade, automated risk system computing Basel III FRTB Internal-Models-Approach capital — Value at Risk, Expected Shortfall, stress calibration, and a Standardised-Approach comparison — on a multi-asset book, served to a live, deployed dashboard.

• 97.5% Historical-Simulation VaR & Expected Shortfall (with parametric & Monte-Carlo Student-t comparisons), stress-calibrated and liquidity-horizon-adjusted, across a 6-asset book (29,000+ daily obs, 2007–26).
• Computes an Internal-Models vs Standardised-Approach capital charge under the Basel III output floor — here the SA floor binds at 11.2% of notional.
• Weekly Acerbi-Szekely ES backtest with Kupiec & Christoffersen tests; surfaced a 54%-of-weeks failure rate revealing the procyclicality of rolling-window ES.
• Engineered end-to-end: pytest + GitHub Actions CI, a Dockerized stack, and a daily cloud run (Actions cron) whose event detector flags regime shifts, ES spikes & backtest breaches, then auto-publishes Claude-written risk-desk recaps with rendered risk cards.
• Architected by directing Claude Code (agentic AI), with every risk function validated against hand-checked results as the human-in-the-loop.

• ES vs VaR — average tail-day loss vs the loss threshold, over 252 days.
• Volatility regime — shaded background: green calm, amber elevated, red stressed.
• Asset contribution — which holding drives today's potential bad-day loss.
• Backtest (Z2) — weekly model check; green passes, red underestimates risk.

An automated macro dashboard that quantifies how the official U-3 unemployment rate systematically understates labor market stress — and tracks the leading indicators that actually signal turning points.

Built and compared three regression families on 4,800+ daily S&P 500 returns (2007–2026) to predict realized volatility and quantify risk persistence — the empirical backbone of any dynamic VaR framework.

• Confirmed statistically significant volatility clustering (p < 0.0001); directly supports dynamic VaR and portfolio risk frameworks used in industry.
• Validated via 70/30 train-test split with full diagnostic suite — heteroscedasticity testing, ACF analysis, Cook's distance influence detection.
• Compared OLS, WLS, and Robust M-estimation to surface which assumptions hold once you stop pretending returns are i.i.d. normal.

Implemented a hierarchical Bayesian model from scratch with a custom Metropolis-Hastings sampler — not a library call — to estimate group-level probabilities and quantify uncertainty rigorously.

• Ran 4 Markov chains × 100,000 iterations each; tuned proposal step sizes for healthy acceptance rates.
• Confirmed convergence via Gelman-Rubin diagnostics (PSRF = 1.0); effective sample sizes exceeding 26,000 across all chains.
• Delivered findings with 95% credible intervals and actionable business recommendations translated for non-technical stakeholders.

Built a 5-year pro forma model and DCF valuation using CAPM-derived WACC across multiple macroeconomic scenarios; stress-tested beta, WACC, and terminal growth via sensitivity tables. Closed with a structured equity research report and a supported “Hold” recommendation.

• Trend, common-size, and ratio analysis across profitability, liquidity, leverage, and default risk.
• Scenario-conditional WACC and terminal growth bands to make the recommendation robust to the macro assumption.

Collaborative research with Prof. Michael Rizzo across 11 universities. Identified a significant correlation between endowment-per-student and U.S. News rankings — presented to Economics faculty with data-driven conclusions on institutional resource-allocation tradeoffs.