Allocation frameworks

Definition

Allocation frameworks decide how capital divides across strategies, assets, and theses — the layer above individual position sizing. The canon runs from Markowitz mean-variance optimization (1952 — the founding mathematics) through risk parity, volatility targeting, and fixed-weight schemes, with one humbling empirical anchor: DeMiguel-Garlappi-Uppal (2009) found naive equal weighting (1/N) beat fourteen optimization models out of sample — estimation error in the inputs routinely exceeds the value of the optimization.

How it works / structure

• Mean-variance and its failure mode: optimal weights from expected returns, variances, and correlations — maximally sensitive to expected-return estimates (the least estimable input); unconstrained optimizers produce extreme, unstable weights (“error maximizers” — the standard critique the 2009 result quantified).
• The practical spectrum (engine-executable): fixed weights (1/N or policy weights — robust, ignores information), risk-based weights (inverse-vol, risk parity — equalize risk contributions, no return estimates needed, risk-volatility-targeting machinery), constrained optimization (bounded weights, shrunk estimates — the professional compromise), and strategy-level budgets (the platform’s frame: capital allocated to STRATEGIES with replay-measured properties, then within-strategy sizing — risk-fixed-fractional).
• The evaluation-window layer: allocation reviewed at thesis horizons, not daily (bias-loss-aversion myopia interaction); drawdown budgets partition the total (risk-max-drawdown-budget).
• Attribution discipline: returns decomposed to allocation vs selection vs timing per period — the accounting that grades the framework itself.

When it applies

Multi-strategy books (the platform’s normal state); the one-decision layer users actually control at long horizons (the allocation literature’s consistent finding: the allocation dwarfs the selection for diversified long-horizon outcomes); regime transitions (frameworks that re-weight on regime states are theses and get replayed as such).

Risk profile & failure modes

• Optimizer worship: feeding noisy expected returns into unconstrained mean-variance produces confident nonsense — the DGU result is the standing rebuttal; constraints and shrinkage or don’t bother.
• Risk-parity leverage import: equalizing risk across low-vol assets historically leveraged bonds — rate-regime concentration wearing diversification’s clothes (regime-rate-environments 2022 lesson).
• Framework churn: switching frameworks after each drawdown is bias-recency at the allocation layer — the framework’s replay window must exceed its rebalance horizon by an order of magnitude.
• Correlation staleness: all frameworks import a correlation matrix; stress rewrites it (port-correlation-budgets).

Evidence & limits

Markowitz is the founding theory (Nobel 1990); DGU (2009) is the peer-reviewed out-of-sample humbling; risk-parity and vol-targeting evidence is documented with regime caveats in their entries. No framework claims dominance — the platform’s requirement is that whichever framework runs, its rules are explicit, replayed, and attributed.

Falsifiable-thesis examples

Illustrations only, not signals:

• “The constrained-optimization allocation will beat 1/N across these strategies on 3-year replay Sharpe” — falsified by the paired replay.
• “Strategy-level risk budgets will keep realized portfolio vol within ±25% of target in three of four quarters” — falsified by the realized series.

Cross-references

• The mathematics: port-diversification-math; the matrix risk: port-correlation-budgets
• The mechanisms: port-rebalancing, risk-volatility-targeting, risk-max-drawdown-budget
• The behavioral interface: bias-loss-aversion (evaluation windows), bias-recency (framework churn)