The predicted daily high for 21 cities → the Polymarket temperature bucket to bet, with an honest read on when there's real edge. Live data, no server.
❓ Guide — what every term & number means🧮 How it's calculated
Predicted daily high = median of 6 weather models + yr/MET Norway (+ NWS/NBM for US cities), then bias-corrected from each city's own track record. The win-odds come from the ensemble distribution (100+ members) and use that city's real spread of past misses, not the optimistic model spread. The bet is the Polymarket bucket a margin above/below that, at each market's resolution airport (settled on the Wunderground recorded high). A city needs ≥3 logged days before calibration kicks in.
How these are calculated — for each selected city we predict the day's high as the
median of the 100+‑member ensemble (a single wild model can't skew it), bias‑corrected from that city's
real recorded misses; models chronically wrong there are skill‑trimmed (load Accuracy). The win% is the
share of ensemble members that make the bet win; Net edge = win% − price − a fee/slippage cost; Stake is
¼‑Kelly for your bankroll. The flag reads the bet against that city's real track record:
opportunity = clear · watch = caution ·
trap = the edge looks big but the models have been unreliable there or the market strongly disagrees, so it's
probably not free money — hover the flag for why. You place the bets yourself.
All picks
What each column is — Median (used) is the bias‑corrected ensemble median that drives the bet;
Mean & yr are reference only. No ↑ / No ↓ are symmetric "Buy No" bets a margin above/below the prediction,
each showing the live price, the models' win chance and the edge (win% − price). Calibration is the bias correction
and σ applied for that city. Resolution: every market settles on the Wunderground station high — the exact
source Polymarket names in its rules — and that's the number every temperature here is measured against
?.
Simple shows the actionable columns; Advanced adds the reference ones (Mean · yr · Cost · Calibration).
↳ How to read a row:Median (used) is the bias-corrected prediction that drives the bet; the two No cells are the actual bets (bucket · price · win% · edge); the green edges that survive a fee — the Net edge on the Top-5 cards above — are the real signal.
ⓘEdge & win% assume the forecast is right.
Details
Forecast stability — has the forecast for this day been moving?
How this is calculated — we pull the day's high as forecast in the last few model runs (now vs 1 / 2 / 3 runs ago) and show the swing (max − min). A big swing = the forecast is still unsettled = the bet is riskier than a calm win% suggests. Uses the models' previous runs.
History
What this shows — for one city, the archived model forecast for each past day vs the bucket it actually resolved to, with a bias-trend bar (red = ran warm, blue = cold, green = stable) and how each model ranked that window.
City
Days back
What this shows — your logged day-before predictions vs the real recorded station high (Wunderground), the No bet at its logged price, and realised P&L — broken down by city and by agree/differ below.
Calibration — did the bets win as often as their price implied?
How this is calculated — from your settled log we bin every No bet by the price you paid and
compare it to how often those bets actually won. Realised edge = win‑rate − price; the Brier score measures
how well‑calibrated the pricing was (0 = perfect). Once bets logged with the models' own win% settle, a model Brier is added too.
Which forecaster is most accurate — by region or place
How this is calculated — for each forecaster we take its average miss in °C against the real recorded station highs in your log, grouped By region or By place (per city). Lower is better; 🏆 marks the most accurate. yr appears once the logger has recorded enough of its forecasts.
View
Data: Open-Meteo (GFS, ECMWF, ICON, GEM, Météo-France, UK Met Office) + yr/MET Norway + live Polymarket Gamma API.
Central estimate is the median across all sources (the 6 models and yr), so a single wild forecast can't skew the pick.
(JMA and CMA were dropped — backtest showed they lowered accuracy. yr's day-before effect isn't backtested — MET doesn't expose past forecasts freely.)
Cost = current price of the predicted bucket (e.g. 34¢ pays $1 if it hits → 2.9×).
US & European cities track the market closely; some hot, humid Asian stations can read a degree or two below the
recorded station high, so treat "differ" rows with extra caution.
The market resolves on one recorded station high — the prediction is a best estimate, not a guarantee. Informational only, not betting advice.