Patricio Ortiz

Basel III makes banks hold liquid assets against the LCR, NSFR and capital adequacy. In practice, though, banks pick their asset mix through rule-based heuristics that don't jointly optimize across those constraints, and that implicitly assume Gaussian return dynamics (very inadequate for modeling financial risk). We instead pose the allocation under fat-tailed scenarios, minimizing Conditional Value-at-Risk (the expected loss conditional on exceeding the β-quantile).

We set up multi-period bank liquidity allocation as a convex program carrying full Basel III LCR, NSFR and capital constraints, plus a hard per-quarter turnover budget (which bounds the rate of rebalancing, i.e. no fire sales). Solved with ADMM, and run against both parametric Gaussian and regime-tagged historical bootstrap scenarios, yielding a quarter-by-quarter feasible rebalancing path, alongside (a much more exciting) a price for the Gaussian assumption (the gap in CVaR when the Gaussian-optimal weights are evaluated against the more-realistic fat-tailed scenarios as simulated by the bootstrap). Finally, it's calibrated to the filings of a real Category III bank (PNC group for the final report, picked fairly arbitrarily, though we also tested it against three similar transpositions of the same standard with HSBC under the HKMA rulebook, DBS under Singapore's MAS, and BNP Paribas under EU CRR).

Estimates the CES substitutability parameter between bank deposits and government bonds in China. The SOE versus joint-stock difference comes out statistically indistinguishable, so the main hypothesis can't be rejected. Interestingly, estimate using distribution of heavily-public (SOE) to mostly private (joint-stock) banks in China yields a substitibulity paramater of ≈ 0.66, which is very similar to a benchmark found by Krishnamurthy and Li’s for the US of 0.65, suggesting imperfect substitutability between bank money and sovereign debt may be a general feature of modern financial systems rather than a US quirk.

Builds a micro model of worker effort under an ownership share λ. Proves the effort game has a unique Nash equilibrium for any λ, and that an optimal share exists which pays the firm strictly more than giving workers nothing at all. Paired with a survey and regression-discontinuity design as well as power estimation for testing it on real data. Ownership Works, the NGO, then was able to use our design to prove their claims for a firm.

Runs Bishop’s empirical-Bayes PCA and Nakajima’s variational-Bayes framework against classical PCA on Taiwanese bankruptcy data. The Bayesian versions lean on automatic relevance determination, which handles the two things classical PCA can’t: picking dimensionality without a hand-set hyperparameter, and saying anything useful about uncertainty.

Reads three primary sources between 1200 and 1253 (Innocent III’s letters to Canterbury, Louis IX’s expense accounts, and a Genoese chartering contract) and follows risk as it moves from spiritual obligation repriced into money, to one sovereign carrying the whole cost on his own ledger, to private counterparties splitting it by contract.

Some of the slides I’m proudest of! Two walk throughs trying to take economic theory into the real world, one in macro and one in finance. First, a trade shock propagating through a small open economy: Chinese steel output falls, Mongolian coking coal prices drop 37.8% y-o-y, the tögrög comes under pressure and the fiscal accounts follow. Second, the securitization chain of the GFC taken layer by layer, from subprime pools up through CDO², explaining leverage and how the correlation assumptions holding those products together gave way.

A four-step pipeline run on Singapore FDI data: PCA down to 95% of variance, purged k-fold cross-validation so nothing leaks across time, sequential-bootstrap bagging, then random forests. Feature importances get mapped back through the PCA rotation so they mean something in terms of the original variables (human capital, population growth, trade openness, i.e. things we might care about).

A summary of a genuinely fantastic(!) paper by Kuznetsov and Mohri in learning theory for time series (pushes Rademacher-complexity generalization bounds past both the i.i.d. and stationary-mixing assumptions, i.e. we now have path-dependent bounds, fast learning rates, and results that can exist even with unbounded losses).