Equarian Markets

A live macro dashboard for futures traders. Rates, credit, carry, the options market and the tape on one screen, read through a rule-based daily read and four quantitative models: a Risk Appetite Index, a growth × inflation regime, the absorption ratio of systemic risk, and the NY Fed recession probability — plus a walk-forward market-bias forecast scored out of sample since 2000 and an SPX option-flow panel built from the live option board.

History: First built in July 2026 and deployed on Netlify. Moved to GitHub in October 2026, so the commit history starts there. Moved to GitLab (eatentam/eqmarkets) on 17 Oct 2026 after the GitHub account was suspended. Moved to Cloudflare Pages on 23 Oct 2026 (see MIGRATION.md). Since 28 Oct 2026 GitLab mirrors every push back to GitHub, read-only.

At a glance

Live	https://markets.ethantam.dev
Hosting	Cloudflare Pages project eqmarkets, connected to this repository; every push to main deploys. Static Vite build, plus Pages Functions: the data proxies (functions/_middleware.js) and /api/explain
Data	Live public sources (Yahoo Finance, FRED, Treasury.gov, the BOJ, the Fed, the CFTC, CBOE) through same-origin proxies; the last good values are saved on the device
Run locally	npm install && npm run dev (see Develop)
Source	gitlab.com/eatentam/eqmarkets. Read-only mirror: github.com/Ethosan/eqmarkets, which GitLab updates on every push

Settings (names only; values live in the Pages project's environment variables): MISTRAL_API_KEY (optional, enables AI explanations), EXPLAIN_MODEL (optional).

Features

• 45 indicators in five groups. Rates: 3M, 2Y, 10Y, 2s10s, 3m10y, 10Y real yield, 10Y breakeven, Kim–Wright term premium, DXY. Risk: VIX, VIX/VIX3M term structure, VXN (Nasdaq vol), 1-day VIX (VIX1D), VVIX (vol of vol), CBOE SKEW, MOVE, HY OAS, Moody's Baa–10Y spread (daily since 1986), IG OAS, Chicago Fed NFCI, St. Louis Fed stress index, oil vol (OVX), gold vol (GVZ), initial jobless claims, Fed net liquidity (balance sheet − TGA − reverse repos), HYG, BTC. Carry: USD/JPY, EFFR, BOJ, CFTC yen positioning, carry-to-vol. Commodities: copper, copper/gold ratio, silver, natural gas, corn, wheat, soybeans, DBC (broad) and DBA (agriculture). Tape: ES, NQ, gold, crude.
• Today's read: fixed rules turn this month's cross-asset moves into plain-English observations (for example, oil and grains up while stocks fall reads as cost-push, checked against breakevens; NDX lagging SPX by more than a month's normal move while the S&P holds reads as leadership damage). Each line shows the exact numbers that triggered it. An optional Explain in plain English button has Mistral AI write a short note from the same numbers (see AI explanations below).
• Macro regime: a growth × inflation map (Goldilocks, Reflation, Stagflation, Slowdown) scored from 3- and 6-month market moves and read weekly. It changes only after a new quadrant holds for three weeks, shows when a change is building, and has a 13-week trail and a 52-week history strip.
• Method notes: every model card has a "How it's calculated" toggle listing its data, formula, weights and rules.
• Market bias: a bullish / neutral / bearish tilt for ES and NQ over the next day, week and month, with a confidence grade, each signal's pull, a walk-forward track record since 2000, and an "after a week like this" line: what actually followed comparable weeks (the current 5-day move's σ-bucket, matched to the volatility regime and momentum direction), so a bullish read after a washout shows its evidence (see Market bias model below).
• Flows and premia: an estimate of trend-follower (CTA) positioning in 8 markets, with the price levels where each signal flips and the forced flow from a ±2σ week; the variance risk premium (VIX minus ES realized vol); the stock–bond correlation; and a calendar of turn-of-month, FOMC, option expiry, holidays and month-end rebalancing pressure.
• SPX option flow: dealer gamma exposure estimated from the live CBOE option board — net gamma in dollars per 1% move, the gamma flip level, the gamma profile across index levels, call/put/gamma walls, 0DTE volume share and put/call ratios — plus the vol curve (VIX9D/VIX, VXN/VIX, VVIX, SKEW) with 1-year percentiles. Index-level options positioning, not single names.

Scope: Markets vs Oracular

Markets and Oracular are separate tools, and new features should stay on their side of the line.

Planned (1 Nov 2026): Markets becomes Oracular's Macro tab during December 2026 and January 2027, and markets.ethantam.dev then points there; until the tab ships this repo takes fixes only. See docs/PLAN.md.

Markets (this repo)	Oracular
Top-down and cross-asset: rates, credit, vol, FX, commodities, index futures	Bottom-up: one company or ticker at a time
Live market data	Live market data, one ticker at a time
Regimes, flows, premia and positioning (RAI, absorption, recession odds, CTA, VRP, stock–bond, SPX dealer gamma)	Investment memo, verdict badge and seven-factor scorecard
Calendar events that move index futures (turn of month, FOMC, expiry, rebalancing), tested as event studies	Per-ticker seasonality tables with t-statistics, lifecycle phase map
Testing whether a signal works (backtest lab, overfitting guards)	Options and derivatives analytics, research chat
Alt signals: fear-and-greed gauges, the Pentagon Pizza Index, the Stripper Index, distress searches, the Coinbase rank and the moon, each tested against the index (moved from Oracular, Oct 2026)

The alt-signal feeds come from Oracular's data worker through the same-origin /odw proxy; the worker turns away browsers from other sites, and a same-origin rewrite sends no Origin header.

So Markets doesn't show single stocks, fundamentals, single-name option chains, buy/hold verdicts or month-of-year seasonality tables — index-level option positioning (dealer gamma, index put/call flow, the vol curve) is macro and lives here — and Oracular doesn't do macro regimes or cross-asset flows.

Data sources

Every number on the page is a real reading. When a source is down, Treasury.gov stands in for the Treasury yields and the breakeven; anything else keeps its last real value, marked stale, or the device's saved copy. There are no typed-in fallback values.

Data	Source	Series	How it's reached
Futures, FX, indices, ETFs, BTC	Yahoo Finance chart API	ES=F NQ=F GC=F CL=F JPY=X DX-Y.NYB HYG BTC-USD ^TNX ^VIX ^VIX3M ^MOVE	/yf/* proxy
Treasuries, real yields, breakevens	FRED (H.15, Fed Board); Treasury.gov daily par and real yield curves when FRED is down (breakeven = 10Y nominal − 10Y real, as FRED defines it)	DGS3MO DGS2 DGS10 DFII10 T10YIE	/fred/* proxy; /ust/* proxy
Term premium	FRED (Kim–Wright, Fed Board)	THREEFYTP10	/fred/*
Credit spreads	FRED (ICE BofA; Moody's)	BAMLH0A0HYM2 BAMLC0A0CM BAA10Y	/fred/*
Labour, Fed balance sheet	FRED (DoL; Fed H.4.1; NY Fed)	ICSA (weekly), WALCL WTREGEN (weekly), RRPONTSYD	/fred/*
Financial conditions	FRED (Chicago Fed, St. Louis Fed)	NFCI STLFSI4 (weekly)	/fred/*
Fed funds	FRED (NY Fed)	DFF	/fred/*
BOJ policy rate	boj.or.jp homepage	scraped	/boj/* proxy; if the scrape fails, the last real reading stays up, marked stale
Yen speculative positioning	CFTC Legacy COT (Socrata 6dca-aqww)	non-commercial net, 3 years	/cftc/* proxy
Commodities	Yahoo Finance chart API	HG=F SI=F NG=F ZC=F ZW=F ZS=F DBC DBA	/yf/* proxy
FOMC meeting dates	federalreserve.gov FOMC calendar page	scheduled statement days	/fed/* proxy, built-in fallback in src/quant/calendar.js
SPX option board	CBOE delayed-quotes API	full chain: strike, expiry, IV, gamma, open interest, volume	/cboe/* proxy, cached 20 min ()

All of these are free and need no API key.

ES and NQ are roll-adjusted. Yahoo's ES=F and NQ=F switch to the next contract at each quarterly expiry, and the next contract trades at a premium, so the raw series jumps 1–3% on the third Friday and the session after. Checked against the cash indices over five years, those days were the only large gaps (25 days over 0.4% for ES, 31 for NQ). Every calculation uses a back-adjusted series instead (src/data/roll.js): from 8 sessions before each quarterly expiry to 1 after (NYSE calendar, so it also covers the CME roll date), each day takes the cash index's return (^GSPC, ^NDX) where the index has printed; every other day keeps the futures' own return, and the path is rebuilt back from today's price. Afterwards no daily gap to the index exceeds 0.4% and the typical one is 0.04%. Returns then roughly match what holding the futures earned: over five years 8.6% a year for ES instead of the raw series' 11.5% (the index did 11.3%; the difference is financing net of dividends). Using index returns on 10 sessions a quarter puts back a little carry on those days, so the adjusted series overstates futures returns by roughly 0.3–0.5% a year. GC and CL futures are not adjusted; the regime uses the rolled funds DBC and DBA instead.

Credit spreads start in October 2023. FRED only serves the ICE BofA high-yield and investment-grade spreads for the last three years, so the HY card's percentile, the Risk Appetite Index's credit input and anything else built on it can only say "wide for the last year or three", never "wide by history". That matters: on 16 Oct 2026 HY OAS was 3.23%, the 96th percentile of its past year after widening 58 bp in three weeks, yet that is well below the long-run norm of roughly 4.5–5% in ICE's full history (which peaked near 20% in 2008 and 11% in 2020). The Baa–10Y card (Moody's, daily since 1986, not licence-limited) is the long-run yardstick: its note gives its percentile since 1986, the HY card's note adds "long run tight / normal / wide" from it, and Today's read says whether a widening is a move from tight levels or real stress. In the long-history market-bias model and the lab, the Risk Appetite Index also lacks Bitcoin before 2014 and VIX3M before 2007, so for much of 2000–2023 it runs on 5 to 7 of its 8 inputs (it needs at least 4).

None of these send CORS headers, so the browser can't call them directly. On Cloudflare Pages, functions/_middleware.js proxies each path to its upstream host, which makes them same-origin (netlify.toml does the same on Netlify, and vite.config.js in development). Anywhere else the app falls back to public CORS proxies.

Methodology

Quantity	Definition
Returns	log returns for prices; first differences for rates (bp) and levels
Realized vol	sample std (n−1) of the last 21 steps × √252 (√365 BTC, √52 weekly)
EWMA vol	RiskMetrics σ²ₜ = λσ²ₜ₋₁ + (1−λ)r²ₜ₋₁, λ = 0.94, seeded with the first 20 obs
Percentile	mid-rank of today's value within the prior year
Robust z	(x − median) / (1.4826 · MAD) over the trailing year; classical z if MAD = 0
RAI trend score	ln(Pₜ / Pₜ₋₂₁) / (σ_EWMA · √21), signed, clipped ±3
RAI level score	robust z of the level over 252 days, signed, clipped ±3
RAI	mean of available scores (≥ 4 of 8); label on the 3-day mean, thresholds ±0.25σ / ±0.75σ
Absorption ratio	Σ top-k eigenvalues / Σ all, k = round(N/5), correlation matrix of 63-day returns (Jacobi eigen-solver)
ΔAR	(15-day mean AR − 252-day mean AR) / 252-day std AR
Recession probability	Φ(−0.5333 − 0.6330 · S̄), S̄ = 21-day mean of DGS10 − DGS3MO in pp
Carry-to-vol	(EFFR − BOJ) / annualized EWMA vol of USD/JPY
Vol-scaled move	Δₕ / (σ_EWMA at window start · √h): log change for prices, point change for rates; clipped ±3
Growth score	mean of ES, copper/gold, −HY OAS (≥ 2 of 3), each the average of its 63- and 126-day vol-scaled moves
Inflation score	mean of the 10Y breakeven, DBC, DBA (≥ 2 of 3), each the average of its 63- and 126-day vol-scaled moves
Regime	read weekly (last close of each week); switches only after the new quadrant holds 3 completed weeks in a row, with a ±0.15 dead band around each axis; dated to the first week of the run
Today's read	19 rules on 21-day vol-scaled moves, VIX/MOVE level z-scores, VXN/VIX and VIX9D/VIX ratio z-scores and NDX/SPX, RSP/SPY relative moves; strength = weakest condition; top 4 shown
Momentum (12−1M)	ln(Pₜ/Pₜ₋₂₁) − ln(Pₜ₋₂₅₂/Pₜ₋₂₅₂)'s window form: (ln Pₜ − ln Pₜ₋₂₁) ÷ (σ_EWMA at the window start × √231), clipped ±3
5-day reversal	(ln Pₜ − ln Pₜ₋₅) ÷ (σ_EWMA at the window start × √5), clipped ±3; its regression slope is restricted to be negative (Jegadeesh 1990)
Weeks like this

All of this lives in src/quant/ as pure functions, covered by tests/quant.test.js and tests/flows.test.js. The tests check them against closed-form cases: EWMA convergence to the true σ, eigenvalues of known matrices, an AR near 1 when one factor drives everything and near k/N for independent assets, the NY Fed probit formula, and bp units for rates.

Market bias model

One read per horizon (next 1, 5 and 21 sessions) for ES and NQ, built so that nothing is fitted to the period it is judged on (src/quant/composite.js, tests in tests/composite.test.js):

1. One forecast per signal. For each of trend followers, the variance risk premium, risk appetite, the absorption shift, the yield-curve recession odds, 12−1-month momentum (Jegadeesh & Titman 1993), and — at the 1- and 5-day horizons only — a 5-day reversal (Jegadeesh 1990; Lehmann 1990) plus turn of month, FOMC and pre-holiday timing, a univariate regression of the forward log return on the signal, estimated on an expanding window of observations whose outcome was known at the time. Goyal & Welch (2008) show big multivariate models fail out of sample; Rapach, Strauss & Zhou (2010) show simple ones combined hold up. Each regression weights recent years more (half-life three years), so relationships can drift (time-varying coefficients, Dangl & Halling 2012).
2. Calm and stressed markets. The market counts as stressed when the index's own EWMA volatility is above its median of the previous year. Each signal gets a slope for each state, shrunk toward its pooled slope by n / (n + 504) effective observations, so a state with little history stays close to the pooled estimate. Predictability is known to be concentrated in bad times (Rapach, Strauss & Zhou 2010; Henkel, Martin & Nardari 2011).
3. Sign restriction (Campbell & Thompson 2008). Each signal may only push the way its research says; a wrong-signed slope is set to zero — and that signal then abstains for the day instead of voting with the historical mean. Without the abstention a zeroed signal still looked "accurate" to the error-weighted combination (a mean forecast is hard to beat on squared error) and diluted the real ones; adding it moved ES's monthly R² from −0.34% to +1.06% on its own.
4. Combination by discounted mean squared forecast error (Stock & Watson 2004), half-life one year; equal shares until a signal has a year of out-of-sample record.
5. Earned weight against the long-run average. Final forecast = M + λ(F − M), with λ the forecast-encompassing weight (Harvey, Leybourne & Newbold 1998) from past out-of-sample errors only (half-life three years, matching the regressions), clipped to [0, 1]. When the models stop helping, λ falls and the read goes back to "no tilt".
6. Label from the tilt in the chance of a higher close versus the long-run base rate, with bands fixed in advance (1, 2, 3 points for day, week, month). Confidence needs past calls of that kind, in the same market state as today when there are 50 or more, to have gone their way more often than on the days the call wasn't made (5 points for high, 2 for medium), most of the weight to agree, and Clark–West p below 0.10 for high.
7. Scored out of sample: R² against the prevailing historical mean, Clark & West (2007) with Newey–West errors, hit rates, calibration, and the certainty-equivalent gain of a mean–variance investor (γ = 3).

It runs in a Web Worker on full daily history from 2000 and is cached on the device for six hours. The scorecard is also split by market state, and the card says in plain words whether the model beats a plain average overall, in calm markets and in stressed ones.

Results and variants tried (out-of-sample R² for day / week / month, full history, every forecastable day counted, re-run 19 Oct 2026 with node scripts/bias-research.mjs; earlier runs differ slightly because the data window moved). Every variant run this round is disclosed:

Variant	ES	NQ
v7 as shipped (5 signals, no abstention)	+0.06 / +0.07 / −0.34%	+0.22 / +1.06 / +3.23%
v7 + abstention (with the all-abstain fallback)	+0.11 / +0.02 / +0.68%	+0.29 / +1.12 / +1.24%
v8 live: v7 + abstention + 12−1 momentum + 5-day reversal (day/week only)	+0.11 / +0.12 / +0.66%	+0.23 / +0.98 / +2.81%
v8 with all 12 candidate signals (see below)	+0.04 / −0.13 / −0.46%	+0.21 / +1.09 / +2.92%

The headline fix this round is the abstention rule: a signal whose sign-restricted slope comes out wrong-sided used to keep voting with its historical mean, and the error-weighted combination rewarded that fake accuracy with a large share. Zeroed signals now abstain. On its own that took ES monthly from −0.34% to +0.68%; adding momentum and reversal gives the live v8 — ES beats its long-run average at all three horizons for the first time (monthly +0.66%, Clark–West p 0.026), and NQ stays significant everywhere (p ≤ 0.006 daily, < 0.001 weekly and monthly). Against the abstention-only variant, momentum and reversal add the most exactly where the complaint was: NQ monthly +1.24% → +2.81%, and daily NQ bearish calls right 58% of the time versus 44% on other days. The honest costs: NQ weekly and monthly R² sit a touch under the no-abstention v7 over the full history (0.98% vs 1.06%, 2.81% vs 3.23%), because the abstaining days now count.

The all-abstain fallback. Abstention first shipped with a hole the size of 2023–26: over the recent window every signal's estimated slope points the wrong way at once (the meltup inverted the usual momentum, trend and risk-appetite relationships), so at the monthly horizon nothing voted and the model went silent for ~700 days — the card showed "Not enough history", which was wrong. Now, when every signal abstains on a day, the combination falls back to the abstaining signals and is set to the long-run average exactly — a neutral read, flagged in the card as "no working signal: recent moves contradict every signal's usual direction". That state is live right now on NQ week/month and ES month; ES week is genuinely bullish on the variance premium and the reversal signal.

21 Oct 2026 round: VIX term structure and credit momentum, rejected. Two more candidates with long free history were tested: VIX/VIX3M (Fassas & Hourvouliades 2019; backwardation has tended to precede rebounds) and credit momentum, the 63-day vol-scaled move of HYG against IEF in total return (Gilchrist & Zakrajšek 2012). The acceptance rule was fixed before the run: ship only if full-sample out-of-sample R² beats v8 at 2 of 3 horizons for both ES and NQ, without lowering the average post-2019 R². Neither passed (day / week / month, scripts/cache/run-2026-10-21.txt):

Variant	ES	NQ
v8 live	+0.11 / +0.12 / +0.66%	+0.23 / +0.98 / +2.81%
v8 + VIX term structure	−0.04 / −0.11 / −1.23%	+0.17 / +0.88 / +1.50%
v8 + credit momentum	+0.09 / +0.02 / +0.42%	+0.22 / +0.94 / +2.75%
v8 + both	−0.05 / −0.18 / −1.25%	+0.16 / +0.83 / +1.32%

Credit momentum did improve the post-2019 cut (NQ monthly −0.66% → −0.23%), but it lowered the full record everywhere, so it stays out. Both remain in the harness so the numbers are reproducible.

The last three years, on the card. The full record since 2002 hides how weak the model has been lately: since 2019 no horizon beats the plain average for ES or NQ. The card now shows a Last 3 years row alongside the full record, and a lean whose horizon hasn't beaten the plain average over those three years is graded Low confidence, whatever its long record says (src/model/bias.js). That rule was set before looking at which reads it would change.

Candidate signals tried and rejected (all estimated on the same walk-forward, all reproducible in the harness): NDX/SPX relative strength ("growth leadership"), RSP/SPY equal-weight breadth, VXN/VIX, VIX9D/VIX and VVIX level z-scores. NDX/SPX is a relative signal — it predicts Nasdaq outperforming the S&P, not the S&P itself, which is what the ES numbers showed. The vol-ratio signals helped NQ's 2019–2026 monthly cut but wrecked ES weekly bearish calls, so none shipped inside the model; VXN, VVIX and SKEW are still on the board as indicators and in the option-flow panel, and the divergence rules in Today's read quote them directly.

Scored on 2019–2026 only, nothing is significant for ES or NQ. Directionally, the skill sits in bearish calls: on full history the index fell short of its average after 43–58% of bearish calls versus 30–45% of other days, while bullish calls did no better than other days. The lab's Market bias model card shows the full scorecard, the Goyal–Welch cumulative error plot, weights over time and calibration.

NQ stays significant at every horizon (Clark–West p ≤ 0.009); ES at none, though monthly now clears p 0.05. Scored on 2019–2026 only, nothing is significant for ES or NQ. Directionally, the skill sits in bearish calls: on full history the index fell short of its average after 44–58% of bearish calls versus 30–44% of other days, while bullish calls did no better than other days. The lab's Market bias model card shows the full scorecard, the Goyal–Welch cumulative error plot, weights over time and calibration.

References for the bias model's new signals: Jegadeesh & Titman (1993), Returns to Buying Winners and Selling Losers, JF. Jegadeesh (1990), Evidence of Predictable Behavior of Security Returns, JF. Lehmann (1990), Fads, Martingales, and Market Efficiency, QJE. Ang, Hodrick, Xing & Zhang (2006), The Cross-Section of Volatility and Expected Returns, JF. Baltussen, van Bekkum & van der Grient (2018), Unknown Unknowns: Uncertainty About Uncertainty and Equity Returns, JFQA (all rejected for the live model; see the results table).

References for the bias model: Dangl & Halling (2012), Predictive Regressions with Time-Varying Coefficients, JFE. Henkel, Martin & Nardari (2011), Time-Varying Short-Horizon Predictability, JFE. Goyal & Welch (2008), A Comprehensive Look at the Empirical Performance of Equity Premium Prediction, RFS. Campbell & Thompson (2008), Predicting Excess Stock Returns Out of Sample, RFS. Rapach, Strauss & Zhou (2010), Out-of-Sample Equity Premium Prediction: Combination Forecasts and Links to the Real Economy, RFS. Stock & Watson (2004), Combination Forecasts of Output Growth in a Seven-Country Data Set, JoF. Harvey, Leybourne & Newbold (1998), Tests for Forecast Encompassing, JBES. Clark & West (2007), Approximately Normal Tests for Equal Predictive Accuracy in Nested Models, JoE. DeMiguel, Garlappi & Uppal (2009), Optimal Versus Naive Diversification, RFS.

References for flows and premia: Moskowitz, Ooi & Pedersen (2012), Time Series Momentum, JFE. Hurst, Ooi & Pedersen (2017), A Century of Evidence on Trend-Following Investing, JPM. Bollerslev, Tauchen & Zhou (2009), Expected Stock Returns and Variance Risk Premia, RFS. Carr & Wu (2009), Variance Risk Premiums, RFS. Campbell, Pflueger & Viceira (2020), Macroeconomic Drivers of Bond and Equity Risks, JPE. Lakonishok & Smidt (1988), Are Seasonal Anomalies Real?, RFS. McConnell & Xu (2008), Equity Returns at the Turn of the Month, FAJ. Lucca & Moench (2015), The Pre-FOMC Announcement Drift, JF. Ariel (1990), High Stock Returns before Holidays, JF.

References: Ilmanen, Maloney & Ross (2014), Exploring Macroeconomic Sensitivities: How Investments Respond to Different Economic Environments, JPM. Bridgewater Associates (2012), The All Weather Story. Moskowitz, Ooi & Pedersen (2012), Time Series Momentum, JFE. Erb & Harvey (2006), The Strategic and Tactical Value of Commodity Futures, FAJ. Gorton & Rouwenhorst (2006), Facts and Fantasies about Commodity Futures, FAJ. Kumar & Persaud (2002), Pure Contagion and Investors' Shifting Risk Appetite, IMF. Kritzman, Li, Page & Rigobon (2011), Principal Components as a Measure of Systemic Risk, JPM. Estrella & Mishkin (1996), The Yield Curve as a Predictor of U.S. Recessions, NY Fed. Estrella & Trubin (2006), The Yield Curve as a Leading Indicator, NY Fed. J.P. Morgan/Reuters (1996), RiskMetrics Technical Document. Gilchrist & Zakrajšek (2012), Credit Spreads and Business Cycle Fluctuations, AER. Kim & Wright (2005), An Arbitrage-Free Three-Factor Term Structure Model, Fed Board.

Caveat: the NY Fed fits its probit to monthly averages and the bond-equivalent 3M bill, so this daily version can differ from the published table by a point or two. Check NYFED_PROBIT in src/quant/models.js against the NY Fed's current release.

Today's read changes day to day now. The monthly rules can hold for weeks, so the same sentences used to lead every morning. Each pattern is tagged with how long it has held (New today, This week, 2+ weeks, from the same rules run on the data as it stood 1, 5 and 10 sessions back), patterns on for two weeks or more give up some rank to fresh news, and three daily rules lead when something happened: moves since the last close of more than twice a normal day, VIX jumps or drops (read together with VVIX), and new 52-week highs and lows. The daily rules use oil and gold ETFs (USO, GLD) and leave out the other unadjusted front-month futures: on 21 Oct 2026 CL=F "fell" 6.3% because the November contract expired, while USO and Brent moved under 1%.

Why the regime uses DBC and DBA instead of the grain and energy futures: Yahoo's continuous front-month contracts (ZC=F and so on) jump on each roll date, by the gap between the expiring and the next contract. Over a 3-month window that can be several percent of fake "move". The funds hold and roll the futures themselves, so their prices don't jump. The commodity cards still show the futures, and the guide warns that big monthly moves there can include roll.

Backtest lab (admin)

A hidden Lab tab for testing the models as trading signals. Open the site with ?admin=1 to turn it on in that browser (?admin=0 turns it off). It only hides UI: there is no secret behind it, and the lab code loads as its own chunk the first time the tab opens.

• Data: 2Y, 5Y, 10Y or full daily history for the model inputs and the instrument (NQ, ES, NDX, QQQ and SPY total return, gold, Bitcoin).
• Point-in-time models: RAI (raw and 3-day mean), absorption ratio and ΔAR, the recession probability, the variance risk premium, the stock–bond correlation and the trend-follower position are rebuilt for every day using only data up to that day. FRED series are treated as known one business day after their date, when they are published.
• Overlay: the price chart is shaded by RAI regime, ΔAR fragility, recession odds, the variance premium, the stock–bond correlation, trend-follower positioning, the turn of the month or any strategy's position, with the model in a synced pane below.
• Strategies: buy and hold, RAI regime, RAI scaled, absorption ΔAR (Kritzman et al.), recession filter, RAI + ΔAR, variance premium, CTA trend, turn of month, a 200-day trend baseline, and a custom JavaScript rule. Options: execution on the signal's close or the next one, cost in bp per side, and an EWMA volatility target.
• Results: growth of $1 (log), drawdowns, and CAGR, vol, Sharpe, Sortino, max drawdown, Calmar, exposure, trades per year, hit rate and beta to buy and hold, plus a CSV export. Every strategy is measured over the same window, starting when all selected strategies have a signal.
• Is it real? Guards against overfitting, the usual way backtests mislead:
  • an in-sample / out-of-sample split (last 30% or 50%), marked on every chart;
  • a 90% interval on Sharpe from 600 stationary block-bootstrap resamples (Politis & Romano 1994, 10-day blocks);
  • the deflated Sharpe ratio (Bailey & López de Prado 2014): the probability the true Sharpe is above zero, after fat tails and after counting every strategy variant tried in the session, including sweep cells;
  • a parameter sweep: any rule that reads ctx.p.a / ctx.p.b runs over a grid, shown as a heatmap scored on the in-sample period, with the best cell's out-of-sample Sharpe. A real edge is a broad plateau, not one bright cell.
• Calendar effects: an event study on the instrument's whole history: turn of month, FOMC day and the day before, option-expiry week, quad-witching week, expiry day, pre-holiday, and month-end after stocks led or lagged Treasuries by 2+ points. Each row has the mean event-day and other-day return in bp, Welch's t, a randomization p-value (2,000 same-size random day sets), the hit rate and the share of total return.
• Console: eqlab.test(ctx => ctx.raiS > 0 ? 1 : 0) prints a rule against buy and hold. eqlab.frame has the aligned data and eqlab.events the calendar study.

A rule is the body of function (ctx), called once per bar. It returns a position (1 long, 0 flat, −1 short, clipped to ±3) or null to hold. ctx exposes only the current and past bars (ctx.price, ctx.raiS, ctx.dAR, ctx.rec, ctx.vrp, ctx.sbCorr, ctx.trend, ctx.sma(n), ctx.get(field, k)…), so a rule can't peek ahead. The one exception is ctx.cal(flag, ahead): calendar dates are known in advance, so a rule may ask whether the next bar is, say, a turn-of-month day. The engine is src/quant/backtest.js, tested in tests/backtest.test.js (lag, costs, holding on null, no look-ahead, vol targeting, ΔAR path).

Caveats: futures use Yahoo's continuous front-month series, so roll gaps count as returns. Flat earns nothing, and Sharpe uses a zero risk-free rate. A backtest over a few years of one market is a sanity check, not evidence of an edge.

AI explanations (optional)

The Explain in plain English button calls functions/api/explain.mjs (a Cloudflare Pages Function; netlify/functions/explain.mjs is the Netlify version), which asks Mistral (mistral-medium-latest, Mistral Medium 3.5) for a two-paragraph note.

• Turn it on: Cloudflare → Workers & Pages → eqmarkets → Settings → Variables and secrets → add MISTRAL_API_KEY as a secret, then redeploy. Create the key at console.mistral.ai → API Keys. Without it the button says it isn't set up.
• Optional: EXPLAIN_MODEL to use a different Mistral model, for example mistral-small-latest (cheaper and faster) or mistral-large-latest.
• What it sends: only numbers. The browser posts the scores behind Today's read. The function checks them against a fixed schema (known ids, bounded finite numbers, ISO dates), rebuilds every sentence itself in src/ai/explain.js, and only then calls the API, so nothing typed by a visitor can reach the prompt.
• Cost control: a per-IP rate limit of 6 a minute (kept in memory, so it resets when the function restarts), a 30-minute cache per identical data block, a 25-second timeout and at most 700 output tokens. A call costs about half a US cent on Medium 3.5 ($1.5/M input, $7.5/M output). Set a monthly spend limit in the Mistral console (Admin → Limits) as the backstop, because the endpoint is public.
• Errors: a rejected key, rate limit, timeout or empty answer each shows a short notice instead of partial text.

Public API

Two read-only endpoints expose the dashboard's own models, computed with the same code the site runs (functions/api/_macro.mjs), so the API and the dashboard always agree. Free for now; no key, no account.

• GET /api/regime → the growth × inflation regime: { quadrant, label, confidence, growth, inflation, readingQuadrant, provisional, weeksIn, since, pending, switchesYear, trail: [{ date, growth, inflation }], asOf, updated, stale }. trail is the 13 finished weekly readings the regime map draws; growth and inflation are this week's reading.
• GET /api/risk → the Risk Appetite Index: { value, latest, label, inputs, total, contributions: [{ id, label, score }], history: [{ date, value }], asOf, updated, stale }. history is the daily index the dashboard charts (about a year); value is the mean of its last three days.

stale: true means a refresh failed and the last good read is being served. asOf is the latest market session the numbers cover.

Call them from anywhere — no key, no signup, CORS-open:

curl https://markets.ethantam.dev/api/regime
curl https://markets.ethantam.dev/api/risk

const [regime, risk] = await Promise.all([
  fetch("https://markets.ethantam.dev/api/regime").then(r => r.json()),
  fetch("https://markets.ethantam.dev/api/risk").then(r => r.json()),
]);

• Freshness: one snapshot is computed per five minutes per function instance; responses are edge-cached for 60 seconds. Fields are additive — new ones may appear without notice.
• Limits: 30 requests a minute per IP (kept in memory, so it resets when the function restarts). CORS-open; Cache-Control: public, max-age=60.
• Abuse: the endpoints take no input, hold no secrets and change no state. A flood is absorbed by the edge cache, the five-minute snapshot cache and a one-minute failure backoff, so the upstream data feeds and the rest of the site are not put at risk. If the API ever costs real money to run, it gets a key or a tighter limit before anything else changes.
• Errors: 429 { error: "busy" } when rate-limited, 503 { error: "unavailable" } when no read exists yet. An error body is the only thing an error ever says.

Project structure

index.html              page shell and static markup
src/
  main.js               boot, refresh loop
  pwa.js                service worker, install bar, update notice
  config.js             indicator list, refresh interval
  state.js              latest series per indicator
  data/
    transport.js        proxy fallbacks and timeouts
    sources.js          Yahoo, FRED, BOJ and CFTC fetchers
    fomc.js             FOMC dates from federalreserve.gov, cached a week
    roll.js             quarterly roll adjustment for ES and NQ
    snapshot.js         last good data saved on the device for instant and offline launch
  quant/
    stats.js            statistics: EWMA, robust z, percentiles, eigenvalues, alignment
    models.js           RAI, absorption ratio, recession probit, per-series stats
    regime.js           vol-scaled moves, growth × inflation regime
    narrative.js        rules behind Today's read
    flows.js            variance risk premium, stock–bond correlation, CTA trend positioning
    optionflow.js       SPX option board: dealer gamma, profile, flip level, walls, put/call, 0DTE
    conditional.js      "after a week like this": bucketed conditional forward-return history
    calendar.js         NYSE holidays, FOMC dates, turn of month, option expiry, event study
    backtest.js         backtest engine: simulate, metrics, vol target, rules, presets, bootstrap, deflated Sharpe
    composite.js        market bias: walk-forward forecast combination and its out-of-sample scorecard
  lab/                  admin backtest lab (separate chunk)
    lab.js              controls, strategies, results, custom rule editor
    plot.js             multi-series chart with shading and synced crosshairs
    insights.js         calendar event study and parameter-sweep heatmap
  ai/
    explain.js          payload validation and prompt building (shared with the function)
  model/
    derived.js          curve spreads, VIX term structure, carry-to-vol
    models.js           builds the models from fetched state
    facts.js            the numbers behind Today's read and the AI note
    history.js          long history, the point-in-time model frame and the momentum, reversal,
                       leadership and options-vol signal paths (lab and market bias)
    bias.js             market-bias build, cache and worker hand-off
    bias.worker.js      runs the market-bias build off the main thread
  ui/
    board.js            indicator cards
    detail.js           detail sheet (dialog)
    models-panel.js     models section rendering
    flows-panel.js      CTA positioning, variance premium, stock–bond correlation, calendar
    bias-panel.js       market bias card
    optionflow-panel.js SPX option-flow card (chain load, gamma profile chart, vol-curve gauges)
    methods.js          "How it's calculated" notes on each model card
    read-panel.js       Today's read and the AI button
    regime-panel.js     regime map
    chrome.js           clocks, session chips, tabs, progress bar
    notes.js            one-line reads per indicator
    chart.js            interactive chart: axes, crosshair, tooltip, keyboard
    tour.js             first-run spotlight tour with live values
    guide.js            "how to read this" guide sheet
    sheet.js            swipe-to-dismiss for bottom sheets
    format.js           number formatting and sparklines
  styles/               base tokens, layout, components
public/                 manifest, icons, service worker, _headers, _routes.json (copied as-is)
functions/              Cloudflare Pages Functions
  _middleware.js        same-origin data proxies (/yf, /fred, /ust, /boj, /fed, /cftc, /cboe, /odw)
  api/explain.mjs       POST /api/explain (Mistral)
netlify/functions/
  explain.mjs           the same endpoint for Netlify
scripts/
  bias-research.mjs     research harness: downloads history, runs the walk-forward offline, prints and saves the results table in the README
netlify.toml            the Netlify build settings, data proxies and headers (kept so the site can go back)

Develop

Needs Node 20 or newer.

npm install
npm run dev          # http://localhost:5173

vite.config.js mirrors the proxies in functions/_middleware.js, so npm run dev and npm run preview fetch data same-origin like the deployed site. Open http://localhost:5173/?admin=1&tab=lab for the lab.

Script	What it does
npm run dev	Vite dev server with hot reload
npm run build	Production build into dist/
npm run preview	Serve the production build
node scripts/bias-research.mjs	Re-run the walk-forward research offline (fetches full history, caches in scripts/cache/)
npm run lint	ESLint
npm test	Vitest unit tests for the quant library and backtest engine
npm run audit:design	impeccable design anti-pattern scan (npx impeccable if not installed)

Deploy

The live site is the Cloudflare Pages project eqmarkets, connected to this repository: build command npm run build, output directory dist. Every push to main deploys. functions/ is picked up automatically, and public/_routes.json runs Functions only on the proxy paths and /api/*. Setup steps are in MIGRATION.md.

To go back to Netlify, import the repository there: it reads netlify.toml (build command npm run build, publish directory dist).

.gitlab-ci.yml runs lint, tests and the build on every merge request and every push to main.

After a BOJ rate decision, if the live scrape ever fails, update BOJ_RATE in src/config.js.

Notes

• Built for context, not trading advice.
• The service worker never caches market data; it only keeps the app shell for offline launch.

Security

Found a security problem? See markets.ethantam.dev/.well-known/security.txt, or email [email protected].