Methodology

Universe & Data Sources

The stock universe is the large-cap slice of a VTI-derived warehouse — roughly 1,000 large-cap names tracked intraday, with small/mid strata followed at daily resolution. A ~95-symbol benchmark board covers the macro read: ETF proxies plus a set of bellwether single names.

Benchmark-board groups

• index — US broad benchmarks (SPY, QQQ, DIA, IWM, MDY, IJR, RSP, VTI).
• rates_credit — the duration + credit-quality ladder (SHY → IEF → TLT; LQD/VCIT/VCLT, JAAA, SRLN, FALN/ANGL, HYG, PFF).
• commodities_fx — gold, silver, oil, the dollar, bitcoin (GLD, SLV, USO, UUP, IBIT).
• global — international / EM equity (EFA, EEM, VXUS, single-country and FT funds, KWEB).
• style / sector / theme — factor, GICS-sector and industry-theme ETFs.
• single — bellwether single names, rendered last on the board: they are color, not signal.
• cross_asset — a legacy tag kept only so archived pre-expansion editions re-render with their proper label.

• Warehouse — a local DuckDB market-data store (daily bars, SMAs, RSI, MACD, Bollinger, money-flow indicators), snapshotted read-only once per day before the market opens.
• Polygon — 15-minute intraday bars (15-minute delayed) for the ETF proxies and the large-cap universe during the trading day.
• Equity-screener — wave/chart-pattern/opportunity surfaces consumed from its published artifact where tickers overlap.

Editions are generated five times per trading day: premarket 07:00 morning 10:00 midday 13:00 close 16:00 wrap 19:00 (all ET).

Factor Baskets — How They're Built

Every stock is scored nightly on momentum, trend structure, volatility posture and institutional-accumulation traits, then assigned to exactly one production basket — the behavioral taxonomy that headlines the dashboard:

Four more dimensions slice the same universe independently: theme (keyword/industry baskets), sector, size and style. A stock therefore appears in one basket per dimension — the drill-down drawer on any card shows the exact membership behind that cell.

Return Math

Basket returns are computed three ways and shown where each is honest: equal-weight (the default headline — every member votes equally), cap-weight (what the dollars did), and median (robust to outliers).

• Per-member returns are winsorized at ±400% before aggregation, so one parabolic name cannot blow up a basket mean; the median is left untouched.
• The 12-horizon registry spans intraday legs to two years: ID 1D 3D WTD 1W MTD 1M 3M QTD YTD 1Y 2Y — calendar anchors follow the NYSE calendar (holidays and half-days honored).
• Intraday legs decompose the day: gap (prior close → open), since-open, and since-prior-close.
• Relative volume (rvol) is measured against the time-of-day curve, not a naive elapsed fraction — premarket rvol is deliberately null.

The Regime Engine

Six components — trend, breadth, volatility, rotation, size, style — are each scored from banded market inputs (SPY vs moving averages, % of the universe above the 200-day, advance/decline, sector-rank churn, small-vs-mega spreads, high-beta-vs-low-vol spreads). A weighted composite maps to one of five labels:

The published label flips only after the raw label disagrees on two consecutive runs — that confirmation collapses daily churn from ~40% to ~16% and makes the label publishable five times a day.

Lifecycle stats on the hero card

Every stat is empirical, fitted on 2022→2026 daily history (~1,050 confirmed days, refit annually): p(stay) is the one-day transition-matrix diagonal, typical dwell the mean run length, the implied end date is the start date plus typical dwell in NYSE trading days, and the exit paths are where the label historically went conditional on leaving. Exits form a clean adjacency ladder — the tail states never jump straight across Choppy Neutral.

Regime progression timeline

The Regime Assessment card plots the composite trajectory of every published run over the trailing month (up to five editions a day) against the five label bands, so a same-day flip is visible as a line crossing a band edge rather than a single word that changed. The series is assembled at report-build time from the local run archive — never interpolated: an edition with no run is a gap, and the count of missing runs is printed. An amber ring on a point marks a run where the raw label disagreed with the published one — confirmation-lag pressure that has not yet (or may never) flip the published label. The vertical axis is clamped to ±1.3 (the composite's realized range with headroom to show all four band edges); a point beyond the clamp keeps its true value in its hover title. The hero also carries a compact three-day tape of the same data that deep-links here.

Analytics Panels

• Sigma Board — today's key metrics against their own 60- and 252-observation history, expressed in z-scores; |z| ≥ 2 is flagged extreme.
• Cross-Asset Confirmation — a composite of rates/credit/commodity/FX tapes that either confirms or contradicts the equity regime label, plus duration-sensitivity betas.
• Cross-Sectional Dispersion — sector and style return dispersion (stdev/IQR) vs its trailing 60/252-observation distribution: wide dispersion pays stock-picking, compressed dispersion favors index exposure.
• Actionable Trading Ideas — a fixed table of backtested playbooks (entry rule + direction + horizon + exit + invalidation) matched against today's live tape; deliberately assembled without LLM involvement.

Each playbook ships only after clearing a standardized offline backtest gate (5-year lookback, split-adjusted daily bars, non-overlapping events per name): n ≥ 80, |t| ≥ 2.0, an economic floor above assumed trading cost (0.20% multi-day / 0.06% same-day, ~10 bps round-trip), a consistent per-year sign, and — for reversion plays whose edge historically inverted in the 2022 risk-off tape — a regime block that suppresses the setup in Distribution Risk / Risk-Off Stress. Every card prints a stats bar (sample size, win rate, average and median forward return, t-stat, lookback span) straight from that backtest, plus a collapsible method/provenance receipt; the numbers are baked at refit time (quarterly) and never recomputed on the page. Three honesty disclosures are explicit: the backtest measures a fixed-horizon hold with no intraday stop (the live invalidation line is a risk overlay, not part of the measured edge); the population is the current liquid cross-section (survivorship disclosed — there is no point-in-time membership in the warehouse); and the t-stat is the naive mean/(std/√n) with per-name non-overlap only — same-day cross-name correlation inflates it on the large pooled samples, so treat |t| ≥ 2.0 as a screen, not proof, and weight the per-year sign consistency at least as heavily. When nothing triggers, the panel shows the armed playbooks and the nearest miss rather than inventing a setup.

Single-Stock Setups

A long-only confluence screen over a 4-hour-to-1-week window: a name qualifies only when its factor basket and its technical state agree. Two archetypes — momentum continuation and oversold reversion — with transparent 0–1 sub-scores (basket alignment, factor quintile, trend structure vs the 50/200-day, RSI zone, 1-week push) capped at 0.86, plus up to +0.14 of additive money-flow confirmation (MACD, CMF, Force Index, MFI, Bollinger position). Flow-confirmed names outrank flow-less ones by construction.

Guardrails — Why You Can Trust the Numbers

• Fail-closed validation: every report is structurally validated before publish; a violated invariant kills the run rather than shipping a wrong page.
• Numeric-consistency gate: every percentage in LLM narration must exist in the underlying report data (±0.15) — the model can rephrase, it cannot invent.
• Honest empty states: a missing block renders "no data this run", never a fabricated value; degraded and thin baskets are labelled as such.
• Claim eligibility: thin/degraded/disclosure baskets stay visible but can never headline best/worst claims.
• Deterministic core: all analytics are pure functions of the snapshot; the LLM narrates on top and can never fail or alter a run.