Who makes money on Polymarket weather markets: what 11,237 wallets show

680 resolved daily markets on the highest and lowest temperature in nine cities, from October 1 to November 9, 2026 (Singapore has highs only). They are split into 7,480 bucket markets, and we counted all 1.97 million fills: both the resting order (the maker) and the order that hit it (the taker). 11,237 wallets traded; taker volume was $13.8M.

City	Daily markets	Taker volume
Hong Kong	80	$2.30M
Seoul	80	$1.80M
Paris	80	$1.73M
Tokyo	80	$1.71M
London	80	$1.59M
Shanghai	80	$1.45M
New York	80	$1.34M
Singapore	40	$0.96M
Busan	80	$0.93M

Profit is measured to resolution: each fill against the price it settled at, 1 or 0.

Between makers and takers the trading is zero-sum before fees. The resting side lost $88.1k to resolution; takers won $88.1k. After an estimated $65.6k of taker fees, takers kept about $22.5k.

The same split shows up wallet by wallet, among the 1,717 wallets with at least $1k traded:

Wallet style	Wallets	Trading result
Mostly taker (under 20% of volume resting)	704	+$33.3k
Mixed (20–80% resting)	724	+$7.2k
Mostly maker (80% or more resting)	289	−$53.7k

The large market makers are paid by Polymarket's liquidity programme, not by trading. One of the biggest lost about $6.4k trading these markets while its wallet collected about $297k in liquidity rewards from October 1 to November 11, across all its markets. Another made $2.3k on trading and received $53k in maker rebates.

We split the six weeks at October 21 and ranked wallets by return in each half. The rank correlation between the halves was 0.20 across 904 wallets with at least 5 markets and $500 in each half, and 0.32 across the 431 with at least 10 markets and $2,000. Both are well clear of chance (p < 0.001).

The top tenth of first-half wallets made $105.7k in the first half and another $44.9k in the second; 68% of them were positive again. Of the first-half winners, 70% stayed positive; of the first-half losers, 45% turned positive.

That is not what we found in Elon Musk post-count markets. The same test on 602 wallets gave a correlation of 0.005, and the top tenth of first-half wallets made $134k, then lost $87k.

93 wallets passed a consistency screen: at least 15 markets, positive in both halves, at least $200 in total, and no single market more than 30% of the profit. Together they made $125.8k, 76% of it as takers. By timing:

• +$62.7k before the observation day began.
• +$63.9k during the day, on buckets the temperature could still reach.
• +$2.3k on buckets the day's temperature had already passed, which can no longer win.

The last group is the "speed" trade: when a report shows the high has passed a bucket, buy NO on it. Market-wide, takers made 2.4% on $258k of this trade in the first 2 minutes after the observation time, about nothing from 2 to 8 minutes, and lost money from 8 minutes on. Airport reports are published several minutes after their observation time, so whatever there is to win is gone before the report is out.

It also carries a settlement risk. On 599 of 600 city-days with airport data, the day's highest or lowest airport report fell in the winning bucket. The exception was Paris on October 7: a report read 19 °C, and the market settled at 18 °C. NO on the 18 °C bucket looked certain and lost. Over the six weeks this trade lost $17.7k in Paris, more than the $15.5k it made in the other seven cities combined.

Hong Kong settles on the Hong Kong Observatory, not an airport report, so it has no airport data here.

Two wallets traded in an identical style and never took the same position, so we count them as one operator. Together they made $26.3k; the larger wallet was positive on 29 of 31 trading days. About 98% of their volume was taken as a taker, all of it in high-temperature markets, held to resolution. The most profitable single wallet, $15.3k, traded in much the same way, with lumpier results.

This operator's picks won 43.8% of the time at an average price of 34.4¢ (505 picks in 242 markets); the single wallet's won 43.3% at 36.3¢. Of those 505 picks, 403 were placed one or two days before the observation day and the rest in the first hours of the day.

We compared their picks with 12 public forecast models, each taken as it was issued: a model run that started at least 8 hours before the pick, in 320 resolved high-temperature markets in eight cities. The best model, ICON, matched their bucket 32.7% of the time. The market favourite matched it 30.9% of the time, so no model explained the picks better than the price did.

On their own, 12 hours before the local day began, the models' buckets won 14–34% of the time. With a temperature correction fitted on the first half, the best (the average of eight models) reached 37.5% in the second. The market favourite won 53% and 59% in the two halves.

Public deterministic forecasts do not explain the winners, and the market price already contains them.

• The settlement source and its rounding decide the market. Know which station settles each city, and that a report and the settled value can still differ.
• On average, resting orders were the losing side of the trade.
• Speed only mattered inside the minutes before a report is published.
• The market favourite was hard to beat with public forecast models.

This describes what happened in these markets; it is not trading advice.