How the numbers are made.

Each chart form is licensed to make exactly one kind of claim. Picking the form is picking the argument, so the mapping is fixed:

Titles state the finding as a sentence, computed from the data where possible, so the headline can never disagree with the chart. The subtitle is the metric plus its units. A source line appears on every chart, always. Series get direct labels at their endpoints rather than a legend the eye has to shuttle to and from.

The dashboards print these beside their charts; this is the canonical list. Reproducibility is the brand: anyone with the named sources can rebuild every number below.

Coverage tiers

Tier 1 = the original desk universe plus the highest observed US-listed dollar-volume additions until n=250. Tier 2 = the next 3,000 listed securities by liquidity rank. Tier 3 = every remaining registry entry.

Liquidity rank is descending last sale multiplied by same-observation share volume. A missing comparable observation stays null. Tier 1 is prerendered. Tier 2 and Tier 3 HTML is rendered at the edge and cached for 24 hours. Quotes enhance on request through the 30-second quote cache.

Source: SEC company tickers, Nasdaq Symbol Directory, and Nasdaq stock screener

Delta Layer

Change = current minus the latest final observation at or before each calendar boundary. Z-score = (current minus prior-window mean) divided by prior-window sample standard deviation. Percentile = midrank of current against the prior final-observation window.

The default window is 252 observations with a minimum n of 20. The current observation never enters its own reference distribution. Partial observations are displayed as current context but excluded from history. Exact retained Machine Floor dates provide content-addressed regime, breadth, and SPX options evidence with a payload checksum. Other historical requests name their latest-history or FRED-vintage revision policy. Every result reports its actual sample size and baseline date.

Source: The source named on each underlying metric

Confluence Engine

One ordered daily vector: five-leg regime composite, SPX net GEX, SPX max-pain distance from spot, VIX3M/VIX slope, 2s10s, 5s30s, commercial COT percentile for each covered contract, HYG/LQD, broad dollar, and RSP/SPY breadth.

The artifact stores raw history once at /data/confluence-vector.json. Current dimensions add z-score, percentile, lookback, source date, and sample size from the shared Delta Layer. Missing observations remain null. Options history begins with observed daily builds because the delayed public feed does not expose a historical chain archive.

Source: FRED, Yahoo Finance, CFTC, and CBOE delayed options

Analog Engine

Cosine distance across shared finite dimensions after each daily state vector is standardized against only its prior 252 sessions. Candidates need at least 18 shared dimensions and are selected with a five-session separation.

Returns are measured 1, 5, and 21 sessions after each analog. The interface prints the full histogram, positive hit rate with a 95% Wilson interval, median, worst decile, range, and sample size. Net GEX and max-pain distance stay out of historical matching until point-in-time options-chain history exists. Conditional history, not prediction.

Source: Confluence Engine, S&P 500 cash index, and E-mini S&P 500 continuous futures via Yahoo Finance

Falsification Ledger

Freeze each published signal definition, reserve the first 1,260 sessions as walk-forward warmup, evaluate successive 252-session test folds, and retain at most one event in any 21-session window. Publish performance only when at least 50 independent events remain.

Hit rate uses direction-adjusted returns above zero with a two-sided 95% Wilson interval. Average raw return and average direction-adjusted return are both shown at 1, 5, and 21 sessions. Maximum drawdown compounds the non-overlapping 21-session event sequence. Failure regimes need at least 10 events. Withheld signals expose counts and the refusal reason, never performance statistics.

Source: Published Pine definitions, Confluence Engine history, CFTC, FRED, and SPX/ES daily closes via Yahoo Finance

Daily Thesis

Select the finite current state-vector dimension with the highest absolute z-score. Transform mapped component z-scores into risk direction and classify absolute component z-scores below 0.50 as no vote. Attach the 50-state Analog Engine distribution and set the anomaly prior-window mean, z=0.00, as the invalidation level.

The broad dollar z-score is sign-inverted for risk direction. Other single-direction components retain their sign. Cross-asset COT receives no vote because its contracts do not share one risk direction. Agreement counts exclude the anomaly component. The falsifier uses the source cadence and the same prior observation window as the anomaly z-score.

Source: Confluence Engine and Analog Engine artifacts

Oracle

Classify each question as lookup, screen, or explain. Lookups resolve an exact ticker or company name against the weekly registry and read its tiered symbol contract. Screens compile to the public screener query grammar, apply the same registry filters and sort, and publish the reproducible URL. Explanations resolve one attributed state-vector dimension, compute its retained history and analog distribution, check the falsification ledger, and state the next threshold that changes the read.

Every assertion carries an inline evidence link and formula. Lookup answers state the coverage tier. Screen answers retain the compiled filter URL. Directional explanations include a falsifier. A missing contract field, licensed source, historical series, or required artifact returns a named honest null. The event stream sends each citation record before the answer token that references it.

Source: Canonical universe registry, tiered symbol contracts, Confluence Engine, Analog Engine, Falsification Ledger, and each fact's named primary source

Machine Floor v2

Remove volatile build timestamps from each institutional payload, serialize the remaining payload with JSON.stringify, encode it as UTF-8, and compute SHA-256. Store the payload in an immutable URL containing that hash, then record the exact State Vector, Analog Engine, Falsification Ledger, and Daily Thesis references in one daily snapshot manifest.

The stable latest URLs are mutable pointers. Historical evidence uses content-addressed archive URLs. The thesis scorecard evaluates the first finite state-vector observation after a daily thesis. Weekly or intraday falsifiers wait for the first distinct finite value so a carried observation is not counted as a new print. A current thesis remains pending with n=0 until an eligible observation exists.

Source: Confluence Engine, Analog Engine, Falsification Ledger, and Daily Thesis build artifacts

Macro regime composite

Equal-weight mean of the finite reporting legs: volatility, breadth, high-yield credit spread, 2s10s yield curve, and broad dollar index, each normalized to 0-100

Non-finite observations do not enter the mean. Two or more missing legs mark the composite degraded.

Source: FRED, Yahoo Finance breadth

Cross-asset correlation

Pearson correlation over up to 60 paired daily log returns, aligned from the latest shared positional window

A return is excluded from both sides when either paired close is nonpositive or non-finite.

Source: Yahoo Finance daily closes

Risk Dial

Equal-weight mean of six 0-100 legs: VIX vs its 50d mean · VIX9D/VIX ratio · HY OAS 20d change · RSP/SPY 20d relative return · SPY put/call volume vs a 1.25 baseline · SPY−TLT 20d return spread

Needs at least 3 of 6 legs to print; otherwise it shows a degraded state, not a number.

Source: Yahoo Finance, FRED, CBOE

Net liquidity

WALCL − WTREGEN − RRPONTSYD, in $T, weekly, forward-filled

The desk retains each component beside the result. Sample size is the aligned date count after all three series have a prior observation.

Source: FRED

Macro revision

Latest FRED value − ALFRED first-print value for the same observation period

Release and revision dates come from the observation real-time fields. Only matched first-print and current periods enter n.

Source: FRED and ALFRED

Recession panel

Four checks: Sahm rule ≥ 0.50 · 2s10s < 0 · claims 4-wk avg > 1.10x its 6-mo low · NFCI > 0. Score 0-1 clear, 2 watch, 3-4 elevated.

Source: FRED

Surprise index

Per release type, z = (actual − forecast) / sigma of that type’s historical surprises, clamped to ±3, sign-flipped for unemployment-class releases, summed with an 84-day half-life decay

Source: Finnhub economic calendar

Max pain

The strike that minimizes total in-the-money payout across calls and puts at the front expiry

Source: CBOE delayed quotes

GEX

Naive sum of gamma × OI × 100 × spot² × 1%, calls positive, puts negative

Dealer positioning is assumed, not observed; treat the sign convention as a convention.

Source: CBOE delayed quotes

Options IV surface

ATM term point = mean nearest-strike call and put IV per expiry; 25-delta skew = put IV − call IV

Every term point reports its contract sample. The strike surface retains front-expiry call and put OI, volume, and IV within 20% of spot.

Source: CBOE delayed options chains

Institutional COT

Category net = long − short; percentile uses up to 52 weekly observations

Financial contracts use Dealer, Asset Manager, and Leveraged Fund categories. Physical commodities use Producer/Merchant, Swap Dealer, and Managed Money.

Source: CFTC TFF Futures Only and Disaggregated Futures Only

Treasury auction demand

Indirect share = indirect accepted ÷ total accepted; dispersion = high yield − average median yield

True auction tail needs the pre-auction when-issued yield. FiscalData does not publish it, so tail is null.

Source: Treasury FiscalData auctions_query

Breadth oscillator

McClellan-style: EMA19 − EMA39 of daily net 20-day highs across the curated universe

Source: Yahoo Finance candles

Story clustering

Headlines merge at token-set Jaccard ≥ 0.45 with a ticker-overlap boost; score = ln(1+sources) × sourceWeight × (1+velocity) × exp(−age/9h)

Source: Aggregated news wire

Seasonality

Mean calendar-month return plus hit rate over the full history

History, not prediction. Nonpositive and non-finite closes are excluded before returns are calculated.

Source: Yahoo Finance

News velocity

GDELT 24-hour coverage volume for a query, refreshed on its 15-minute cycle

Source: GDELT

Insider flow

The last 8 Form 4 ownership documents; 90-day flow counts code P buys and code S sells; conviction = transaction shares ÷ holdings immediately before the trade

High conviction is at least 10% of pre-trade holdings, medium is 2% to under 10%, low is under 2%. Missing holdings remain null and do not enter n.

Source: SEC EDGAR

ETF primary-market flow

Change in issuer-reported shares outstanding × latest issuer NAV

Two distinct issuer dates are required. This estimates net share creation or redemption, not secondary-market trading volume.

Source: iShares fund pages

Congressional trades

STOCK Act periodic transaction reports, both chambers, transaction-level; amounts stay the statutory ranges; lag = disclosure date − trade date, flagged late past the 45-day window

Upstream return-since-trade columns are not shown: the desk does not attribute performance to disclosed ranges.

Source: House Clerk + Senate eFD, via the open Congress Trading Monitor dataset

GDPNow strip

The Atlanta Fed model estimate of current-quarter annualized real GDP growth; drift is the change against the prior model vintage

A nowcast reads the data as it lands; it is not a forecast survey. Updates around major releases, roughly 6-7 times a month.

Source: Atlanta Fed via FRED (GDPNOW)

Whale books (13F)

The two most recent 13F-HR information tables per tracked filer, manager rows aggregated by CUSIP and instrument; quarter-over-quarter diff into new stakes, exits, adds, and trims; top 60 lines by value

Long US equity snapshots only, up to 45 days stale by law; puts and calls labelled, never hidden; unmapped issuers keep their filed names.

Source: SEC EDGAR 13F-HR

Seasonality explorer

Per calendar month across the universe: mean month-over-month return and hit rate from up to 15y of monthly closes, baked at deploy; names under 8 observed years render muted

The Years column is the true sample size. A seasonal pattern describes the past; the desk does not forecast with it.

Source: Yahoo Finance monthly candles

Book drawdown

Today's lots valued over up to 3 years of shared sessions; the deepest peak-to-trough decline with its recovery date

Assumes current lots held throughout; deposits and exits are not modelled. Non-finite valuations are skipped.

Source: Yahoo Finance closes, computed in this browser

Monte Carlo fan

Gaussian geometric-brownian simulation seeded from the book's daily log returns; 500 paths, 1y horizon, percentile bands

History, not prediction; fat tails are underrepresented by construction.

Source: Computed in this browser

Session Clock

Eleven exchange cash sessions on a 24-hour UTC dial, offsets resolved per render from the IANA timezone database; arc ink scales with the day's move to a ±2% cap

Lunch halts draw as one continuous arc; exchange holidays are not modelled.

Source: Yahoo Finance quotes

Everything on the site runs on free or keyless public data. The full bill of materials:

Visible degradation beats silent failure. When a source is down or a leg is missing, the chart says so in place of the number; nothing is interpolated to look healthy.

No invented numbers. Where a language model writes prose, its output is validated against a typed contract before it ships, and it never produces a figure that did not come from a named source above.

Sources are chosen for durability, not convenience. Where a series has no dependable open feed, the panel that would use it degrades to a calm, related read rather than a broken widget, and nothing on the page is fabricated to fill a gap.

Everything is reproducible from the formulas on this page. If a number here cannot be rebuilt from the registry and the sources table, that is a bug.