RateWalk: Forecasting Fed Decisions
Naive inflation-regime conditioning overfits; empirical-Bayes shrinkage fixes it in the US, UK and Germany, and a blend with a market proxy reaches log-loss 1.056 and a 72.7% hit rate.
The U.S. Sovereign Debt Doom Loop: A Point-in-Time Framework for Identification, Market Pricing, and Policy Response
When do US debt dynamics turn self-reinforcing, and what does each way out cost?
Higher debt can push rates up, and higher rates push debt service and future debt up further. The thesis does not forecast a crisis. It measures, in real time, when the dynamics become explosive, how markets price that risk, and what each policy response costs, using only information that was available at each historical date.
Three questions, one framework
Identification
Tests in real time when debt dynamics turn self-reinforcing. The marketable Treasury debt is rebuilt bond by bond from 11,000+ auctions since 1979, and 39 macro series are read from their full revision histories, so no historical judgment sees data that had not yet been released.
Market pricing
Reads fiscal stress through the rates market: the Treasury volatility risk premium (implied versus realized) and a regime-switching stress state, each tested as a leading signal of the fiscal gap under multiple-testing control.
Policy response
Simulates 30-year debt paths under two independently estimated macro engines, a vector autoregression and a regime-switching hidden Markov model, on common random numbers, pricing five policy responses against four shocks, down to sector employment and recovery times.
Scale
11,000+
Treasury auctions since 1979, rebuilt bond by bond
39
macro series with full revision histories
2
independently estimated macro engines
5 x 4
policy responses by shocks, on common random numbers
How the work is held to account
- Every assumption is either tested out of sample and corrected for multiple testing, or stated as a range from the literature.
- Design choices are challenged by an LLM agent over a purpose-built 147-paper retrieval corpus (hybrid vector and keyword search with an in-corpus citation graph).
- No-lookahead invariance tests guard the data layer; the engine carries 48 passing tests and stamps a configuration hash on every result.
- Paths that break hard feasibility are rejected and reported with their reasons, never clipped.
Results will appear here once they pass the thesis quality gate: every claim is tested, stated conditionally, and reviewed by my advisor before it is published anywhere.
Built on
The research that led here
The thesis combines methods developed in these papers: first-release macro data, time-varying econometrics, and the volatility risk premium.
Too Big to Co-Move
Concentrated sectors co-move less with the rest of the market: a one-point rise in top-three share lowers monthly R² by 0.75 points (permutation p = 0.04), though concentration does not explain why markets disconnected.
VolEdge: The Variance Risk Premium
A full-stack platform that extracts risk-neutral moments from live option chains without a pricing model; a GARCH vol-targeted sleeve beat SPY on risk (Sharpe 0.85 vs 0.68) while naive short-variance harvesting lost money.