Argento: A Governed Framework for Multi-Portfolio Package Construction

Problem definition

A broad strategy catalog creates a finite allocation problem: which products may be combined, at what fixed weights, and for which account sizes? A useful solution must distinguish portfolio construction from customer selection. Argento addresses the first question. It constructs a governed shelf of packages; a separate fit process may later consider a customer’s objective, time horizon, tolerance for temporary loss, funding plan, withdrawal needs and exclusions. Time horizon and risk tolerance are personal inputs to allocation decisions,^[3] but they are not inferred from portfolio return data.

The method follows four requirements: use only approved model portfolios; compare all candidates on the same history; prevent free-form weights and excessive concentration; and reject packages that cannot satisfy the recorded product minimums. The resulting labels—Conservative, Balanced and High Growth / High Risk—are relative positions within the governed catalog, not universal risk classifications.

Eligible universe and common history

Eligibility is evaluated from the latest metadata record for each product. The initial model accepts only records with product_type = PORTFOLIO. Baskets are excluded before all counts, metrics and rankings. A portfolio must be public or subscriber-visible, not deleted, marked investible and IBKR-tradeable, and have a parsable positive USD minimum. Missing or invalid data-quality records fail closed; a valid record must not be marked low-coverage.

All history-qualified portfolios are restricted to the latest common complete window, capped at 180 XNYS sessions. This prevents candidate statistics from being computed over different market intervals. It also makes the effective sample equal to the shortest qualifying common history.

Return normalization and risk estimation

Source returns may occur on calendar dates. Argento maps each observed return to the same or next active XNYS session and compounds returns sharing a session. Missing source observations are not imputed as zero.

For a fixed weight vector w, session package return and annualized volatility are:

Σ̂ is the sample covariance matrix of aligned session returns with denominator T − 1.

Historical annualized return is elapsed-time CAGR rather than a 252/session-count scaling. If R is cumulative package return and D is inclusive calendar-day coverage:

The output also records maximum drawdown, zero-rate Sharpe, SPY beta, diversification ratio and maximum pairwise correlation. These are descriptive historical statistics. Only annualized return and volatility determine Pareto dominance.

Governed candidate set

Argento enumerates a finite candidate set rather than optimizing unrestricted continuous weights. Each package contains two or three distinct portfolios. Weights are positive multiples of 10%, sum to 100%, and may not exceed 60% for any portfolio. Aggregate weight from one provider is also capped at 60%. The maximum pairwise historical correlation must not exceed 0.70.

Strategy and risk tags do not define eligibility or package composition. A Growth tag may break a tie for High Growth, and a Balanced tag may break a tie for Balanced, but only after distance to the volatility target. Diversification ratio, zero-rate Sharpe and deterministic package signature provide subsequent tie-breaks.

Frontier selection and objective mapping

Let candidate b dominate candidate a when b has no greater historical volatility and no lower historical annualized return, with at least one strict improvement:

At least one inequality is strict; the numerical comparison tolerance is 10⁻¹².

The non-dominated candidates form the governed frontier, consistent with the risk–return tradition of portfolio selection.^[1] Catalog-wide volatility targets are derived once from the full governed frontier:

For each account size, Argento first filters the candidate set for dollar feasibility and rebuilds the frontier. Conservative is its minimum-volatility point. Balanced is the nearest distinct internal point to τ_B. High Growth / High Risk is the nearest later point to τ_H. A complete ladder must contain three distinct, non-dominated packages with strictly increasing historical volatility and return.

Income is not treated as another volatility rung. The verified input has no cash-yield or distribution series; therefore Income is always emitted as OBJECTIVE_DATA_GAP. A title or metadata tag is insufficient evidence of distributable income.

Dollar feasibility and reproducibility

For account amount A, percentage weight q_i and current product minimum m_i, every component must resolve to a whole-dollar allocation and clear its minimum:

A change in account size may expand the feasible candidate set but does not alter candidate return histories. The catalog-wide targets remain fixed for the snapshot, while the feasible frontier is account-specific.

Every row records the snapshot run and timestamp, source identifier, input SHA-256, policy version, account amount, selection status and reason, universe counts, component allocations, historical window, metrics and calculation rules. Stable ordering and package signatures ensure identical inputs and policy produce identical output.

Verified implementation snapshot

The following values describe one frozen run. They are evidence of model execution, not current availability or performance.

The common window contains 136 completed sessions and 197 calendar days. Four otherwise eligible portfolios were excluded for insufficient contiguous history; no portfolio was excluded by the data-quality flag. Because the observation window is short, its annualized returns are not presented as marketing performance.

Limitations

The risk model is estimated from a short historical window and may reflect a narrow market regime. Covariance, correlation, beta, drawdown and frontier membership are not stable parameters. Historical CAGR is used to describe and order observed candidates; it is not an expected-return estimate.

The analysis is conducted at portfolio-return level. It does not inspect securities, sectors, factors, leverage or liquidity inside a component. Provider and pairwise-correlation caps reduce specified forms of concentration but do not establish holdings-level diversification. Minimum feasibility does not establish capacity, liquidity, customer eligibility or low implementation cost.

Argento does not independently model advisory fees, transaction costs, slippage, taxes, cash flows or future distributions. SPY beta is descriptive and may not be an appropriate benchmark for every product or objective. Metadata and minimums are point-in-time values and require refresh before use.

References and implementation

1. Markowitz, H. (1952). “Portfolio Selection.” The Journal of Finance, 7(1), 77–91. doi:10.1111/j.1540-6261.1952.tb01525.x.
2. Sharpe, W. F. (1966). “Mutual Fund Performance.” The Journal of Business, 39(1), 119–138. Author-hosted reprint.
3. U.S. Securities and Exchange Commission, Investor.gov. “Asset Allocation and Diversification.”
4. U.S. Securities and Exchange Commission. “Investment Adviser Marketing: A Small Entity Compliance Guide.”
5. Implementation: research/pm_research/portfolio_packages.py. Verified output: docs/results/model-portfolio-packages-20260806T1200Z.csv.