Every weather market on Polymarket — "Will the daily high in London exceed 22°C tomorrow?", "Will Paris hit 28°C this weekend?" — resolves to a single observable number from a single station at a single moment. Yet between the last forecast revision and resolution, the odds drift, oscillate, and reset multiple times. That drift is not random. It's structured by the cadence at which the underlying forecasts update.
Forecast cadences are predictable. Markets are reactive.
Model schedules at a glance
The four major numerical weather prediction models that WIN Weather Bot consumes run on schedules that haven't meaningfully changed in years:
- GFS (NOAA, USA) — full run every 6 hours (00, 06, 12, 18 UTC), available ~3.5 h after init
- ECMWF (Reading, UK) — full run every 12 hours, available ~7 h after init, the most respected for medium-range
- ICON (DWD Germany) — full run every 3 hours, fast updates on the European domain
- NAM/HRRR (NOAA short-range) — sub-hourly for the contiguous US
Where the asymmetry comes from
Polymarket order books, on the other hand, can update in milliseconds. The asymmetry: the underlying truth refreshes a few times a day, the price refreshes a few times a second. In theory the market should perfectly anticipate the next model run. In practice, this is rarely the case.
The "reprice lag" — measured in minutes, not seconds
The 30 to 90 minute reprice window
When ECMWF releases its 12 UTC run at around 19 UTC (8 PM London time), the new forecast values are public on a handful of paid feeds within 10–15 minutes. The first retail traders to react do so 30 to 90 minutes later, often via Twitter screenshots or weather-newsletter pushes. That 30–90 minute window — during which the order book is still pricing tomorrow's high using the previous run — is the structural source of edge.
What the bot watches for
The bot doesn't try to predict the weather better than ECMWF. It tries to detect when a fresh model run has materially changed the expected outcome before the market reprices.
What "material" means in practice
For a binary "Will London exceed 22°C?" market, a forecast shift of 0.3°C is noise. A shift of 1.5°C across multiple models, near the threshold, with consistent direction — that's signal. The bot computes a weighted ensemble (ECMWF + GFS + ICON), the spread vs. the current market mid, and a confidence score from disagreement among models. When the score crosses a threshold and the spread remains open, it submits.
Why this edge persists
Why retail still dominates this market
Retail dominates weather markets in volume. Professional weather desks at hedge funds don't compete here — the size is too small to justify the infrastructure cost. Polymarket's geographic restrictions further thin the pool of sophisticated participants. The result is a market with persistent, structural mispricings — measurable in minutes per day per city.
How long does the edge last?
The window is closing slowly as more bots arrive. But for now, it remains the cleanest available form of weather-derived alpha in size-constrained markets.