How the models work

Every number in this platform is the output of a disclosed model stack — the math is never hidden, and neither are the fallbacks. These are models: they simplify, they lag turning points, and they will sometimes be wrong. The same data is shown at every experience level; only the annotation changes. VYLOS research supports advisor due diligence. It is not investment advice, a forecast, or a guarantee of performance.

Tier construction

Five tiers from Ultra Conservative (1% crypto ceiling) through Ultra Aggressive (8%), with Conservative / Moderate / Aggressive at 2 / 3 / 5%. Ceilings were recalibrated 2026-05-27 to institutional fiduciary norms. Each tier mixes spot exposure, defined-outcome wrappers and income instruments; ceiling, max-drawdown cap and rebalance cadence are encoded in core/risk_tiers.py and applied before the 30% single-position cap and renormalization to 100%.

Mean-variance optimizer

Runs the real Σ⁻¹μ mean-variance allocation with a Ledoit-Wolf shrinkage covariance built from a live returns panel of universe representatives. When the full covariance is unavailable it falls back to a diagonal heuristic — and says so, never silently.

Black-Litterman posterior

A curated view library plus a custom views editor. Advisor views tilt equilibrium expected returns into a Black-Litterman posterior over the same live Ledoit-Wolf covariance, producing posterior weights shown next to the market baseline.

Regime · 2-state HMM

A Gaussian Hidden Markov Model (Baum-Welch fit, Viterbi decoding) on a ticker’s live two-year daily returns. Unlike an independent classifier it models the persistence of risk-on / risk-off states through a transition matrix, and drives the regime-conditional stress recommendation.

Stress replay

Named historical and hypothetical scenarios replayed as per-asset shocks against the tier-sized crypto sleeve; total impact is the sleeve-weighted loss against current AUM. Because the sleeve is capped at the tier ceiling, the whole-portfolio hit stays contained.

Brinson attribution

Decomposes excess return into allocation (was the tier’s sleeve mix right?), selection (did the chosen ETFs beat their category?) and interaction (the cross-term). Returns are illustrative until live ETF return history is wired — and are labeled as such wherever shown.

Research assessment v0

Leveraged structure, or fee ≥ 2× its category median▼ UNFAVORABLE

Spot structure at or below the category-median fee▲ FAVORABLE

Everything else■ MONITOR

A transparent, deterministic fee-and-structure screen computed from real universe facts (structure + expense ratio) — no invented signals. A disclosed placeholder until per-ETF composite scoring ships. Fund-level output is always a research assessment, never a buy/sell call.

Reader levels

Beginner — every jargon metric carries a plain-English annotation.

Intermediate — condensed interpretations with the key values visible.

Professional — raw values, formulas and model diagnostics inline.

Same data at every level — nothing is dumbed down, only annotated. The Cornish-Fisher card below changes with this toggle.

Technical composite (engine)

The Phase-1 signal blends three technical indicators per ETF — RSI(14), MACD histogram and 20-day momentum — into a unified score in [−1, +1], with ±0.30 thresholds. A Phase-2 model extends it with on-chain and macro layers weighted by regime state. In this advisor UI it is a research input only; it never surfaces as a fund-level buy/sell instruction.

Risk measurement · Cornish-Fisher VaR

We estimate worst-case loss with a method that adjusts for crypto's non-normal return distribution — more accurate than standard VaR for fat-tailed assets.

Fiduciary-appropriate filter

By default, leveraged ETFs and single-stock covered-call wrappers are excluded from the candidate set at basket-build time, before optimization, regardless of tier. Toggleable in Settings for clients whose IPS authorizes an aggressive sleeve. The filter keeps generated baskets within the bounds typical FA fiduciary duty allows without further client-specific authorization.

Data sources

SEC EDGAR · N-PORT — ETF composition and AUM data
yfinance — price history, expense ratio, volume averages
VYLOS composite signal — per-ETF technical signal (RSI / MACD / momentum)
News feed — 15-minute cached headlines; non-fiduciary input
Broker — mock for demo; alpaca_paper for paper trading. Live execution (alpaca_live) is architected but hard-disabled today; all order paths are paper/mock only.

Performance-display compliance

Per the SEC Marketing Rule, every performance figure shown in this app is hypothetical and for illustration only; it is presented across multiple horizons (1Y / 3Y / 5Y / since inception), compared against a benchmark, and accompanied by max drawdown. Past performance does not guarantee future results.

Hypothetical results. Past performance is not indicative of future results.

Why this matters

The stack exists so an advisor can see why a number is what it is before acting on it. Every figure traces to a disclosed model with disclosed inputs, and when a model falls back — diagonal covariance, thin ETF history, illustrative returns — the page says so. The models will be wrong sometimes; the decomposition and the disclosures are there so the wrongness is visible, not hidden. VYLOS research supports advisor due diligence. It is not investment advice, a forecast, or a guarantee of performance.