METHODOLOGY · ENGINEERING MANUAL

Open data. Real math. Public receipts.

Prexora's signals are not opinions. They are the output of numerical weather models compared to live market pricing. Every step of the pipeline is auditable — we tell you the data sources, the gates each signal must pass, and the entire historical track record. What we don't expose is the recipe: the exact thresholds and detector weightings stay internal so the edge doesn't decay.

Forecast models
5+
NWS · ECMWF · GFS · ICON · METAR
Cities at launch
12
Tier-classified, model-tuned
Quality gates
8
Independent · all must pass
Learning system
Auto
Stricter over time, not looser
THE ROADMAP

From a weather signal to a full prediction-market engine

Prexora starts narrow and proves it works — weather on Polymarket, public receipts, an autonomous agent — then widens to every market and every venue. Here is the path.

Live now
Weather signals
PolymarketPublic track recordTelegram
This week
Autonomous agent
Executes for youNon-custodialStake / Pro
Next
Every market
SportsPoliticsCryptoElectionsEconomy
Then
Everywhere you are
Telegram appGoogle PlayApple App Store
Vision
Every venue
PolymarketKalshiAll prediction markets
01 · WHERE SIGNALS COME FROM

Five forecast models, one ground truth

Most weather signal services run one or two models. Prexora blends five numerical weather prediction (NWP) systems plus live observations — and adapts the blend to each city based on which model performs best for that climate.

⛅ ECMWF
European Centre for Medium-Range Weather Forecasts. The world's most accurate global model. Primary for international cities.
🇺🇸 NWS
National Weather Service. US authoritative forecast — what most US prediction markets settle against.
📡 GFS
NOAA Global Forecast System. Cross-check for US cities. Updated four times daily.
🌐 ICON
DWD German model. 6.5km resolution over Europe — best signal for European cities.
🌡 Wunderground
Forecast aggregator with hourly updates. Drives our drift detection signals.
✈ METAR
Real-time aviation weather observations from airport stations. The ground truth for resolution.
02 · HOW WE SCORE

Probability vs price = edge

For every market on Polymarket, we compute the probability the model says it should be at, and compare to the market's actual price. The gap is your edge. Sizing follows half-Kelly, capped to a small fraction of bankroll per signal.

edge = our_prob market_price
stake = ½ × Kelly(our_prob, price) × bankroll
Capped at 8% of bankroll per signal.
Market
35¢
Model
62¢
↑ +27¢ EDGE
03 · HOW WE PROVE IT

Every signal logged. Every outcome auto-resolved.

The moment a signal is generated, it's stamped with a timestamp and the market price at publication. The moment Polymarket settles the underlying market, our system records the outcome against the CLOB winner flag — no human intervention, no retroactive picks, no edits. The track record on this site is what came out of that process.

Signal fires
Timestamp logged
Market settles
Auto-resolved
→ See every signal
04 · CITY TIERS

We don't pretend Seoul and Atlanta are the same problem

Each city is classified by terrain, climate volatility, and which numerical model performs best for it. Different cities use different default models (Tokyo → ECMWF; Dallas → GFS; Hong Kong → ECMWF). Stakes are scaled down for high-volatility climates regardless of edge magnitude.

EASY tier

Atlanta Dallas Chicago Houston
Flat terrain, single climate regime, models agree tightly.

MEDIUM tier

NYC Boston Miami Austin Phoenix DC
Mixed terrain, two regimes, model disagreement common.

HARD tier

Los Angeles Seattle
Orography, microclimates, models often disagree.

12 US cities at Phase 1 launch. International coverage (London, Paris, Tokyo, Seoul) ships in Phase 1.5 once the city-specific calibration is locked.

05 · AUTO-LEARNING

The system gets stricter over time, not looser

Prexora reads every resolved bet and adjusts itself. If a detector loses money for 30 days, its confidence multiplier is automatically reduced. If a (city, direction) combination loses 3+ times below a 30% win rate, it's automatically blacklisted until manually cleared. If a city loses 2 bets in a single day, signals on that city pause until tomorrow. Discipline is built in.

A

Signal fires

Engine generates a pick with confidence × edge.

B

Outcome logged

Polymarket CLOB resolves the market. Win or loss recorded.

C

Detector adjusted

Losing detectors get auto-down-weighted; bad cities pause.

D

Stricter next time

Rolling-window discipline. The model gets harder to please.

06 · THE 8 QUALITY GATES

Every signal must pass all eight

Before any signal is published, it has to pass eight independent gates. Any one rejection and the signal is discarded silently. The exact thresholds stay internal — the gate names below are the categories, not the parameters.

01
Confidence floor. Below the threshold, signal is discarded.
02
Edge floor. Volume-adjusted edge must clear a minimum.
03
Volume floor. Book depth must be sufficient to absorb retail flow.
04
Entry price band. Must be in a tradeable range (not phantom-priced sub-cent or asymmetric Zone 4/5).
05
Expected value. Math gate — EV per dollar staked must clear a minimum.
06
City cold-streak. Recent rolling WR on the city above threshold.
07
Daily stop-loss. Anti-revenge: city pauses after losses today.
08
Direction asymmetry. Block directional bets where Kelly asymmetry kills the trade.
TRUST STATEMENT

Prexora is research and education, not financial advice. We do not place bets for you. We do not custody your funds. Every signal is timestamped at publication and auto-resolved against Polymarket CLOB. The track record is computed on Prexora's standard $2,500/day model portfolio — fully auditable, nothing deleted, nothing backfilled.