Zephyr

Zephyr is a weather-driven prediction market trading system focused on exploiting probability mispricing in event-based contracts.

Instead of “guessing” outcomes, Zephyr treats prediction market prices as implied probabilities and compares them to forecast probabilities derived from professional meteorological models.

---

Core Idea

Prediction markets price contracts like:

• YES = $0.60 → market implies a 60% chance
• NO = $0.40 → market implies a 40% chance

Meanwhile, modern ensemble weather models (NOAA GEFS, ECMWF, HRRR) provide probabilistic forecasts.

Zephyr finds situations where:

Forecast probability significantly differs from market probability

Example:

• Forecast says 80% chance of >5" snow
• Market is priced at 60%
• Edge = +20% expected value

Zephyr trades only when this gap is large and stable.

---

Strategy Overview

1. Forecast → Probability

Zephyr converts ensemble outputs into probabilities:

[ P(event) = \frac{#runs\ exceeding\ threshold}{N} ]

Example:

• 42 of 50 runs predict >5 inches snow
  → Forecast probability = 84%

---

2. Market Price → Probability

Market price is treated directly as probability:

[ P_{market} = price ]

---

3. Trade Only When Mispricing Is Large

Zephyr enters positions only if:

• Probability gap exceeds 10–20%
• Multiple models agree
• Forecast confidence is increasing

---

Best Target Markets

Zephyr performs best on clean threshold contracts:

• Temperature cutoffs
  “Will NYC exceed 85°F tomorrow?”
• Snowfall totals
  “Will Boston get >3 inches this week?”
• Hurricane landfall events

These markets are more structured and forecastable than vague outcomes.

---

Risk Management

Weather has variance, even with strong models.

Zephyr enforces strict bankroll controls:

• Max 2–3% exposure per contract
• Portfolio of many small edges
• No single-event oversized bets

Long-term profitability comes from repetition, not jackpots.

---

Data Sources

Planned integrations include:

• NOAA GEFS ensemble forecasts
• ECMWF ensemble probabilities
• HRRR radar nowcasts
• Emergency bulletins and alerts
• Market pricing feeds (Polymarket)

---

Roadmap

• NOAA ingestion + ensemble probability computation (prototype for one test event)
• Market price scraper + implied probability engine (Polymarket public quote pull)
• Edge detection + trade signal generator
• Backtesting framework (CSV-driven)
• Automated execution + bankroll rules (paper-order logging + fractional Kelly + hard cap)

---

Run The Prototype

This repository now includes a first working step for GEFS ingestion and event-probability computation:

scripts/gefs_event_probability.py

It pulls the latest GEFS ensemble run from NOAA NOMADS and computes:

[ P(event) = \frac{runs\ exceeding\ threshold}{total\ runs} ]

Default test event:

• NYC max 2m temperature tomorrow (America/New_York) >= 85F

Precipitation events are also supported in the pipeline (e.g., total precip >= 0.1 in), and can be logged via scripts/log_snapshots.py when Polymarket markets match the filters.

Run:

python3 scripts/gefs_event_probability.py

Example with custom threshold:

python3 scripts/gefs_event_probability.py --threshold-f 30

---

Generate A Trade Signal

Build a forecast probability from GEFS and compare it to a market implied probability.

Option A (manual market probability):

python3 scripts/generate_signal.py --market-probability 0.42

Option B (live Polymarket quote by market slug):

python3 scripts/generate_signal.py --polymarket-slug your-market-slug

Precipitation example:

python3 scripts/generate_signal.py \
  --event-type precip_total \
  --threshold-in 0.1 \
  --polymarket-slug your-market-slug

Paper-log orders when a trade is triggered:

python3 scripts/generate_signal.py \
  --market-probability 0.42 \
  --paper-ledger data/paper_orders.csv

---

Run A Backtest

Run the built-in sample backtest:

python3 scripts/run_backtest.py

Show per-trade detail:

python3 scripts/run_backtest.py --show-trades

Use your own CSV:

python3 scripts/run_backtest.py --csv path/to/your_backtest.csv

CSV columns:

• event_id
• contract_ticker
• forecast_probability
• market_probability
• outcome (1 for YES outcome, 0 for NO outcome)
• optional: timestamp

---

Log Recent Snapshots (SQLite)

Discover Polymarket weather/temperature markets (filtered by volume, city, and date window), compute GEFS probabilities, and store snapshots in SQLite:

python3 scripts/log_snapshots.py --db data/zephyr.sqlite

Use data/market_universe.json to tune cities, volume, and date window.

Record a resolved outcome:

python3 scripts/record_outcome.py --market-slug your-market-slug --outcome 1

Export joined snapshots + outcomes for backtesting:

python3 scripts/build_backtest_from_db.py --db data/zephyr.sqlite

Then run the backtest:

python3 scripts/run_backtest.py --csv data/backtest_from_db.csv

---

Philosophy

Zephyr is built on one principle:

Trade weather markets like quant probability arbitrage, not gambling.

Forecasts are data. Prices are probabilities. Profit is the gap.

---