Hierarchical Risk Parity (capped)

A methodology's `rules` and `universe_filters` are declared, merged, and hashed into its identity, but no pipeline step evaluates them. They have zero references in the executor. A reader seeing rules: - id: concentration_cap expr: max_asset_weight <= 0.25 reasonably concludes the constraint is enforced by that rule. It is not. In hrp@3 the constraint happens to hold because `contract.weight_bounds.max` is also 0.25 — the rule is decorative, and the two could disagree silently. This inverts the defect class the project has found repeatedly. The usual pattern is a hidden choice that takes effect without being declared. Here a declared rule appears to take effect and does not, which is arguably worse: declaring it creates the belief that it is checked. Surfaced by adding a second methodology family. xsmom@1 declares `has_full_formation_history` and a `min_history` rule; neither runs, and the history requirement is enforced only incidentally by an ExecutionError inside compute_momentum.

On this version: Declares concentration_cap <= 0.25, which is satisfied only because contract.weight_bounds independently enforces the same number.

Hierarchical risk parity degenerates into a cash proxy when the universe contains a near-zero-variance instrument and no per-asset ceiling is imposed. Recursive bisection allocates inverse-variance, so an instrument whose variance approaches zero absorbs an unbounded share of the portfolio. This is a property of the method, not an implementation error, and the source paper does not address it because it demonstrates HRP on assets of comparable risk.

On this version: Created in response, imposing a 25% per-asset ceiling — the smallest change that makes the allocation describe the methodology rather than describing BIL.

A price panel joined across instruments with different trading calendars inherits the union of their sessions. Forward-filling a five-session instrument onto a seven-day index manufactures zero-return observations that are not market data — deflating realized volatility and mis-scaling any annualization that assumes a session count the data does not have.

On this version: Annualized 1.3664% -> 2.2892% and volatility 3.1879% -> 4.0939% once weekend padding was removed; a further correction to 2.2094% once NYSE holidays were excluded.

Claims this rests on

Assumptions it depends on

lookback

504 trading_days

linkage method

single

correlation distance

sqrt_half_one_minus_rho

covariance estimator

sample

max turnover

0.25 fraction_of_notional

rebalance cadence

21B

weight bounds

[0.0, 0.25]

gross leverage max

1.0

universe

SPY, SH, TLT, TBT, LQD, DBC, GLD, HYG, BIL

content hash

20238973e509efc2bc5aa03be9e8dd6d…

• Close-to-close execution at the next session's price.
• Correlations estimated over the trailing window are informative for the next rebalance.
• No borrow constraints or short availability limits (the universe is long-only ETFs).
• A 25% ceiling is binding often enough to diversify without overriding the method.

• Long-only; expresses bearish views solely through inverse ETFs already in the universe.
• No regime conditioning — allocation responds to correlation structure alone.
• Estimated on daily Adj Close, which is restated over time by the data vendor.
• Ignores capacity and market impact.
• The 504-day lookback needs two years of history before the first allocation.
• The per-asset ceiling is an imposed constraint, not part of Lopez de Prado's published method; results are therefore not a replication of that paper.