The Hong Kong Lock: We Stopped Forecasting the City and Started Reading It
Polymarket settles every Hong Kong temperature market on a single number — the official daily maximum logged by the Hong Kong Observatory (HKO). For one city, that turns weather trading from a forecasting problem into a reading problem. So we stopped predicting a grid cell near Hong Kong and wired the engine straight into the instrument that decides the payout. Three things shipped today, and together they make Hong Kong the sharpest market we run.
Three upgrades, one source
Dominant model — 50% weight
HKO's official 9-day forecast now carries more weight in the Hong Kong consensus than GFS, ECMWF and UKMO combined. The globals stay on as a cross-check; GFS is trimmed for its subtropical warm bias.
~90% accuracy, lead-time aware
HKO's own published hit rate is roughly 90% at 1-3 days, tapering to 85% (4-7d) and 80% (8-9d). Its weight tapers on the same curve, so the engine leans hardest on it exactly when it is most right.
Live gauge — polled hourly
On resolution day the engine reads the HKO headquarters temperature every few minutes and tracks the running daily maximum: the literal settlement number, watched live, not forecast.
Why the gauge beats every model
A global model has two jobs: forecast the weather, and forecast the offset between its grid cell and the real station. It never sees the second error, so it never fully corrects it. HKO has one job — its own station — and on resolution day it has zero jobs, because it stops forecasting and simply reports. You cannot out-model the thing that defines the answer.
| Global models GFS · ECMWF · UKMO |
HKO Hong Kong Observatory |
|
|---|---|---|
| Predicts | A grid cell near the city | The exact gauge that settles the market |
| Station error | Unknown, persistent offset | Zero — it is the gauge |
| Day 1-3 accuracy | Good | ~90% (official) |
| Live intraday read | None | Yes — hourly |
| HK weight | Secondary cross-check | Dominant (50%) |
The lock: where the edge lives
A daily maximum only ratchets upward — it never falls back. So the moment the live HKO reading touches, say, 33°, the market line "33° or higher" is settled before settlement. If the order book is still asleep and pricing it at 70¢, the engine takes it and holds to $1.00. A winning share pays $1.00, so buying a locked line at price P books $1 − P per share — a return of (1 − P) / P on capital:
| Price of a locked line | Profit / share | Return on capital |
|---|---|---|
| 35¢ | 65¢ | +186% |
| 50¢ | 50¢ | +100% |
| 65¢ | 35¢ | +54% |
| 75¢ | 25¢ | +33% |
| 83¢ (ceiling) | 17¢ | +20.5% |
The engine refuses to pay above 83¢ — past that, the run to $1.00 stops covering trading costs. It also refuses anything under a 35¢ floor: a "locked" line trading that cheap is more likely a bad read than free money, so it trusts the book and skips. And it always reaches for the cheapest locked line first, so real fills tend to sit well below the ceiling.
The mirror trade: after the peak
Hong Kong tops out in the early-to-mid afternoon, then cools. Once it is evening and the live read has fallen a clear margin off the day's high, the maximum is effectively frozen. The engine then buys the bucket the day actually landed in — the "no hotter today" line — and treats everything above it as a near-certain NO. Same lock logic, opposite end of the day, a second mispricing window every session.
Live on the dashboard
HKO now sits in the data-source rail next to GFS, ECMWF, UKMO, NWS and NCEI — own live status, a per-market temperature readout for Hong Kong, and the official running maximum updating in real time. Open the app on a Hong Kong resolution day and you are watching the settlement number form, tick by tick. Hong Kong has always been one of our deepest markets; now it is also the most precise one we run.
StormBot is a tool, not financial advice. Markets carry risk, a "locked" line assumes clean resolution, and past results never guarantee future ones. Trade responsibly.