Everything in the last three chapters — the forecast, the probability, the edge, the bet — depends on one thing being true: that everyone agrees, in the end, on what actually happened. That sounds obvious. It is not. It is the part almost every amateur gets wrong, and it is the quiet reason a careful bot beats a sloppy one.
Let's define the word we'll use all chapter. A market settles (or resolves) when the outcome becomes known and the winning side gets paid. In Chapter 2 we saw that a prediction market is just a bet with two sides: "yes" and "no." When the real answer arrives, one side is worth one dollar and the other is worth nothing. Settlement is the moment that verdict is handed down. There is no partial credit and no "close enough": the winning share pays exactly $1.00, the losing share $0.00, and everything turns on which side of a single line the final number falls.
Here is the trap. The question a market asks — "Will it hit 90°F in New York tomorrow?" — sounds like plain English anyone understands. But the rule that decides the answer is nothing like plain English. It is a precise, unforgiving little machine. And if your forecast is answering a slightly different question than the rule is asking, you can be right about the weather and still lose the bet.
You are not betting on the weather. You are betting on the exact number a specific rule produces. If your model predicts the weather but not that number, your edge is an illusion.
01 — Three details that decide everything
When a real weather market says "Will New York hit 90°F?", the fine print pins down three things that the headline hides. Get any one of them wrong and you are, in effect, answering a different question.
1. One specific station — not "the city"
"New York" is not a temperature. A city is a hundred square miles of parks, rivers, rooftops, and pavement, and the temperature is different in every one of them. So the market doesn't measure "New York." It measures one official weather station — almost always a specific airport — and ignores everywhere else on Earth.
A weather station is a fixed instrument that records temperature, and the airport one can read several degrees hotter or cooler than downtown, the suburbs, or the reading on your phone (which is usually a blend from nearby sources). The difference between the airport and midtown on a given afternoon can easily be the difference between "yes" and "no." So the very first job is to know which station this market resolves on — and forecast that exact spot, not "the city" in general.
Concretely: the New York market resolves on Central Park (code KNYC), while a Newark contract resolves on the airport (KEWR) a dozen miles away — and on a summer afternoon those two can differ by 3–4°F, the airport's runways baking while the park's lawns stay cool. If the line is "90°F" and your model forecasts the airport's 91 but the market settles on the park's 88, you were right about the region and wrong about the bet. Knowing the station is choosing which thermometer you are actually forecasting.
A weather market resolves on a single named station, usually an airport. Forecasting "the city" instead of that one thermometer is a small-sounding mistake that flips real bets.
2. A specific rounding rule — where does 89.6 go?
The instrument records something like 89.6°F. Does that count as "90"? It depends entirely on the rounding rule — the recipe for turning the raw measurement into the whole number the market cares about. Some rules round to the nearest whole degree (89.6 becomes 90 — a "yes"). Some chop off the decimal and round down (89.6 becomes 89 — a "no"). Same weather, opposite payout.
Worse, official readings are often reported in Celsius first and converted to Fahrenheit (the two temperature scales; water freezes at 0°C / 32°F). Rounding in Celsius and then converting gives a subtly different answer than rounding in Fahrenheit. These are the kinds of details that live in a footnote and quietly move money.
Here is a worked example of that Celsius trap. Suppose the station's underlying record is right on the edge at 32.4°C. Two honest-looking paths:
The lesson isn't the specific numbers — it's that "89.6 rounds to 90" is a claim you must verify, not assume, because the wrong recipe silently flips a bet you thought you'd won.
3. A local-day boundary — when does "tomorrow" start and end?
A market asks about the high temperature on a given day. But a day has to start and stop somewhere, and that "somewhere" is a local-day boundary — the local midnight-to-midnight window, in the airport's own time zone, that the reading is measured over.
This matters more than it sounds. Imagine the temperature peaks at 11:50 p.m. — does that peak belong to today or tomorrow? If you use the wrong time zone, or measure from midnight in your zone instead of the station's, you can grab a peak from the wrong day entirely. Heat near the edges of a day is exactly where forecasts and settlements disagree, and it is exactly where a careless bot leaks money.
A concrete way this bites: a bot on a London server reads times in UTC by default, and New York in summer is UTC−4. If that bot uses "the calendar day in UTC," its window ends at 8:00 p.m. New York time — four hours early. A 9:30 p.m. warm spell lands in the wrong bucket, and the forecast can be perfect while the answer is still wrong, purely because the day was defined in the wrong zone. The market measures over the station's own local midnight-to-midnight window; the only safe move is to compute in the station's local time, every time.
One more subtlety: the settled high is the maximum over the station's hourly observations within the local day — not the continuous "true peak" a forecast imagines. You forecast the hottest of the discrete hourly readings inside the window, on the right clock.
02 — Why small details mean big money
Recall the whole point of this project, from Chapter 5: the crowd tends to overpay for surprises — it prices uncertainty at roughly 1.3× what it should. Our edge is small and statistical, and it comes from our probabilities being well-calibrated — not from forecasting genius or speed. And calibration is measured against the exact number the settlement rule produces: mis-specify the rule and you're precisely tuned to a question nobody is paying out on.
An edge that small has no room for self-inflicted errors. If you are right on the weather 55% of the time but a settlement mistake silently flips 6% of your bets, you have handed your entire advantage back — and then some. A rounding rule you assumed but never checked isn't a rounding rule; it's a coin flip you didn't know you were making — invisible in the moment, discovered only at settlement, one lost bet at a time.
This is also why the near-the-line cases dominate the risk. A blowout — "exceed 90°F" on a day forecast for 102°F — settles the same under almost any rule. The bets where money is made and lost sit near the boundary, where 89.6 versus 90.4 decides everything — precisely where a rounding or time-zone slip does its damage. The settlement details matter most exactly where the edge lives.
Getting the settlement rules right adds nothing exciting to the story — it just means you were answering the real question. Getting them wrong doesn't cost you a little; it can quietly erase an edge that took a 122-member forecast ensemble to earn.
03 — The moat: verified on 220 real days
A moat (borrowed from the water around a castle) is business-speak for a durable advantage a competitor can't easily copy. Ours is not a secret formula. It is boring, unglamorous correctness: for each of roughly 80 curated stations we track, we've nailed down exactly which thermometer, exactly which rounding rule, and exactly which local-day boundary the market uses.
And we didn't just assume we got it right. We back-tested the settlement logic — replayed history to check it — against 220 real market-days: 220 days where we already knew both what the raw weather did and how the market actually resolved. Our reconstructed answer had to match the real one, day after day. That's the difference between "we think we know the rule" and "we've watched our rule reproduce reality 220 times."
The back-test is a tight loop: for each historical day we apply our candidate rule — right station, tenths-°C-to-Fahrenheit-then-round, the station's local-day window — and compare what we say should have resolved to how the market actually resolved. A single mismatch is a red flag that would eventually surface as a lost bet. Reproducing all 220 cleanly is how a hunch about a footnote becomes a rule we'll stake money on — the same discipline that lets the system score itself over time so calibration is a measured fact, not a hope.
The moat isn't cleverness — it's verified precision. Settlement rules confirmed against 220 real market-days is why we can trust a thin, honest edge instead of getting quietly robbed by our own assumptions.
This is why a naive bot loses to a careful one even when both have a good forecast. The naive bot forecasts "New York, around 90, tomorrow." The careful bot forecasts "the station, rounded this specific way, over the station's own local day" — and has receipts proving that's the right question. Same weather model, very different bank balance.
Settlement is also the moment the whole paper-trading loop closes. A position opens only when a real edge clears the gates, is held across cycles until the market's real settlement arrives, and at that moment the realized profit or loss flows straight into the shown balance and equity. That last step is settlement doing its job — and if the settlement logic were wrong, the paper P&L would be wrong in the same silent way a real account would be.
This is still paper trading — no real money, no track record. And getting the rule right does not manufacture opportunities. These markets are dominated by one-sided market-maker sell walls with thin bids; where prices are genuinely two-sided they're already efficiently priced. Most model "edges" are against sub-penny dust our price gates correctly refuse, so only a couple of real edges typically clear — and often the disciplined answer is zero trades. Correct settlement makes the few real trades honest; it can't conjure trades that aren't there.
04 — Where this fits
In Chapter 6 we turned a forecast into a bet. This chapter added the fine print that makes that bet honest: settlement is the exact rule that decides who's right. Next, in Chapter 8, we deal with the other half of survival — making sure that even when we're wrong (and we will be), we don't lose more than we can afford. A real edge only compounds if you're still in the game to collect it.
Predicting the weather is the flashy part. Predicting the number a footnote produces is the part that pays.