Trust & Transparency

Portfolio Health Score

Model portfolio_health_score@2.1.0 (provisional-basis release). A 0–100 composite of six components: volatility (0.20), tail risk / CVaR (0.25), concentration (0.15), correlation stress (0.15), drawdown depth & proximity (0.15), and data confidence (0.10). Data confidence is coverage-scaled — when the covariance path can only measure a subset of holdings, the component is scaled by that coverage rather than silently reported at full weight.

Last internally reviewed: 2026-07-04

Decision Readiness

Evaluates regime clarity, risk budget utilization, and market timing factors to indicate when conditions favor deliberate decision-making.

Risk Metrics

Value at Risk (VaR) via historical simulation, portfolio volatility, and maximum drawdown calculated from your actual positions.

Stress Tests & Scenarios (deterministic)

Reviewed factor shocks are applied to portfolio weights and reconciled to holding and factor contributions in percent and dollars. Built-in, custom, and story scenarios use the same engine; story AI only structures editable assumptions. Rates and credit require separate duration inputs, missing coverage is explicit, and the result assigns no probability, confidence interval, correlation multiplier, or calibrated uncertainty band.

Last internally reviewed: 2026-07-04

Monte Carlo Simulations (probabilistic)

Path simulation from your holdings' actual return history: the covariance matrix Σ is estimated with Ledoit-Wolf shrinkage on a 252-day window, with a one-shot regime uplift applied to Σ up front. Shocks are drawn as a Student-t (df = 5) variance-mixtureto model fat tails, and paths are generated with antithetic sampling for variance reduction. Default run: 10,000 paths. When per-holding history is insufficient, coverage is disclosed and an asset-bucket fallback is used for the affected holdings — never silently.

Last internally reviewed: 2026-07-04

Portfolio Optimizer

Solver-based Min-Variance / Risk-Parity on the measured covariance matrix, with hard position and sector caps. When a sector cap is binding at the optimum, we surface it on the result ("binding: {Sector}").

Last internally reviewed: 2026-06-30

Behavioral Analytics

Confidence-tiered constructs: fewer than 10 recorded actions → no numeric score (patterns only); 10–29 → directional bands (e.g. "80–95"); 30+ → numeric scores. We never fake precision at low sample sizes.

Last internally reviewed: 2026-06-30

Tax-Loss Harvesting

Losses are bucketed into Short-term, Long-term, and Term unverified. Positions with estimated (rather than real) purchase dates are never folded into Short/Long — they surface with a fix note explaining how to promote them.

Last internally reviewed: 2026-06-30