ForecastEx UHLGA: settlement grounded + KLGA oracle fit (#2217)

TL;DR

1. PROVEN — ForecastEx's only live NYC daily-high product is the U-series UHLGA,

settling on Weather Underground's daily high for the LaGuardia-coded region. The probe's station inference was right; its "confirm in writing" step is now done. The DH-series (NWS-CLI-settled, NYC = Central Park) that the IBKR Campus article describes is no longer listed (0 DH products in the venue's product summary).

1. MEASURED — venue settlement ≡ round(max METAR tmpf) for LGA: 14/14 settled

days match (Jun 2026). The NWS CLI daily max disagrees 6/13 days, by up to 4°F (2026-06-12: WUvs CLI 98). The CLI is the wrong settled leg for this venue.

1. NEW CAPABILITY — the venue publishes its full board publicly:

https://www.forecastex.com/api/download?type={prices|pairs|summary}&date=YYYYMMDD (EOD prices including settlement values, no auth, dates ≥ 2026-02-03). This unblocks the Observe leg without the IBKR EC entitlement that blocked #2216.

1. MEASURED — the KLGA station fit (the "re-fit, not fee swap" the probe demanded):

n=1825 pairs, OOS RPS 0.0356 (fitted) vs 0.0508 (defaults), 30% gain, promoted to data/kalshi/weather-oracle-params-klga.json.

1. HONEST NEGATIVE — ceilingSupport: []: five summers of KLGA data contain too few

99–104°F forecast days to fit a ≥100 ceiling. The Kalshi ≥100-fade edge does not transfer as fitted. Any UHLGA edge must be found forward, against live prices.

1. Settlement: primary evidence

• **CFTC 40.2(a) product filing (2026-01-23,

ptc01232637912.pdf), Appendix A "U Contract Terms and Conditions":**

• Event question: *"Will the [highest/lowest/average] temperature in [region]

[exceed/be below] [##] F on [date]?"* → exceed semantics.

• Product code: U[H/L/A][three letter region code] → UHLGA = U·High·LGA.
• "Source Agency: Weather Underground"; resolves on WU's *"final daily temperature

values for the applicable station and date"* (the "High Temp / Actual" figure), once the first next-day observation publishes. Tick $0.01; binary $1.00/$0.00 settlement.

• Venue product summary (public CSV, 2026-06-16):products;UH* +UL*

temperature products; zero DH*. NYC appears only as UHLGA/ULLGA. The IBKR-Campus "NYC = Central Park" description applies to the DH-series generation and is stale for what actually trades.

Consequence (unchanged from the probe, now proven): Kalshi KXHIGHNY (KNYC/Central Park) vs ForecastEx UHLGA (LGA) is a station-basis pair, not an arb. LGA runs systematically warmer; cross-venue-monitor.js's caveat stands.

2. Settlement semantics, measured against the venue's own numbers

From the public prices CSVs (settlement column), per contract date: maxYes = highest threshold settling YES=1, minNo = lowest settling YES=0. On clean flips (minNo == maxYes+1) the settled high is pinned exactly (strict-exceed ⇒ high = minNo).

Jun 3–16 2026, KLGA (14 clean days):

The CLI↔WU gap is physical: the CLI max comes from the ASOS continuous (5-min) record and catches spikes between reports; WU's "Actual" high tracks the METAR feed. Over summers 2021-2025 the gap averages −0.89°F (ASOS-max − CLI, n=463 shared days).

3. KLGA oracle fit (station-generic pipeline)

scripts/validate-weather-oracle-fit.js now takes --mos-station --asos-station --asos-network --settled cli|asos --normals-cli-year --out (defaults reproduce the KNYC run byte-identically; all pre-existing tests green). Normals come from the station's own NWS CLI high_normal (NCEI 1991-2020;days fetched, KLGA summer mean 84.2°F).

Run: --mos-station KLGA --asos-station LGA --settled asos --normals-cli-year--train 2021,2022,2023 --test 2024,2025 --promote --out data/kalshi/weather-oracle-params-klga.json

Ceiling: ceilingSupport: []. No 99–104°F forecast bin reached the 12-sample support floor in five summers. lib/forecastex-weather.js therefore substitutes a non-binding ceiling whenever the station fit carries none — the KNYC ceiling table (the entire Kalshi edge) can never silently apply to LGA and fabricate a fade. certified: false until forward verification.

4. End-to-end read-only smoke (plumbing, not signals)

Public prices (2026-07-09 file) → settled join 2026-07-08 high=87 (clean) → thresholdBoard 77..89 → 2°F ladder via cumulative differences → KLGA params + flat 1¢ fee through robustEdgeReport. The pipe composes. The "robust edges" it prints against EOD closes are artifacts — closing prices on a same-day board already know the outcome a morning forecast doesn't. Real evaluation needs day-ahead boards nightly (available in this feed) or live intraday quotes (needs the EC entitlement).

5. What remains before any UHLGA trade

1. EC entitlement on a ForecastEx-permissioned IBKR account (#2216 found STKNOPT

only) — human/IBKR action; nothing in code can fix it.

1. Forward verification (paper): nightly join — KLGA-params calibratedDistribution

vs next-day board (from this feed) and vs settledHighs — accrue RPS/PIT + any robust-edge hit-rate to n≥20 before believing any edge. The feed's history (≥2026-02-03) also allows a retrospective day-ahead backtest first.

1. Act path: ADR-gated order code (order entry is intentionally absent from

ibkr-cpapi.js per ADR-0019); scope-gate + kill-file discipline as on Kalshi.

Evidence classes

• PROVEN: U-terms (CFTC filing), product list (venue CSV), exceed-strict convention.
• MEASURED: 14/14 ASOS-max settlement match; CLI divergence; fit + OOS numbers; −0.89°F

ASOS−CLI history gap; feed reach (2026-02-03 probe).

• HEURISTIC: flat ~1¢/contract fee (spread-embedded; #2216's measurement still pending a

permissioned account — FORECASTEX_FEE_CENTS overrides).

• UNIMPLEMENTED: deck wiring, forward-verification job, any order path.