WeatherBetter

How to read this

• Model P — P(YES) from Normal(μ, σ) vs Kalshi strike rules (strict >/<)
• μ — NWS daytime forecast high; settlement is KNYC Daily Climate Report max
• σ / bias — uncalibrated priors (2.0–3.5°F) unless Calibration tab is enabled; empirical values range 3.8–6.4°F with +1.6°F bias
• Signal — Buy YES/NO when edge ≥10pp and EV ≥5¢ per $1 (after fee buffer)
• Enable empirical calibration in the Calibration tab for historically-fit σ and bias.

Quick start

1. Hit ↻ Refresh (top right) to load the latest NWS forecast and Kalshi snapshot.
2. On Forecast, scan the 7-day high/low and rain outlook.
3. On Edge finder, each row is one Kalshi contract. The Signal column is the headline: green Buy YES / teal Buy NO badges mark where the model sees value.
4. Type a price in a YES ¢ / NO ¢ box to test “what if the ask were X?” scenarios.
5. Click a contract name to expand its exact resolution rules.

Forecast tab

• Metric cards — today’s forecast high (μ), tonight’s low, rain chance (NWS POP), wind.
• 7-day temperature — daily forecast high/low range (capped at the ~7-day NWS horizon).
• Precipitation — NWS probability of precipitation per day.

Edge finder — columns

• Contract — the Kalshi strike (e.g. “78° to 79°”). Click to expand resolution rules.
• Model P — the model’s P(YES) for that contract. Shown clamped 1–99% for readability; the edge/EV uses the exact value.
• YES/NO ask — current Kalshi ask prices (¢). “—” means that side has no real liquidity.
• YES ¢ / NO ¢ — type an override to price a hypothetical ask (persists; bypasses the illiquid flag).
• Signal — the recommendation (see badges below).

Edge finder — date headers & badges

• Date header — shows μ, its source (period = NWS daytime forecast, grid = gridded-max fallback), σ, and lead days. “μ unavailable” = NWS down; “beyond forecast horizon” = date past the ~7-day NWS reach.
• Buy YES / Buy NO — edge ≥10pp and EV ≥5¢ per $1 after the 2¢ fee buffer.
• No edge — model and market roughly agree, or the edge is too small.
• Illiquid — no tradeable two-way book (grayed out). Enter an override to test it anyway.
• Stale — settlement date is in the past (from a stale snapshot).
• Banners — a red banner warns when the snapshot is >45 min old or NWS is unavailable.

How the model works

1. μ — NWS daytime forecast high for the date (°F). Settlement is the official KNYC Daily Climate Report max.
2. σ — forecast-error spread by lead time: 2.0°F day-of, 2.5°F (+1d), 3.5°F (2d+).
3. P(YES) — treats the high as Normal(μ, σ) and computes each contract’s strike probability, with a ±0.5°F continuity correction (temps settle in whole degrees). Three rule types: greater, less, inclusive between.
4. Signal — compares model P(YES) to the market ask; flags a buy when edge ≥10pp and EV ≥5¢ per $1 after a 2¢ fee/spread buffer.

Known limitations — read before trusting a signal

• Priors are uncalibrated by default — enable the Calibration tab toggle to use historically-fit σ (3.8–6.4°F) and +1.59°F bias. With priors on, long-lead signals are overconfident.
• Long-lead overconfidence (priors only) — prior σ is flat 3.5°F for leads ≥2d; real error grows to ~6°F by day 7. Calibrated σ fixes this.
• Calibration window — fitted on Jan–May 2026; may not generalize to summer/winter temperature ranges where forecast behavior differs.
• Per-contract pricing — each contract is scored independently. A complete strike ladder sums to ~100%, but a partial snapshot won’t, and cross-date correlation isn’t modeled.
• Buy-only — signals evaluate buying at the ask. Exiting an existing position fills at the bid and isn’t modeled.