agiprolabs/claude-trading-skills/kalshi-weather-markets

Kalshi Weather Markets — Daily Temperature High/Low

Kalshi lists daily high and low temperature options for ~20 US cities as binary contracts that settle YES ($1.00) or NO ($0.00). This skill covers the market structure, the forecast-to-probability map, exact settlement mechanics, and hard-won pitfalls. It builds on the exchange layer — for Kalshi API mechanics (host, auth, orders, order book, candlesticks) see the kalshi-api skill; for strategy, sizing, and backtesting see prediction-market-strategy.

Contract Types

Brackets — B<center>

A bracket ticker B<center> is a 2°F-wide, both-ends-inclusive window.

• B74.5 covers the two integers {74, 75}°F.
• YES iff the settled temperature is exactly 74 or 75.
• Brackets in one event are mutually exclusive and (with two open tail markets) collectively exhaustive.
• Their YES prices sum to the overround (fair = 1.0; > 1.0 = aggregate overpricing).

Thresholds — T<strike>

A threshold ticker T<strike> is a one-sided binary.

• greater → YES iff cli >= strike + 1
• less → YES iff cli <= strike - 1
• Critical: strike_type ("greater" / "less") is not inferable from the ticker. Read it from the API strike_type field every time.

Ticker Format

KXHIGH<CITY>-<YYMONDD>-B<center>     # bracket high
KXLOW<CITY>-<YYMONDD>-T<strike>      # threshold low

The date is encoded in the ticker, not derivable from `close_time`. KXHIGHNY-26JUN21 settles 2026-06-21 LST. close_time is next-day UTC (~00:59 ET). Joining on close_time off-by-ones every label — use the ticker date.

Forecast → P(YES)

Given a forecast distribution N(μ, σ) for the day's extreme, apply the half-integer continuity correction (mandatory — settlement is on integers, not a continuous scale):

# Bracket B<center>, covering integers {floor, cap}
P(YES) = Φ((cap + 0.5 − μ) / σ) − Φ((floor − 0.5 − μ) / σ)
 
# Threshold "greater":
P(YES) = 1 − Φ((T + 0.5 − μ) / σ)
 
# Threshold "less":
P(YES) =     Φ((T − 0.5 − μ) / σ)
 
Φ(x) = 0.5 · (1 + erf(x / √2))   # stdlib only, no scipy needed

The ±0.5 shift is not optional. Dropping it biases every bracket. Treating 2°F brackets as 1°F half-open windows produced a +1640% phantom backtest in one project.

See scripts/weather_brackets.py for runnable implementations of all four functions.

Deriving (μ, σ) from Ensemble Quantiles

sigma_raw = max((p90 − p10) / 2.56, 0.5) · sigma_scale · sigma_mult
mu        = p50                          # or nowcast-blended (see forecasting.md)
sigma     = max(sigma_raw, 0.1)          # hard floor against degeneracy

The 2.56 divisor is the 10th–90th percentile span of a standard normal (2 × 1.28σ).

CLI-Space Bias Correction

The settlement value (NWS CLI integer °F, LST day) is not the same as raw ASOS/METAR hourly max/min — CLI applies QC, backup-station fallback, and LST aggregation. Shift μ before computing P(YES):

mu_cli = mu_metar + bias_city_season     # bias = oracle_extreme − asos_extreme, fit per city + season

Fit bias_max / bias_min as seasonal (circular) curves per city. Skipping this systematically misprices every bracket for cities with a structural CLI/METAR gap.

Settlement Rules

Kalshi

• Source: NWS Climatological Report (CLI) — the official daily climate summary issued by each WFO.
• Fallback: IEM ASOS daily download matches CLI 100% and is available programmatically.
• Window: LST (Local Standard Time), no DST adjustment. The day runs midnight-to-midnight LST year-round.
• Value: Integer °F maximum (HIGH) or minimum (LOW) temperature for that LST day.
• Bracket: YES iff cli ∈ {floor, cap} (both ends inclusive).
• Threshold greater: YES iff cli >= strike + 1.
• Threshold less: YES iff cli <= strike - 1.

Settlement-Source References

Read each market's own rulebook before scoring or trading. Settlement source, station, and day-window are per-market contract terms that can change.

Cross-Venue Divergence

The same metro on the same date can settle to different values across venues — both because of the station and the DST window in spring/fall.

Any cross-venue analysis must settle each leg on its own source.

Nowcast Blending (Same-Day Path)

Once an intraday observation is available, pull μ toward reality and shrink σ:

• HIGH: clamp μ to [obs, obs + drift · hours_remaining]
• LOW: clamp μ to [obs − drift · hours_remaining, obs]
• σ shrinks as sigma_raw · sqrt(hours_remaining / 24), floored at sigma_floor (≈ 0.5)
• drift ≈ 3.0°F/hr default

Optional NWP prior blend: new_p50 = w · hrrr + (1−w) · p50 (w ≈ 0.5), then rebuild symmetric quantiles using a calibrated σ.

Calibrated Model Performance (Reference Numbers)

Per-city OOS Brier scores across 22 highs + 22 lows (v1.5, 2026-06-17 baseline):

Forecast skill ≠ trading edge. A calibrated model that beats climatology by 0.05 Brier does not guarantee positive EV at market prices — the market already incorporates NWP. The practical edge is maker-side fading of mispriced longshot brackets (favorite–longshot bias), not raw directional forecasting.

Weather Pitfalls

1. Wrong settlement source. Scoring against a derived truth that correlates with but differs from the venue's resolution flips ~10% of outcomes. Settle on the venue's own result.

1. Bracket off-by-one (phantom +1640%). Treating 2°F inclusive brackets {floor, cap} as 1°F half-open [floor, cap) manufactures a large phantom backtest edge. The bracket is both-ends-inclusive.

1. strike_type not inferable from ticker. T74 on a low market might be greater or less. Always read strike_type from the API. Never guess.

1. Date-in-ticker, not `close_time`. Use the date embedded in the ticker string for settlement-date joins, not close_time (which is next-day UTC).

1. LST ≠ local clock. Kalshi settles on LST (no DST). In spring/fall, the LST window shifts relative to local time. Cross-referencing WU (which uses local clock) against CLI on DST-transition days will produce mismatches.

1. CLI ≠ METAR. Raw ASOS hourly max/min is not the settlement value. CLI applies QC, backup-station fallback, and LST aggregation. Fit per-city seasonal bias corrections before computing P(YES).

1. UTC vs local-day feature aggregation. Aggregating forecast features over UTC days instead of LST days misaligns labels — cost ~14 percentage points of accuracy in one study.

1. Clock-mismatch look-ahead. Filling at an 18:00Z book snapshot while features are cut at 14:00 LST trades non-Eastern cities on future information. Use each city's own local decision time.

1. Phantom penny asks. 1¢ ask levels are frequently spoofed; assuming you fill them over-credits PnL ~23×. Count only depth that persists across snapshots and is corroborated by trade prints.

1. Overround as a diagnostic. Sum the YES prices across an event's full bracket set. overround > 1.0 is normal (house edge); overround >> 1.1 signals a mispriced event (or data error).

Files

References

• references/brackets-and-settlement.md — Bracket/threshold structure, P(YES) formulas, settlement rules, cross-venue divergence table, overround
• references/forecasting.md — Ensemble quantiles → (μ, σ), nowcast blending, CLI-space bias correction, model skill numbers, forecast ≠ edge

Scripts

• scripts/weather_brackets.py — Gaussian bracket/threshold P(YES), settlement resolution, and quantile→(μ,σ) functions (pure stdlib, runs offline)