Regime Alpha

The problem

Full-period averages can hide changes in portfolio behavior. Holdings that appear diversified across an entire history can become more closely related in particular market conditions. The research question is how those changes affect risk contributions, volatility, and drawdowns.

My contribution

I develop the quantitative models, Python data pipelines, and analytical workflows, and review AI-assisted code changes in GitHub. The platform connects market and macroeconomic inputs with portfolio factor decomposition and diagnostics conditioned on market regimes.

My research includes Gaussian and Bayesian mixture models, hidden Markov models, stochastic state transitions, and interpretation of regime probabilities. The FRED and Alpha Vantage pipelines include quality and readiness controls intended to keep stale or invalid inputs out of model builds.

A technical decision: separate edits from results

I design the workflow around an explicit model build, with draft inputs kept separate from a completed analysis. That distinction matters: editing a holding or model setting must not silently change the meaning of a result that was produced earlier.

The implementation compares a signature of the draft configuration with the last successful run. A successful build records its configuration, result artifacts, build time, and data date. Navigation or a failed rebuild preserves the prior successful artifacts.

Existing lifecycle tests cover deterministic signatures, changed-input detection, valid explicit build requests, preservation after a failed rebuild, and snapshot metadata. These checks address consistency of the analytical workflow.

Inspect the public example

The read-only sample follows two fictional holdings, each starting at half the portfolio’s capital. It makes the difference between capital allocation and signed risk contribution visible, then compares the same history across saved market states.

1. Start with the overall review: compare capital weights with each holding’s contribution to modeled variance.
2. Compare State 0 and State 1. Read absolute contributions and total risk alongside percentage shares; inspect the reported sample support.
3. Open the exact tables to see signed allocations, dated inputs, and the factor, residual, and cross-term decomposition.

What the evidence supports

These diagnostics are historical and model-dependent. The public sample is precomputed and makes the calculations inspectable without access to the research workspace; it does not establish causality, forecast calibration, or investment outperformance.