Author: Changan I Biteye Content Team

Weather is not like elections, which have positions; it is not like the NBA, which has home teams. Yet it is this market that attracts domestic users. The reason is simple: everyone has feelings about it, and everyone thinks they understand Shanghai's weather.

But "feeling like you understand" and "being able to make money" are two different things.

Biteye shares three things today:

1. Understand the settlement rules

2. Establish a method for weather prediction

3. Use a system to find trading opportunities that others cannot see

I. First, clarify: how does this weather market settle?

1. The settlement temperature is not what you think

Many people have a misconception when participating for the first time: they compare the highest temperature forecast with a weather app on their phone, but the app shows the temperature for downtown Shanghai, while Polymarket settles using actual data from Shanghai Pudong Airport (ZSPD weather station). This data is publicly available through the American weather platform Wunderground, and PM directly reads the records on WU as the basis for settlement.

Two places, two numbers. Pudong Airport is located on the eastern side of the city, close to the Yangtze River estuary, and is affected by sea breezes, so the temperature is usually lower than in the downtown area. This difference may not be noticeable most of the time, but at the boundary of temperature ranges, it could mean the difference between winning and losing a bet.

So you can see this confusion in the comments section of the weather market: "It clearly feels warmer today than yesterday, why is the displayed highest temperature lower?"

2. The numbers are correct, but the units are not what you think

WU's data comes directly from the METAR reports submitted hourly by the airport (a globally used meteorological telegram format in civil aviation).

There is a detail hidden here: METAR records are in whole Fahrenheit degrees, and WU displays this number directly, without conversion or correction.

Most weather forecasting systems and meteorological models output temperatures with decimal points. The more refined your model is, the easier it is to overlook this coarsest detail.

3. Shanghai temperature patterns

After analyzing nearly 1900 days of data from the ZSPD station, it turns out that the occurrence of the highest temperatures in Shanghai is more concentrated than expected:

• All four seasons are highly concentrated between 11:00 AM and 1:00 PM,

• The concentration is highest at 12:00 PM in summer, with a single hour accounting for 27.6% of the entire season.

• The peak period in autumn is slightly earlier, with 10:00 AM also being one of the high-frequency periods.

Knowing the patterns is the first step, but the patterns won't monitor the market by themselves. When does the highest temperature appear each day, has it refreshed, and how far is it from the range?

So the editor built this system: before daily settlement, it predicts as accurately as possible which Celsius range the day's highest temperature will fall into.

II. Five methods, three successful

After clarifying the market rules, the next question is: how to predict the highest temperature of the day?

As a weather novice, the first step is to ask ChatGPT: how does the meteorological industry calculate the highest temperature of the day, and what mature methods are available? ChatGPT provided a theoretical framework, and Claude turned the framework into code. The two AIs worked together to set up the system over a weekend.

A total of five methods were tried, and ultimately only three were successful.

Successful methods:

1️⃣ WC + ECMWF integrated forecast

To predict the highest temperature, data is needed. Two sources were used:

• Weather Company (WC) is a commercial weather API that provides hourly forecast data with high accuracy;

• ECMWF is the European Centre for Medium-Range Weather Forecasts' global weather model, which is more sensitive to large-scale weather systems.

Each source has its strengths and weaknesses, so they were weighted for voting. The weights were dynamically adjusted based on the weather type of the day: trust WC more on sunny days, and trust ECMWF more on cloudy, windy weather.

2️⃣ Real-time correction: using warming data to estimate peak values

The forecast is calculated the night before, but today's weather is constantly changing. So this module's job is to use the actual data that has already occurred this morning to estimate how high today's temperature can go.

The logic is not complicated; the editor found that the fastest warming period in Shanghai is between 8-9 AM. After obtaining the actual temperature at this time, the system checks historical data: on average, how many more degrees could it rise at the same season and time in the past.

Then two corrections are added:

• If it's cloudy, multiply by a discount factor; the thicker the cloud cover, the more the warming is hindered.

• If it's windy, also multiply by a discount factor; strong winds accelerate heat loss. This results in an "extrapolated estimate."

Air pressure, dew point, and humidity were also included in the calculations, but after backtesting, it was found that these factors had a smaller impact and lower correlation, so they were removed.

However, relying solely on extrapolation is not stable enough, so a Kalman gain concept was used, which essentially takes a weighted average between the "extrapolated result" and the "original forecast," with the weight automatically changing over time.

• At 6 AM, extrapolation accounts for only 20%, with most still trusting the forecast.

• By 12 PM, extrapolation accounts for 72%.

• After 1 PM, it almost completely trusts the actual measurement, accounting for 85%.

The later it gets, the more important what is happening now becomes; the earlier it is, the more reference value historical forecasts have.

After 2 PM, the system determines that the peak value has likely passed and directly takes today's highest temperature from historical records to lock in the result, no longer extrapolating.

3️⃣ Is today a warming day?

This is the most satisfying module in the entire system, making a judgment every morning: will today's highest temperature be higher than yesterday's?

Every morning between 2-4 AM, the system collects a batch of meteorological data and feeds it to this model:

• Changes in air pressure over the past 3 hours and 12 hours

• Wind direction and speed, cloud conditions in the early morning

• Yesterday's temperature fluctuations, temperature trends over the past three days, whether yesterday's temperature was high or low

• Plus the month, season, what day of the year it is, and whether it rained yesterday

The model outputs five categories: warming day, slightly warming, stable, slightly cooling, cooling day, along with a confidence level.

However, this method has significant accuracy differences across different seasons.

• Most accurate in winter: when cold air arrives, air pressure rises sharply, and north winds strengthen, the signals are extremely clear, and the model can easily identify them.

• Least accurate in autumn: cold and warm air masses repeatedly tugging at each other, warming today and cooling tomorrow, historical patterns fail the fastest in this season.

Methods eliminated:

1. Fourier numerical prediction

Initially, an attempt was made to use Fourier analysis to fit the cyclical patterns of historical temperatures to see if it could directly predict the highest temperature of the day.

The result was that it could only tell you "what the average temperature is for this season historically." The randomness of Shanghai's weather is too high; Fourier fitting produces a smooth average curve, not the real daily fluctuations. The error was 3.6°C, and it systematically underestimated 100%, so it was eliminated.

2. ERA5 peak time prediction

ERA5 is a global historical reanalysis dataset from the European Climate Centre, used to predict when the highest temperature of the day will occur.

Backtesting results showed:

• ≤1 hour accuracy rate of 59.6%

• ≤2 hours accuracy rate of 81.3%

It sounds okay, but the problem is that PM's accuracy is higher, leaving a very short time window for traders to make judgments. If it cannot achieve peak judgments within half an hour, it is better to look at Polymarket's data, so this method was eliminated.

III. System practice: two cases and reflections on shortcomings

Polymarket's weather market opens trading four days in advance, and popular temperature ranges are usually fully priced early in the market. Buying directly at high-probability ranges has a poor risk-reward ratio.

So the strategy adopted by the editor is to wait for signals and enter the market during the warming time window.

Based on the self-built weather system, the following two operations were performed:

Case One:

In the early morning of the 16th, the Telegram channel pushed a report in night mode: tomorrow will be a cooling day. The reason is that the cloud cover is thick that night, and both seasonal and yearly characteristics point towards cooling.

At this point, the editor did not place a bet immediately. The early morning signal was only the first layer of reference.

By 11 AM, the system pushed a real-time report for the warming period. The actual highest temperature had already reached 12°C, and the +1°C probability score indicated that the chance of rising another degree today was 42%, leaning towards no further warming.

Combining this with the early morning logic regression's cooling signal, both modules pointed in the same direction, making the signal much clearer than in the early morning. So a bet was placed that the highest temperature on the 16th would not exceed 13°C.

Settlement for the day: 12°C. The previous day, the 15th, was 15°C, a drop of a full 3 degrees.

Case Two:

For example, today's weather in Shanghai on the 17th, the weather system can also serve as an early warning: the push received at 7 AM indicated an abnormal peak time: 10 PM.

Normally, the highest temperature on a sunny day occurs between 1-3 PM, but today the peak is at 10 PM, indicating that this is not warming from sunlight but rather warm, moist air being transported at night. It has been raining all day, with cloud cover at 97-100%, and almost no sunlight.

At this point, opening Polymarket, the pricing for 12°C was still at 53%. Some in the community were confused: it is already afternoon, and the temperature is only 11°C; the normal peak time has long passed, so why are people still buying 12°C?

Behind this confusion is that everyone is still using sunny day logic to judge the rainy day market.

The system is not confused. It identified the type of weather for today early in the morning, with an abnormal peak time, and there is a significant deviation between the current temperature and market expectations. This is an information gap, and an information gap is a trading opportunity.

This is precisely the significance of building this system: it makes it easier to identify opportunities; it provides quicker warnings in the face of risks.

What shortcomings does the system still have?

Building a system over a weekend inevitably has its flaws:

• The accuracy in autumn is only 63.7%, close to flipping a coin. Cold and warm air masses repeatedly tug at each other in this season; warming today and cooling tomorrow, historical patterns fail the fastest in autumn.

• Air pressure characteristics cannot be obtained in real-time trading. During model training, air pressure changes were used as features, and the backtesting results were good. The signals for cold air passing through are very clear. However, in real-time operation, the current interface cannot access real-time air pressure data.

• Coastal corrections are still waiting for data activation. The sea breeze effect at Pudong Airport is real, and the system has built a corresponding correction module, but the backtesting sample size is still insufficient.

Finding these issues in a system that has only run for a weekend is already a gain. Moving forward, it will be improved while running.

Conclusion

Meteorology has developed for hundreds of years, utilizing satellites, supercomputers, and global models, yet weather forecasts still cannot guarantee 100% accuracy for tomorrow. It is not that scientists are not working hard; the atmospheric system itself is chaotic, and a one-degree difference in initial conditions can lead to completely different results.

This system that has only run for a weekend will certainly make mistakes. The accuracy in autumn is close to flipping a coin, and if cold air arrives early, the system may not respond in time; the sea breeze effect has not yet been fully captured.

But that is not important. Making predictions in the market does not require being right every time; it only requires looking at one layer of information more than the market when the odds are advantageous.

The Shanghai weather market is still in its early stages, and the editor will continue to track this system, improving it while running. If you are also participating in the Polymarket weather market, feel free to discuss in the comments: what methods do you use to determine entry timing? What unexpected settlement results have you encountered?