VolatilityRegimes: Reproducible Macro-Finance

View Repository Source

A content-hashed, reproducible macro-finance pipeline that estimated a latent uncertainty factor and decisively rejected a consumption pricing kernel.

Empirical finance is plagued by irreproducible results. This project built a fully reproducible macro-finance research pipeline driven by a single declarative registry covering sources like FRED, Yahoo Finance, and the Kenneth French library.

Finding and Failing

Using this pipeline, I estimated a latent uncertainty factor that works as a usable stress index. The regime layer built on it is not validated: its probabilistic scores are worse than a constant base rate, and its headline event-overlay evidence does not survive a placebo test. The consumption pricing kernel failed outright: it failed the Hansen-Jagannathan bound (0.275 against 0.419 required, Hansen J $p = 2.7 \times 10^{-10}$), reproducing the equity-premium puzzle.

Instead of quietly tuning the parameters to “fix” the failure, I reported the outright rejection—alongside a placebo test that invalidated my own regime-validation claim.

Engineering Against Silent Failures

The infrastructure handles the hidden traps that quietly invalidate empirical work:

  • Data revisions and point-in-time correctness (preventing look-ahead bias).
  • Mixed reporting frequencies.
  • Missing-data gaps.
  • Automatic non-stationarity screening.
  • Quarantine procedures for series failing quality checks.

To make runs checkable, finished panels are content-hashed (SHA-256) into a manifest that the data layer verifies on load, backed by a truncation test that rebuilds the panel on a shorter history and confirms it matches the full-panel slice. Carrying that same verification through to the modeling stages is the open item.