Documentation

Complete reference for WeatherRadar AI — the precision weather prediction market trading platform.

Getting Started

WeatherRadar AI is a quantitative trading platform built for Kalshi temperature prediction markets. It ingests meteorological ensemble models, computes probability distributions over market strike buckets, and derives edge-based trade signals — all in a unified interface designed for both manual and autonomous execution.

The platform combines forecasts from GEFS, ECMWF, HRRR, NBM, and NDFD into a single blended fair value (FV) estimate per market series. By comparing this FV against live Kalshi bid/ask spreads, WeatherRadar AI surfaces actionable edges before market settlement. Calibration data is maintained continuously so that the signal quality score reflects real-world model performance.

Whether you are placing manual trades using the Trader interface or running fully autonomous execution on Elite, WeatherRadar AI is designed to give you a systematic, data-driven edge in short-term temperature markets.

Account Creation and Plan Selection

Create an account at weatherradar.ai and select a plan that matches your trading activity:

Your First Forecast

1. Log in to your account and navigate to the Trader page from the top navigation.
2. Select a city from the sidebar series list. Active markets display with live Bid and Ask prices pulled from Kalshi.
3. Review the Fair Value column. This is the platform's probability estimate for each market bucket, derived from the blended ensemble models.
4. Check the Edge column. Positive edge (FV above market mid) indicates an underpriced market; negative edge indicates an overpriced market.
5. Click any row to open the detail panel. Review the FV gauge, observation trend, and signal quality score before placing a trade.
6. If a signal qualifies (see Buy & Sell Signals), click through to Kalshi to execute the trade manually, or configure auto-trading if on Elite.

Reading the Interface

The Trader page is divided into three regions: the series sidebar on the left, the market table in the center, and the detail panel on the right. Each region updates at different intervals based on data source freshness.

The series sidebar lists all tracked cities and date combinations. Markets are grouped by settlement date. Clicking a series loads its full market table and activates the detail panel for the selected row. The sidebar also displays the last model update timestamp and the current calibration grade for each city.

The detail panel expands to show the FV gauge, the full probability distribution over all strike buckets, recent ASOS observations, model blend breakdown, and active signals. Elite users also see auto-trading status and pending order information for the selected series.

Market Table Columns

The FV Gauge

The FV gauge in the detail panel visualizes the full probability distribution produced by the fair value engine across all strike buckets for the selected series. Each bucket is rendered as a bar proportional to its assigned probability, forming a temperature probability curve.

The highlighted bucket corresponds to the currently selected market row. The gauge also overlays the market mid-price as a horizontal reference line so you can visually compare the model's distribution center against where the market is pricing probability mass.

When the distribution is narrow and centered away from the market mid, edge tends to be large and signal quality tends to be higher. When the distribution is broad, edge values are typically smaller and the Q score will reflect increased uncertainty.

Observation Panel: ASOS Temperatures

The observation panel displays recent temperature readings from the nearest ASOS (Automated Surface Observing System) station associated with the market's city. Readings are ingested at hourly intervals and displayed as a timeline chart with labeled data points in degrees Fahrenheit.

The trend indicator shows the directional temperature movement over the last three observations: rising, falling, or steady. This is used as a secondary signal filter — when the observed trend is directionally consistent with the model's predicted peak temperature, signal confidence increases.

Fair Value Engine

Ensemble Blending

The fair value engine ingests forecast data from six primary model sources and blends them into a single consensus forecast using performance-weighted averaging:

• GEFS — Global Ensemble Forecast System (NOAA, 31-member ensemble)
• ECMWF — European Centre for Medium-Range Weather Forecasts (high-skill global model)
• HRRR — High-Resolution Rapid Refresh (NOAA, hourly convection-allowing model)
• NBM — National Blend of Models (NWS, post-processed multi-model blend)
• NDFD — National Digital Forecast Database (NWS gridded forecasts)
• Additional regional and experimental models ingested as available

Model weights are updated daily based on rolling CRPS scores computed against ASOS verification data. Models with lower (better) CRPS receive higher blend weights. For near-term forecasts (6–18 hours), HRRR typically carries the largest weight due to its high resolution and frequent update cycle. Beyond 48 hours, ECMWF and GEFS weights increase.

Gaussian Integration over Market Buckets

Once the blended temperature forecast (mean μ and standard deviation σ) is computed for a city and valid time, the engine integrates a Gaussian distribution over each Kalshi market bucket to assign probabilities.

For a bucket defined by lower bound L and upper bound U (in °F), the fair value probability P is:

Where Φ is the standard normal CDF. The resulting probabilities across all buckets for a series sum to 1.0. This probability (expressed in cents, 0–99) becomes the Fair Value displayed in the market table.

The engine applies a sigma compression factor to prevent overconfidence at extreme forecast horizons and during high-uncertainty regimes detected by the regime classifier. A compressed sigma results in a flatter distribution and lower edge magnitudes.

Bias Calibration and CRPS Grades

Forecast calibration is measured using CRPS (Continuous Ranked Probability Score) — a proper scoring rule that rewards probabilistic forecasts that are both accurate and sharp. Lower CRPS indicates better calibration.

The platform computes rolling 30-day CRPS for each city and model combination and assigns a letter grade displayed in the interface:

Signal Quality Score (Q)

The signal quality score Q is a composite metric on a 0–1 scale that reflects overall confidence in the current FV estimate and trade signal. It is computed from three sub-components:

• Calibration component — derived from the current CRPS grade (A = 1.0, D ≈ 0.2)
• Model agreement component — inverse of the inter-model spread; high agreement across GEFS, ECMWF, and HRRR increases Q
• Observation consistency component — measures agreement between the latest ASOS reading and the model's predicted trajectory

A Q score above 0.70 is required for a signal to display in the market table. Auto-trading has a higher threshold of 0.75 for autonomous order entry. The Q score is shown in the detail panel for every market row.

Trading Guide

Understanding Edge

Edge is the fundamental measure of expected value in a prediction market trade. It quantifies how far the platform's fair value diverges from where the market is currently priced.

edge = FV - market_mid

where:
  FV          = platform fair value estimate (in cents)
  market_mid  = (bid + ask) / 2
  edge > 0    = market underpriced — model expects higher probability
  edge < 0    = market overpriced — model expects lower probability

Edge alone does not determine whether a trade is worth taking. Small edge values may not justify transaction costs or spread capture risk. The signal thresholds below incorporate empirically derived minimums based on historical execution data.

Buy and Sell Signals

Signals are generated when edge exceeds defined thresholds and the market mid is within a range where execution is practical:

Kelly Sizing Formula

Position sizing follows a fractional Kelly criterion. For a binary market (YES/NO outcome), the Kelly fraction is:

WeatherRadar AI applies a half-Kelly multiplier by default to account for model uncertainty and reduce variance. The recommended stake size per trade is therefore 0.5 × f* of your total active capital.

Kelly fractions below 2% are treated as noise and no position is recommended. Fractions above 8% are capped at 8% to prevent over-concentration in a single market.

Risk Management

The platform enforces two hard risk limits regardless of signal strength:

• 4% per-ticker cap — no single market series may represent more than 4% of total portfolio value at entry. Auto-trading will not enter a position that would breach this limit.
• Max drawdown gate — if the portfolio experiences a drawdown exceeding a user-configured threshold (default 15% from peak), auto-trading suspends all new entries until the session is manually reset. Manual trading signals continue to display.

Auto-Trading (Elite)

Elite plan members can connect their Kalshi account via API key and enable autonomous trade execution. When active, the auto-trading engine monitors all tracked markets, computes edge and signal quality in real time, and submits orders without manual intervention.

Prerequisites

• Active Elite subscription
• Kalshi API key (generated in your Kalshi account settings under API Access)
• API key entered in WeatherRadar account settings under Integrations > Kalshi
• Auto-trading toggle enabled per city in the Trader sidebar

How Autonomous Execution Works

Once enabled, the engine runs a continuous evaluation loop on a per-city basis. Each evaluation cycle:

1. Fetch latest model blend and recompute FV for all active buckets in the series.
2. Pull live bid/ask from Kalshi order book. Compute edge and Q score.
3. If a qualifying signal is detected (edge threshold met, Q ≥ 0.75, drawdown gate open, position cap not breached), compute Kelly-sized stake.
4. Submit a FAK limit order at the ask (for buys) or bid (for sells). Order expires immediately if not filled at that price.
5. Log execution result, update position tracking, and re-evaluate in the next cycle.

Trading Windows

Auto-trading does not run 24 hours a day. Each city has a trading window defined around its expected peak temperature time — typically a 4–6 hour window centered on the afternoon peak for high-temperature markets, or early morning for low-temperature markets.

Outside the trading window, the engine enters monitor-only mode: it tracks positions and may exit unfavorable positions, but will not enter new ones. This prevents execution during low-liquidity overnight periods where spread costs erode edge.

FAK (Fill-and-Kill) Limit Orders

All auto-trading orders are submitted as Fill-and-Kill limit orders. This order type attempts immediate execution at the specified limit price and cancels any unfilled portion instantly. WeatherRadar AI uses FAK exclusively — no resting orders, no GTC (Good-Till-Cancelled) orders.

The FAK limit price is set at the current best ask (for buys) or best bid (for sells) at the time of submission. If the book has moved and the order cannot fill at that price, the order is cancelled and the engine waits for the next evaluation cycle.

Position monitoring and exit triggers: Open positions are monitored continuously. An exit order is triggered if:

• The edge reverses and becomes significantly negative (FV has moved against the position)
• The Q score drops below 0.50 (model confidence has degraded materially)
• The market is within 30 minutes of settlement cutoff
• The drawdown gate is triggered at the portfolio level

API Reference

Authentication

The WeatherRadar AI API uses JWT-based session authentication. Authenticate by sending a POST request to /auth/login with your credentials. The server returns a session cookie that must be included in all subsequent requests.

POST /auth/login
Content-Type: application/json

{
  "email": "[email protected]",
  "password": "your_password"
}

// Response sets HttpOnly cookie: wr_session=<jwt_token>

For programmatic access, pass the cookie in all subsequent requests using your HTTP client's cookie jar, or extract the token and send it as a Cookie: wr_session=<token> header.

Endpoints

GET /api/fair-value

GET /api/fair-value?series=HIGHNY&date=2026-04-12
Cookie: wr_session=<token>

{
  "series": "HIGHNY",
  "date": "2026-04-12",
  "model_blend": {
    "mu": 68.4,
    "sigma": 3.2,
    "weights": { "HRRR": 0.38, "ECMWF": 0.29, "GEFS": 0.18, "NBM": 0.10, "NDFD": 0.05 }
  },
  "signal_quality": 0.82,
  "buckets": [
    { "ticker": "HIGHNY-26APR12-T60", "lower": 58, "upper": 62, "fv": 3 },
    { "ticker": "HIGHNY-26APR12-T64", "lower": 62, "upper": 66, "fv": 12 },
    { "ticker": "HIGHNY-26APR12-T68", "lower": 66, "upper": 70, "fv": 41 },
    { "ticker": "HIGHNY-26APR12-T72", "lower": 70, "upper": 74, "fv": 33 },
    { "ticker": "HIGHNY-26APR12-T76", "lower": 74, "upper": 78, "fv": 9 },
    { "ticker": "HIGHNY-26APR12-T80", "lower": 78, "upper": 999, "fv": 2 }
  ],
  "updated_at": "2026-04-12T06:15:00Z"
}

GET /api/obs-history

GET /api/obs-history?series=HIGHNY
Cookie: wr_session=<token>

{
  "series": "HIGHNY",
  "station": "KJFK",
  "observations": [
    { "time": "2026-04-12T06:00:00Z", "temp_f": 61.0 },
    { "time": "2026-04-12T05:00:00Z", "temp_f": 59.4 },
    { "time": "2026-04-12T04:00:00Z", "temp_f": 57.9 }
  ],
  "trend": "rising"
}

GET /api/public/stats

{
  "trades_total": 4821,
  "win_rate": 0.613,
  "avg_edge_captured": 8.4,
  "cities_tracked": 18,
  "models_active": 5,
  "last_updated": "2026-04-12T07:00:00Z"
}

Response Format

All endpoints return application/json. Error responses follow a consistent structure:

{
  "error": {
    "code": "UNAUTHORIZED",
    "message": "Valid session required for this endpoint.",
    "status": 401
  }
}

FAQ

Changelog

• v8.0
  April 2026

  SaaS Platform Launch

  • Full SaaS platform with account management, billing, and plan tiers
  • Pro and Elite subscription plans with feature gating
  • Dashboard with portfolio performance and calibration status views
  • API endpoints for external integrations (Pro+)
  • Email notification system for signal alerts and auto-trading events
• v7.0
  April 2026

  FAK Execution and Tranche Entry System

  • Migrated order submission to Fill-and-Kill limit orders exclusively
  • Tranche entry system: large positions split into 3 sequential fills to reduce market impact
  • Per-city trading windows with configurable offsets
  • Position monitoring loop with automated exit triggers
  • Max drawdown gate added to auto-trading engine
• v6.0
  April 2026

  FV Engine Overhaul

  • Complete rewrite of fair value computation pipeline
  • Sigma compression factor introduced via regime detector module
  • CRPS-based model weighting replaces fixed weights
  • Signal quality score Q introduced as composite confidence metric
  • ECMWF integration added; model count increased from 4 to 6
  • Calibration grades (A/B/C/D) added to interface and API
• v1.0
  March 2026

  Initial Kalshi Trading Engine Launch

  • Initial release of the Kalshi temperature market trading engine
  • GEFS, HRRR, NBM, and NDFD model sources integrated
  • Gaussian bucket integration for fair value computation
  • Manual signal display in Trader interface
  • ASOS observation ingestion and trend detection