Argento: Risk–Reward Portfolio Construction

The risk–reward problem

A portfolio with a high historical return is not automatically a better portfolio. The result may have required substantially more volatility, deeper losses or concentrated exposure to a single return pattern. Conversely, the lowest-risk portfolio may not provide enough expected reward for the intended objective.

The construction problem therefore has two dimensions: seek greater reward while controlling risk. The relevant unit is the complete package, because combining portfolios changes risk according to both the volatility of each component and the way their returns move together.

This is the central role of the efficient frontier: it narrows a large set of possible combinations to those that provide a distinct risk–reward trade-off. Argento then maps those candidates to a fixed risk scale rather than selecting one portfolio for every investor.

Methodological framework

Argento separates evidence, construction and implementation. This avoids treating a strong historical record, a diversified-looking mix or a technically feasible allocation as if each were sufficient on its own.

The same sequence is used for standard and retirement accounts. The retirement version adds an eligibility requirement before construction; it does not use a different return or risk method.

Balancing longer and recent evidence

Longer histories provide context, but they can obscure a meaningful change in recent performance. Shorter histories react more quickly, but they are noisier and can overstate temporary strength. Argento keeps both views.

Let the recent annualized return differ from the full-history annualized return by Δ. Argento converts that change into a signed recency vector v(Δ): positive change points upward, negative change points downward, and the downward vector has greater magnitude for an equally sized change.

Risk is also measured over two views of the data. One estimate uses the longer available histories; the other uses the same recent history for every admitted portfolio. Package risk includes covariance among the selected components. The higher estimate governs:

Governed portfolio construction

Argento evaluates a finite set of combinations instead of solving for unrestricted continuous weights. This makes the construction rules explicit and the output reproducible.

Interpreting the output

The output is a mapping from account amount, investment track and risk level to a feasible portfolio package. It should be interpreted within the following boundaries:

• Risk levels are model targets. A higher level means a higher modeled-volatility target, not a guaranteed return or a statement about how much loss a customer can tolerate.
• Account size can change the package. Product minimums can make a candidate feasible at one amount and infeasible at another, even when the selected track and risk level are the same.
• Recent evidence affects ranking. Recent deterioration receives a larger adjustment than recent improvement, but neither is treated as a prediction.
• Retirement status changes eligibility. The retirement version restricts the starting universe to portfolios marked retirement eligible; the core construction method remains the same.
• Evaluation and implementation follow different clocks. Argento continuously evaluates new evidence and changes to the governed frontier. Portfolio allocation changes are implemented through a scheduled monthly rebalance rather than immediately when the frontier changes.

Assumptions and limitations

Argento is based on historical portfolio-level returns. Its estimates depend on the available full and recent history, the fixed risk scale, the cash assumption and the governed candidate rules. Different policy choices could produce a different frontier.

The model does not inspect the securities, sectors, factors, leverage or liquidity inside each portfolio. Covariance measures historical co-movement at the portfolio level; it does not prove holdings-level diversification. Source returns are gross of fees. Net-of-fee interpretation deducts the uniform 1% annualized advisory fee; because the same fee applies to every candidate, it lowers reported returns without changing relative rankings or frontier membership. The model does not independently account for taxes, transaction costs, capacity or future distributions.

Historical means, volatility and covariance can change. A package that was efficient in the observed data may not remain efficient in the future. CASH_HURDLE is a fixed model assumption rather than a quoted or guaranteed yield.

Problem definition and notation

A broad model-portfolio catalog creates a finite allocation problem: which portfolios may be combined, at what weights, and for which account amounts? Argento solves this catalog-construction problem. It does not infer a customer’s objective, time horizon, tolerance for loss or liquidity needs from portfolio return data.

The method is governed by five requirements: use only eligible model portfolios; retain each portfolio’s available history while evaluating common recent history; enumerate a finite weight set rather than unconstrained weights; evaluate return, uncertainty, volatility and covariance at package level; and reject packages that fail the return hurdle or product-minimum test.

Eligible universe and return normalization

Eligibility is evaluated from the latest metadata record for each product. The model accepts portfolios, not baskets. A portfolio must be visible to the applicable customer population, active, marked investible and tradeable, and supported by a parsable positive minimum investment. Versioned governance exclusions are applied before all metrics, counts and rankings.

Missing, invalid or low-coverage data-quality records fail closed. A portfolio must also have sufficient current and contiguous history through the latest completed market date. The retirement variant adds one requirement before construction: the latest retirement-eligibility field must explicitly allow the portfolio.

Source returns may occur on calendar dates. Argento maps each observed return to the same or next active market date and compounds observations that map to the same date. Missing observations are not imputed as zero:

A qualified portfolio retains its full normalized history. Pre-inception returns are not created, and the newest portfolio does not truncate the evidence retained for older portfolios. Package descriptive history uses only the actual overlap of its selected components.

Selection return and recency treatment

For portfolio i, μ̂_F,i is the annualized arithmetic mean over its own full eligible history. The model separately calculates μ̂_R,i over common recent history shared by every admitted portfolio. Define the change in evidence as Δ_i = μ̂_R,i − μ̂_F,i.

Recent evidence is represented as a signed vector v(Δ_i). Positive change creates a moderate positive vector, while negative change creates a stronger negative vector. This asymmetry allows weakening evidence to matter without allowing a short strong period to replace the longer record:

Component selection returns are combined linearly. Cash contributes its governed return assumption:

For stability review, common recent history is divided into prior and latest non-overlapping portions. Their annualized arithmetic means are emitted for each component and package. These fields are diagnostics only: they do not change the selection-return equation, covariance estimate, return hurdle, frontier or tie-breaks.

The selection return is a historical ordering statistic. It is not a forecast of future return and is not presented as an expected customer outcome.

Return uncertainty, covariance and selection risk

Return uncertainty

A deterministic BOOTSTRAP_DRAWS-draw moving-block bootstrap resamples blocks of BOOTSTRAP_BLOCK_LENGTH observations within periods that have a constant availability mask. This preserves portfolio inception boundaries and contemporaneous observations. The random seed is derived deterministically so identical inputs produce identical diagnostics.

If V̂_μ is the bootstrap covariance of annualized component means, package return uncertainty is:

The standard error is reported separately. Under the current policy, UNCERTAINTY_PENALTY is non-binding, so uncertainty does not affect admission, frontier ranking, target matching or tie-breaks. This distinction prevents a diagnostic from being mistaken for a selection input.

Cash assumption and admission score

The governed cash input is the fixed effective annual policy constant CASH_HURDLE. Its normalized per-observation return is (1 + CASH_HURDLE)^{1/ANNUALIZATION_FACTOR} − 1. Modeled cash volatility, covariance and return uncertainty are fixed by CASH_RISK_ASSUMPTIONS. The cash input is neither a live rate nor a guaranteed yield.

Unequal-history covariance

The full-history covariance matrix Σ̂_F is estimated with a monotone-ragged Gaussian method that retains valid longer histories rather than truncating all portfolios to the newest inception. A second covariance matrix Σ̂_R is estimated from complete common recent history.

For portfolio-weight vector w, each covariance matrix produces an annualized package volatility. The larger estimate binds selection:

This rule prevents either a quiet longer record or quiet recent history from making the package appear safer than the other estimate supports. If the estimates are equal, the full-history basis is retained for audit labeling.

Descriptive metrics

Historical cumulative return, compound annual growth, realized volatility, maximum drawdown, SHARPE_REFERENCE_RATE-adjusted Sharpe ratio and SPY beta are calculated separately on an initial-weight buy-and-hold path over actual component overlap. Synthetic cash is included when present. These measures explain the package but do not drive Pareto dominance or target selection.

Governed candidate set and track rules

Argento enumerates a finite candidate set rather than optimizing unrestricted continuous weights. Every candidate has a deterministic signature based on its components and weights.

The absence of an independent concentration cap is deliberate. When multiple portfolios are included, the component floors create the practical weight ceiling. A quantitatively better-fitting single portfolio may therefore beat a feasible mix. The model does not force diversification at the cost of a worse frontier position.

Total Return considers every governed candidate, regardless of strategy or risk tag. Income applies the same calculations but restricts every non-cash component to the approved Income universe. The Income label is an objective filter, not a forecast of distributions; the source does not contain a validated distribution series.

The retirement variant clones the standard policy and changes only the starting eligibility universe. Return estimation, covariance, cash treatment, frontier logic, risk targets, feasibility and tie-breaks remain the same.

Frontier construction and risk-level mapping

After candidates below the cash hurdle are removed, candidate b dominates candidate a when b has no greater selection risk and no lower cash-excess score, with at least one strict improvement:

At least one inequality must be strict. Numerical comparisons use FRONTIER_TOLERANCE.

The non-dominated candidates form the governed frontier. Near-identical candidates are ordered deterministically by lower selection volatility, higher cash-excess score, multi-portfolio status, higher diversification ratio and lexical package signature.

RISK_LEVEL_COUNT fixed annualized-volatility targets are evenly spaced from RISK_SCALE_MIN to RISK_SCALE_MAX:

For each account-amount breakpoint and track, Argento rebuilds the minimum-feasible frontier. It selects the nearest allowed frontier point to each target. Target matching first minimizes absolute volatility distance, then prefers lower volatility, higher score, a multi-portfolio package, higher diversification ratio and finally lexical signature.

The allowed frontier index never moves backward as risk level rises. Selection risk and cash-excess score must therefore remain non-decreasing. If two adjacent levels use different packages, both risk and score must increase strictly. Adjacent levels may share one package when minimums or frontier sparsity prevent a distinct improvement.

Dollar feasibility and implementation mapping

Let A be a whole-dollar account amount, q_i an integer portfolio weight in percentage points, q_c the cash weight, and m_i the current whole-dollar product minimum. Because multiplying dollars by percentage points yields cents, each component allocation can be tested without floating-point rounding:

The minimum account amount required by a package is the greatest component-specific amount needed to clear every product minimum, bounded below by the policy’s catalog floor:

Amount breakpoints begin at the catalog floor and at each cash-hurdle-qualified candidate’s minimum feasible amount. At every breakpoint, the model reconstructs the track-specific feasible set and risk ladder. Consecutive risk levels sharing the same package are grouped, and contiguous amount ranges with the same selection are merged.

A change in account amount can expand the candidate set, but it does not alter portfolio histories, selection returns, covariance estimates or fixed risk targets. This is why the same track and risk level can map to a different package at a different amount.

Each unique allocation is represented by a deterministic package signature and delivery mix ID. The application resolves account type, track, risk level and amount to the mix ID; the managed-strategy configuration then resolves that ID to portfolio weights. Retirement mappings are additive and do not alter existing standard mappings.

Reproducibility, output fields and validation

The model is deterministic. Identical source inputs and policy rules must produce identical candidate signatures, diagnostics, frontier ordering, selections and output rows. Policy changes receive a new policy version rather than silently changing prior output.

Argento runs continuously as new portfolio history and metadata become available, reevaluating eligibility, package metrics, implementation feasibility and the governed frontier. Evaluation does not itself trigger trading. Portfolio changes are accumulated for the scheduled monthly rebalance, when the latest governed mapping is applied subject to operational and customer-level controls.

Validation requires both Total Return and Income coverage across every level in RISK_LEVEL_SET. Standard and retirement catalogs are built separately, then consolidated only when they share the same source provenance. Selected packages must be cash-hurdle qualified, minimum feasible, non-dominated and monotonic in both risk and score.

Physical output rows are collapsed by repeated risk selections and contiguous amount intervals. Row counts are therefore run-specific and should not be interpreted as a fixed property of the model.

Statistical and implementation limitations

Available histories differ and may reflect different market conditions. Historical return, risk and diversification estimates can therefore change as new information becomes available. Results will also reflect the model’s policies for history, recency, risk levels and the cash-relative hurdle.

Continuous evaluation can identify a frontier change between rebalance dates. Until the next scheduled monthly rebalance, a live account may therefore retain the package selected at the prior rebalance.

The recency adjustment can still react to a temporary regime. Prior-versus-latest history fields expose changes but do not constrain selection. Bootstrap standard error is reported as a diagnostic and is not a selection penalty. The cash-excess score is a historical ranking statistic, not a return forecast.

The analysis operates on portfolio-level returns. It does not inspect underlying securities, sectors, factors, leverage or liquidity. Covariance, provider count and a multi-portfolio tie preference do not establish holdings-level diversification. Product-minimum feasibility does not establish capacity, liquidity, customer eligibility or low implementation cost.

All supplied returns are gross of fees. Net-of-fee interpretation deducts a uniform 1% annualized advisory fee from reported annualized return. Because this deduction is common to every candidate, it shifts returns downward without changing relative ordering or frontier membership. The model does not independently estimate transaction costs, slippage, taxes, funding flows or future distributions. SPY beta is descriptive and may not be an appropriate benchmark for every portfolio.

Metadata, eligibility and investment minimums are point-in-time inputs and must be refreshed before use. Model or hypothetical performance requires separate compliance review and fair presentation of material risks and limitations.

References

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.”