Options Desk — Final Ruling: Data Sources & Advanced Metrics

Panel: Opus, Sonnet, Grok 3, Gemini Pro 3.1 Judge: Opus (via synthesis) Date: 2026-03-25 Grade: A (strong consensus on architecture and core metrics, minor dissent on OCC and Vanna/Charm)

Consensus Points

1. Polygon Snapshot as Foundation (UNANIMOUS — all 4)

Use GET /v3/snapshot/options/{underlying} — one call returns full chain with pre-calculated Greeks (delta, gamma, theta, vega, IV). No need to calculate Greeks from raw prices. Fix the broken options-collector.ts first (imports non-existent options-filter.ts). Replace with direct Polygon snapshot + DB insert.

2. 3-Tier Ticker Architecture (UNANIMOUS — all 4)

All panelists converged on the same tiering:

• Tier 1 (every run): SPY, QQQ — index proxies for all macro Kalshi markets
• Tier 2 (3-4x daily): NVDA, TSLA, AAPL, META — highest vol, most Kalshi single-name activity
• Tier 3 (1-2x daily): MSFT, GOOGL, AMZN, JPM, XLE, IWM — complete picture

Add TLT, GLD, HYG (Opus) — rate vol, inflation hedge, credit stress proxies. Takes list from 13 to 16.

3. CME FedWatch Probabilities (STRONG — Opus, Sonnet, Gemini)

The sharpest anchor for Kalshi Fed rate markets. CME publishes institutional-grade probability from Fed Funds futures. Compare directly to Kalshi price — CME is institutional money, Kalshi is retail. The gap is the edge.

• JSON endpoint behind CME page, or derive from ZQ futures via Polygon
• Sonnet notes: if DIY calc diverges from CME's published number by >1-2%, the difference is noise you introduced, not edge

4. GEX with Flip Point (STRONG — Opus, Sonnet, Gemini)

All panelists agree: equity GEX is the most actionable mechanical signal.

• GEX(K) = OI_calls(K) * Gamma(K) * 100 * S - OI_puts(K) * Gamma(K) * 100 * S
• Flip point: strike where cumulative GEX crosses zero
• Above flip → dealers long gamma → dampen moves → d2 overestimates big moves
• Below flip → dealers short gamma → amplify moves → d2 underestimates risk
• Polygon returns gamma per contract — no manual calculation needed

5. Skew-Adjusted Binary Probability (STRONG — Opus, Grok)

Standard d2 uses ATM IV, but the probability of S > K depends on IV at strike K:

• P_skew_adjusted = N(d2) using IV(K), not IV(ATM)
• skew_premium = P_skew_adjusted - P_atm_iv
• Most binary calculators use ATM IV. Using strike-specific IV is more accurate, and the difference IS a tradeable signal.

6. VRP (Volatility Risk Premium) as Binary Edge (STRONG — Opus, Sonnet, Gemini)

• VRP = IV_30d - RV_30d
• VRP_zscore = (VRP - mean(VRP_history)) / std(VRP_history)
• When VRP is elevated (z > 1): d2 probabilities are MORE biased than usual. "Will SPY stay above X" markets are UNDERPRICED.
• When VRP compressed (z < -1): options market unusually accurate, less edge.
• Opus moonshot: adjust effective_IV = IV - alpha * VRP (alpha ~0.5, calibrate empirically)

7. Drop CFTC Options COT (UNANIMOUS — all 4)

Futures COT already flows via existing scraper. Options-only supplement (COTR) adds <8% marginal info. Correlation with futures-only COT is >0.92. Not worth separate download.

8. Deprioritize Vanna/Charm (STRONG — Sonnet, Grok, Gemini; Opus includes but flags complexity)

Second-order Greeks require intraday snapshots to be useful. Daily snapshots give stale values. Defer until intraday polling exists. Master GEX first.

Dissent Rulings

OCC Daily CSVs — INCLUDE at reduced scope (Opus+Sonnet say yes, Grok+Gemini say skip)

Opus and Sonnet argue OCC gives definitive clearinghouse OI that differs from real-time by 5-15%. The OI delta (day-over-day change) reveals whether volume opened or closed positions — this changes the signal completely. Grok and Gemini argue Polygon is "99% as good."

RULING: Include OCC, but only for SPY and QQQ (not all 16 tickers). Two CSVs per day. The open/close distinction is worth having for the two tickers that drive all macro binary markets.

Earnings Calendar — INCLUDE (Sonnet+Gemini say yes, others silent)

Need to know when earnings are to calculate event vol isolation. Without a calendar, you can't identify earnings-adjacent expirations.

• EarningsWhispers API (free, unofficial) or FMP economic calendar
• Daily collection

RULING: Include. One daily scrape. Required for event vol stripping.

News Sentiment — DEFER (Gemini only)

Gemini suggests NewsAPI for "news pressure" scoring. Valid concept but adds API cost ($449/mo at scale) and complexity. We already get sentiment signals from VIX/VVIX dynamics.

RULING: Defer. Not worth the cost when VIX suite already captures fear/uncertainty.

ETF Flows — DEFER (Gemini only)

Gemini suggests scraping issuer sites for shares outstanding changes. Valid for institutional flow detection but low ROI relative to GEX which captures the same hedging pressure.

RULING: Defer. GEX is the better signal for the same underlying flow.

Build List

Phase 1: Core Data Collection (1 new file, 1 fix)

Fix: src/pipeline/data/options-collector.ts — Replace broken imports with direct Polygon snapshot:

• GET /v3/snapshot/options/{underlying}?limit=250&apiKey=KEY
• Store to options_chains table: ticker, timestamp, strike, expiry, call_put, iv, oi, volume, delta, gamma, theta, vega, last_price, underlying_price
• 16 tickers: SPY, QQQ, AAPL, MSFT, GOOGL, AMZN, NVDA, META, TSLA, JPM, XLE, IWM, TLT, GLD, HYG, VIX
• Self-initializing table pattern

New: src/pipeline/data/options-occ.ts — OCC daily settlement CSVs:

• SPY and QQQ only
• Store to options_occ_oi table: ticker, date, strike, expiry, call_put, oi, volume, oi_change
• Calculate day-over-day OI delta (open/close indicator)

Phase 2: Calculated Metrics (1 new file)

New: src/pipeline/data/options-metrics.ts — Calculated from chain data:

1. Skew-Adjusted Binary Probability — N(d2) using strike-specific IV, not ATM
2. VRP Z-Score — IV_30d - RV_30d, normalized over 180-day history
3. GEX + Flip Point — Aggregate dealer gamma exposure, find zero-crossing strike
4. Event Vol Isolation — Forward variance stripping around earnings/FOMC dates
5. Implied Probability Term Structure — d2 probability at same strike across all expirations, slope detection
6. Put-Call OI Ratio by Expiration — Per-expiry PC ratio (not just aggregate)
7. Cross-Platform Mispricing Index — |options_prob - kalshi_price| / max(options_prob, kalshi_price) * 100
8. Vol-of-Vol Regime — VVIX/VIX ratio z-score, flags unreliable d2 periods

Phase 3: External Anchors (added to existing files)

CME FedWatch — Add to economics.ts or new fedwatch.ts:

• Scrape CME JSON endpoint for published rate probabilities
• Compare to Kalshi Fed rate markets
• Weekday collection, 4x daily

Earnings Calendar — Add to new earnings-calendar.ts:

• EarningsWhispers or FMP API
• Daily collection, 14-day lookahead

Phase 4: Cron Wiring

Phase 5: Moonshot (Opus) — Breeden-Litzenberger Probability Surface

Build AFTER Phase 1-4 are stable. Uses call spread approximation for model-free probability:

prob(S_T > K) ≈ (C(K) - C(K+dK)) / dK * (1+r)^T

• Requires dense chain (SPY has $1 strike spacing — dense enough)
• No Black-Scholes assumptions — uses actual market prices
• The basis (Kalshi price - surface probability) is the signal
• Both Opus and Sonnet highlighted this as the highest-alpha metric

Implementation Notes

• Polygon Starter plan: 100 req/min (NOT 5/min — Grok was wrong about rate limits)
• 16 tickers × 1 snapshot call = 16 API calls per run = 10 seconds at 100/min
• All new tables use self-initializing pattern with ensureTable() guard
• getDb() from ../../db.js — better-sqlite3, SYNC, no await
• safeFetch() with timeout and User-Agent
• pino logger per module
• ESM with .js extensions on all imports
• updateFreshness() after every collection run