Trading Kalshi Temperature Markets Against a Forecast

How do Kalshi temperature markets actually work?

Kalshi lists daily high-temperature markets for major US cities. A market asks where the day's high will land, and the day is split into a ladder of buckets — a range like 84 to 85 degrees, an above-this-floor bucket at the top, a below-this-cap bucket at the bottom. Each bucket trades as its own contract. A share pays $1 if the settled high falls inside that bucket and $0 if it does not, and the price in cents is the market-implied probability of exactly that.

Settlement is not a matter of opinion. Each series resolves against a specific official climate report — the CLI report for a named station, which is usually an airport instrument rather than the city center. That distinction is the whole game in a coastal city. Downtown San Francisco and SFO can disagree by enough degrees on the same afternoon to flip which bucket settles at a dollar, and the market settles on the airport. Our finder forecasts the coordinates of the settlement station for exactly this reason.

Read the whole ladder together and it forms a probability distribution over the day's high. That is the structural feature that makes these markets different. In a political or cultural market the underlying probability is a judgment call and the crowd price may be the best estimate available. A temperature ladder is a distribution you can estimate independently and then compare, bucket by bucket, against what the market is charging.

What does the Weather Edge finder actually compute?

Be clear about the size of the machine, because the honest description is the useful one. The finder covers seven cities — New York, Las Vegas, Washington DC, San Francisco, Houston, Boston and Oklahoma City. For each one it pulls the open Kalshi markets in that city's high-temperature series, and separately pulls a free Open-Meteo forecast for the settlement station's coordinates in the city's own timezone.

It takes the deterministic daily-maximum forecast as the centre of a normal distribution, and gives that distribution a standard deviation that grows with lead time: about two degrees for today, widening by roughly one degree per additional day out, and it stops looking past a week because the approximation is too loose to say anything by then. It integrates that curve over each bucket, with a half-degree continuity correction at every boundary because settlement is a whole-degree reading, and that integral is the model probability.

The market side of the comparison is the YES price, taken from the last trade, or from the mid only when the book is quoting both a bid and an ask — half of a one-sided quote is not a price anyone traded, so those rows drop out along with anything at a penny or ninety-nine cents. Edge is simply model probability minus market price. Positive means the model reads the bucket as underpriced, a YES lean; negative means overpriced, a NO lean. The board sorts by absolute edge, refreshes on roughly a fifteen-minute cycle, and tells you how many cities actually answered so the count on screen is never larger than the data behind it.

This is a first-pass model and it is described that way on the page. It is one deterministic forecast with an assumed spread, not an ensemble, not a probabilistic product, and not a proprietary meteorological system. Where it is useful is in flagging buckets whose price is hard to reconcile with any reasonable distribution around the forecast. It is a place to look, not a conclusion.

How do you work the board step by step?

Set the minimum edge first. The board offers five, eight, twelve and twenty percent, and the default of eight is a reasonable place to start. Below five percent you are inside the model's own error bars and reading noise. Above twenty you will usually find either a very long-dated bucket or a bucket so thin that nothing has traded in it, which is a different problem.

Read the row before you read the number. Each row gives the city, the settlement date and how many days out it is, the bucket label as a link straight into the market on Kalshi, the forecast high the model used, the model probability, the market price, and the edge. The lead time matters more than almost anything else: a plus-six-day row is a wide, soft distribution being compared to a market nobody has priced carefully yet, while a today row is a tight distribution against a book that has had all morning to look at the same forecast you did.

Then go and look at the market itself. Open it on Kalshi and check the book. The largest edges on this board are routinely in buckets with no meaningful depth, and a twenty-point disagreement you cannot get filled on at anything like the quoted price is not an opportunity. Confirm the bucket boundaries against the market rules while you are there, because a boundary sitting on top of the forecast mean is the most fragile row on the entire board — one degree of observation error flips it completely.

Use the rest of the desk to pressure-test the ladder. The Overround Scanner runs the single-venue check on mutually-exclusive markets: it prices the full set of NO legs at their executable asks and flags when that set costs less than the amount it must pay, which is a structural trade that needs no forecast opinion at all. The Market Browser searches every market on Polymarket, Kalshi and Manifold if you want the neighbouring contracts. The Event Calendar shows the settlement clock. And a large Kalshi print in a series you are watching can be caught with the Kalshi block alert rule, set to a dollar size you choose, which fires as a desktop notification while a WhaleTracks tab is open.

When your side is chosen, the edge cell carries a link into the Kelly Bankroll Calculator that opens it prefilled for the correct contract: a YES lean sends the YES price and the model probability, a NO lean flips both together so the price and the probability describe the same thing you are buying. It arrives on the Kalshi fee schedule, and there is a maker chip for the case where you intend to rest an order rather than cross the spread.

What are the limits of a first-pass forecast model?

The spread is assumed, not measured. Real forecast uncertainty is not a smooth function of lead time — it collapses on a boring high-pressure day and blows out ahead of a front — and this model uses one growth rule for every city and every regime. That means it systematically overstates confidence on volatile days and understates it on placid ones, and the market will happily take the other side of both.

Uncertainty is also not constant within a day. A high-temperature forecast made at dawn carries far more uncertainty than the same forecast at noon, when much of the day's heating has already happened. The model's spread depends on which day it is forecasting, not on what time it is now, so a today row read in the late afternoon is more conservative than it needs to be.

Definition risk is the one that turns a good read into a losing trade. You must know which station settles the series, the observation window, and how rounding works at the boundaries. The finder forecasts the station rather than the city, which removes the largest version of this error, but it cannot remove the sensitivity of a boundary that sits on your forecast mean.

And the data has a shelf life. Forecasts update on their own schedule, the board caches for about fifteen minutes, and the market often reprices on a new run before anything on your screen changes. Treat a number on this board as a hypothesis about fair value with a timestamp attached, not a settled fact. Frequently the market is right and the model is the thing that is wrong.

What is the best way to size a weather edge?

Haircut the model before you size it. The calculator asks for the probability you believe, not the probability the model printed, and those should not be the same number. A model reading of fifty-five percent on a bucket deserves to be typed in as fifty or fifty-two until you have watched enough of these settle to know whether the model is calibrated. Sizing off an unhaircut model probability is sizing off the assumption that the spread rule is exactly right, which section four just spent four paragraphs arguing against.

Then use the sensitivity ladder, which is the single most useful control on the calculator for this strategy. It re-prices the same trade at a few points either side of your estimate. On weather buckets the ladder frequently shows the trade flipping to no trade with a two-point move, which is a precise, unflattering statement about how much model error the position can absorb. If it flips, the correct size is small or zero.

Watch the fees. The Kalshi fee is largest on contracts near fifty cents, which is exactly where the interesting buckets live, and it is charged inside the calculator before Kelly sees the edge — which is why a bucket that looks like a five-point disagreement can size at zero. Thin buckets add slippage on top of that. A model edge that does not survive fees and the spread is not an edge.

Respect correlation across cities. Five city-days driven by the same front is one position held five times, not a diversified book, and the calculator sizes each trade in isolation so that judgment is yours to make. Set the per-market cap before you open the board rather than after you see a tempting row, run the variance check to see what a string of these looks like at your chosen fraction, and only ever commit money you can afford to lose. Nothing here promises a profit, and no forecast is a settled temperature.

Is trading temperature markets actually worthwhile?

It can be a genuinely instructive corner of the market, because it is one of the few places where you can form an independent probability instead of arguing with the crowd about a judgment call. Whether it pays is a different question and not one this article will answer for you. Any figure you compute from past disagreements is hypothetical by construction, past performance does not guarantee future results, and a disagreement you find today may exist because several people who run better models than this one have already decided the market is right.

The failure modes are specific and worth naming. A large edge in a bucket with no depth. A boundary sitting on the forecast mean. A stale model reading against a market that already repriced on a new run. An unhaircut probability sized as if it were certain. And the quiet one: fees on a near-fifty-cent contract eating a five-point edge entirely.

Used well, the loop is repeatable and cheap to run. Price the ladder with the finder, screen at a sensible minimum edge, check the depth and the boundary on the venue, cross-check the ladder's internal consistency on the Overround Scanner, haircut the probability, size it in the calculator under a cap, and then watch how it settles.