A portfolio review, with the evidence attached.

Portfolio risk

Equal capital does not allocate equal risk.

Closing weights as of 2025-02-03. Fictional symbol identifiers are not actual securities or suggested allocations.

OLS coefficients fitted 2024-01-03–2024-10-03 and frozen thereafter. Sleeve loading is decimal asset return per unit decimal fictional-sleeve return; intercept is decimal return per session. Historical association, not a forecast.

Variance contributions: decimal return squared per session. Volatility contributions: decimal return per square-root session. Variance share is a signed fraction of total variance. Negative contributions are retained. Asset variance contributions sum to 1.8495206e-05, the same variance underlying the headline.

Measured fixed-closing-weight variance: 1.8597123e-05. Model-implied variance: 1.8495206e-05. Both use decimal return squared per session; shrinkage accounts for their difference. The actual hypothetical buy-and-hold history has drifting weights and is a separate quantity.

Conditional evidence

What varies across the estimated states?

State IDs have no permanent economic identity. A two-component Gaussian mixture observes absolute fictional sleeve returns. These are same-day historical associations; they are not tradable signals or forward stress forecasts. The training period is in-sample.

Weight ESS is not a count of independent observations. Unconditional moving-block sensitivity uses 100 fixed-model resamples in 10-session blocks. The 5th–95th percentile range for portfolio principal eigenspaces is 1.32–1.48. This does not capture uncertainty in model selection.

Incremental regime value: unavailable. This fixture checks mechanics. It does not test economic predictability, allocation benefit or outperformance.

Review questions

Questions before conclusions.

No previous comparable run available.

• Why do equal initial capital weights produce unequal variance allocations in this generated example? Inspect closing weights, signed loadings and residual covariance.
• Do the historical conditional concentration and stability estimates justify closer review? They are not forward stress forecasts.
• Confirm the portfolio mandate and data coverage. A readiness pass means usable data, not allocation advice.

Method & reproduction

A small, inspectable specification.

Signed OLS with an intercept and one fictional common return sleeve, fitted 2024-01-03 through 2024-10-03; coefficients and training standardization are frozen thereafter. The mixture uses the same training cutoff and seed 426. No favorable-state policy is selected.

Risk uses the full historical window to 2025-02-03: the joint factor/residual covariance includes cross terms, with 20% shrinkage. PCA concentration is 1 / sum(q²) across distinct eigenspaces, where q allocates portfolio variance to principal directions. No causal-factor interpretation.

The generated total-return indices contain no real corporate actions. The example models no fees, transaction costs, deposits, withdrawals or taxes. It is not an account record. No licensed market data is redistributed.