We trade the gap between weather science and market prices.

On venues like Polymarket, anyone can trade on tomorrow's high temperature in a city. Each degree band — say 89–90 °F — is its own contract that pays $1 if that band contains the day's official high and $0 if it does not. A full market is a row of these contracts covering every plausible outcome, and their prices, read together, are the crowd's probability distribution over tomorrow's weather.

It helps to see why a price is a probability here. A contract that pays exactly $1 when an outcome happens and $0 otherwise is worth, to a rational buyer, precisely the chance of that outcome. If the 89–90 band trades at 30 cents, the market is collectively saying "there's about a 30% chance the day's high lands in that band." Line up all the bands in a market and their prices should add to roughly one dollar — because exactly one band will contain the true high. So the row of prices is not a metaphor for the crowd's belief; it is the crowd's belief, written in cents. That is what makes these markets legible to a model: we can compute our own number for each band and hold it directly against the market's number.

Here is the quiet flaw in that distribution. The people trading these markets are, for the most part, casually uncertain. They know weather is hard to predict, so they hedge their bets — and they hedge too much. They spread their money across too many outcomes, buying the unlikely tails "just in case." The result is a market that systematically overprices unlikely outcomes and underprices the likely middle. It charges too much for the surprises and too little for the boring, probable answer.

The behavioral reason is familiar to anyone who has watched people bet. Being wrong on the boring middle feels like a small, forgettable loss; being caught flat-footed by a surprise feels like a large, memorable one. So people overweight the surprise. Across a whole market of such traders, that individual instinct becomes a structural shape: a distribution that is too flat and too fat in the tails. It is not that any one trader is irrational — it is that a crowd of mildly cautious people, added together, produces a curve that is measurably wider than the weather itself.

That 1.3× number is the whole thesis in a single figure. It means a genuinely well-calibrated forecaster — one whose stated probabilities match reality over the long run — can price each outcome more accurately than the crowd, and quietly trade the difference. We are not betting that we know tomorrow's weather better than the national weather service. We are betting that we can turn the same public science into sharper probabilities than a market of over-cautious humans does.

What does "1.3×" actually mean in feet on the ground? Suppose the true day-ahead uncertainty in a city's high is about ±2 °F — that is roughly how far reality tends to land from a good forecast a day out. A market pricing at 1.3× uncertainty behaves as if that spread were closer to ±2.6 °F. That extra six-tenths of a degree of imagined wobble does not sound like much, but it is exactly what pulls probability out of the likely center bands and sprinkles it onto the tails. Our job is not to shrink the real uncertainty — we cannot — but to price against the true ±2, while the crowd prices against an inflated ±2.6. The gap between those two curves is the whole opportunity.

An intuitive example

Imagine the forecast for a city tomorrow points clearly at a high of about 91 °F, with the normal day-ahead wobble of a degree or two. A calibrated model might say the 91–92 band is worth about 38 cents on the dollar — a 38% chance. But the market, nervous about being wrong, has spread its money out: it prices that same band at only 30 cents, and pads the far-off ≤86 and ≥97 tails with money they don't deserve.

Buying the 91–92 band at 30 cents when it is genuinely worth 38 is not a prediction that we will win this particular day. We might not — weather is weather. It is a purchase of a favorable price. Do it once and it's a coin flip with a good edge. Do it across hundreds of independent city-days, and the 8-cent gap between price and value is what shows up in the results.

1 — Forecast

We start where every serious forecaster starts: the public numerical weather models. We pool a ~122-member super-ensemble drawn from the three leading systems — the American GFS, the German ICON, and the European ECMWF. Each member is one plausible run of the atmosphere. Together they don't give us a single guess; they give us a distribution — a spread of possible highs for tomorrow, with the shape of that spread telling us how confident the physics actually is. A tight cluster means a near-certain day; a wide fan means genuine uncertainty. This is the raw material we price against.

2 — Calibrate

Raw ensembles are good but not honest enough to trade. They have two well-known flaws, and we correct both. First, every station has a persistent local bias — a given model may run consistently warm or cool at one specific airport because of terrain, coastline, or urban heat the model can't resolve. We bias-correct to the exact settlement station, not the city in general. Second, ensembles are famously under-dispersed: they act more confident than they should, packing their members too tightly. We widen the spread to match how far reality has actually landed from the ensemble mean, historically. The output is a probability for every temperature band that we would be willing to bet is true, not just plausible.

The bias correction is concrete, not hand-wavy. If, over the last several weeks, the raw ensemble has run 1.4 °F cool at a particular airport, we subtract that known offset before we price — so a raw ensemble mean of 89.6 becomes a corrected 91.0. This is learned per station and refreshed continuously, because the offset itself drifts with the season. It is unglamorous bookkeeping, and it is exactly the kind of detail a naive bot skips and quietly pays for.

Fixing under-dispersion is the more important half, because it is what directly attacks the crowd's mistake. An ensemble that says the high is "91, give or take half a degree" when reality gives or takes two degrees is overconfident in the opposite direction from the crowd — and if we traded it raw, we would be the ones overpaying the center. So we widen it to match reality. There is a clean way to check we did this right: over many past days, the true high should fall inside our stated 50% range about half the time, inside our 80% range about 80% of the time, and so on. When those hit rates line up with the labels, the distribution is honest. That check — comparing stated confidence to realized frequency — is the definition of calibration, and it is the single property this whole business rests on.

3 — Compare

Now we lay our calibrated probability next to the market's price for the same band. The difference — our probability minus the market's price, net of the spread and expected slippage — is the edge on that contract. If we think a band is worth 38 cents and it trades at 30, that's an eight-cent gross edge; subtract realistic execution costs, and what's left is the number we actually act on. Most bands will show no edge, or a negative one, and we simply pass on them.

4 — Trade

Where the edge clears our threshold after costs, we buy — and only there. Size is set by conviction and edge, never all-in, and every position is capped and diversified across cities so no single day, city, or weather system can dominate the book. The system is happiest making many small, favorable purchases and letting the arithmetic of a real edge do the compounding.

It is worth being blunt about how often step 4 actually fires: rarely. On a typical scan, most bands show no edge at all, and most of the ones that look like edges are illusions — a favorable "price" that is really a sliver of dust nobody would sell us in size, which our gates correctly reject (we return to exactly why in §7). Across the real, tradeable edges on a given day, only a small handful survive every filter. A day with zero trades is not a broken pipeline; it is the pipeline working as designed. The discipline is in passing at least as much as in buying.

The pipeline is easiest to trust when you watch it run once. Here is a single, illustrative city-day from the first calibrated probability to the settled result.

The setup. It's the afternoon before settlement for a city market. Our calibrated super-ensemble, bias-corrected to the exact airport this contract settles on, produces a distribution centered near 91 °F. Converted to bands, our probabilities read: 89–90 at 34%, 91–92 at 38%, 93–94 at 16%, and the rest scattered across the tails.

The comparison. We read the order book. The market is pricing 91–92 at 30 cents, 89–90 at 31 cents, and — tellingly — it's paying up for the unlikely ≥97 tail at 6 cents, well above our 2% read. On 91–92, our 38% against a 30-cent price is an eight-cent gross edge. After the spread and expected slippage, call it roughly five cents of net edge per contract. That clears our threshold.

The trade. Sizing runs through fractional Kelly against that five-cent edge, then gets clipped by the per-market cap and checked against how much same-region exposure the book already holds. Suppose that math yields a modest position of, say, 120 contracts at 30 cents — about $36 at risk on this single band. We buy, and only this band; we ignore the padded ≥97 tail rather than short it. One detail that decides whether this edge is real or imaginary: those 120 contracts have to be genuinely available at 30 cents. If the book only shows a handful of contracts at that price and the rest of the depth sits far worse, the true fill price is higher and the edge is smaller — which is precisely the kind of thing our gates check before a single simulated order goes out.

The position is held, not flipped. Having opened, the system doesn't churn. It carries the 120 contracts across subsequent scan cycles, re-checking each cycle that the thesis still holds, and simply waits for the market to settle rather than trying to trade in and out. This is a hold-to-resolution strategy: the edge is the price-versus-value gap at entry, and the way you collect it is to wait for the weather to decide.

The settlement. The next day, the official high at the exact settlement station rounds to 92 °F. The 91–92 contracts pay $1 each: 120 contracts return $120 against the $36 staked — a realized gain of $84 on this position. In our system this is not just a story: the position opens on the cleared edge, is held across cycles, settles on the real METAR observation for that station, and the realized profit or loss flows straight into the paper account's balance and equity. On this particular day, the trade won. The point of the example, though, is not the win — it's that we bought a 38-cent value for 30 cents. Some days that same disciplined purchase settles at zero, and the account debits the full stake. The edge is the price gap, repeated; the individual outcome is noise.

The hard part of this business isn't the forecast — good public forecasts are available to everyone. The hard part is resolution-correctness: knowing, to the letter, how a given market decides who won. Each contract settles on one specific airport weather station, reads temperature rounded a specific way, and draws the boundary of "the day" on a specific local clock. Get any one of those details wrong and every trade you place is subtly, permanently biased — you'll think you have an edge while quietly bleeding it away. Most naive bots make exactly this mistake, and it's invisible until the losses add up.

Our durable advantage is that we did this unglamorous work and verified it against reality. We reverse-engineered the exact settlement rule and checked it against 220 real, resolved market-days across New York, Los Angeles, and London — confirming the precise station each market reads, the round-half-up convention, the use of hourly-only observations, and the correct local timezone for the day boundary. When our reconstructed settlement matched the venue's actual settlement across all 220 days, we knew our forecasts were aimed at the right target.

A few of those details are worth spelling out, because each is a trap. The venue rounds the reported temperature rather than truncating it — so a reading of 91.5 becomes 92, not 91, and a bot that truncates will misclassify exactly the boundary days that matter most. It reads the observation in tenths of a degree Celsius and converts, which changes where the band edges actually fall. It uses hourly observations only, ignoring the special off-hour reports that a careless reader would pick up. And "the day" is drawn on the station's own local clock, so a hot reading just after local midnight belongs to the wrong day if you use the wrong timezone. Any one of these, gotten wrong, does not throw an obvious error — it just quietly aims every forecast a hair off-target, and you lose slowly while believing you have an edge. Verifying all of them across 220 independent days is what lets us trust that our probabilities are pointed at the same thing the market settles on.

Why this is durable when speed is not

There was once a latency game in these markets — a race to react to new observations a few seconds faster than the next bot. That window has largely decayed; it's not where the money is anymore, and chasing it is a treadmill. Calibration accuracy is different. It doesn't get arbitraged away by someone with a faster server, because it isn't about speed — it's about being right about probabilities in a market structurally inclined to be wrong about them. That correctness compounds quietly, and it doesn't evaporate the moment a competitor upgrades their hardware.

We don't compete on raw speed, but the system is still fast, disciplined, and always awake. Speed here buys us cleaner execution and fresher reads — not the edge itself.

The core is not a slow poll but a persistent WebSocket order-book reactor: an open connection to the venue that pushes every change to the book the instant it happens, so the system reacts to a moved price in milliseconds rather than waiting for the next scan to come around. The server sits at a Cloudflare edge roughly 2–3 ms from Polymarket, and once a book update arrives the decision itself — compare our calibrated probability to the new price, check the gates — resolves in well under a microsecond, because all the hard thinking already happened upstream in the forecast and calibration. Alongside the reactor, a five-minute sweep across ~80 curated stations refreshes forecasts and re-reads open markets; forecasts are cached behind a rate-limit circuit breaker so we stay fresh without hammering the upstream weather services.

Here is the honest framing of why we bother being fast, since we insist the edge is not speed. For weather, the useful information arrives in discrete lumps — a new model run, a fresh hourly observation — and the edge lives in the seconds-to-minutes after one of those lumps lands, not in nanoseconds. Being fast does two things and only two: it keeps us from being picked off (someone lifting our stale quote after news we haven't processed yet) and it lets us be first to act on genuinely new information before the price catches up. Neither of those is the source of the edge — the edge is calibration. Speed is what protects the edge from being taxed away by faster, dumber counterparties.

The station list is curated, not exhaustive. We only cover cities where we've verified the settlement rule and where the market has enough two-sided liquidity to be worth trading. Breadth matters — it's what turns a small per-trade edge into a diversified book, and with genuinely tradeable edges as scarce as the liquidity picture above implies, breadth is how we assemble enough independent bets to let the arithmetic work at all. But breadth without verified settlement is just noise, so we grow the coverage deliberately, one confirmed city at a time.

Being always-on matters for a strategy whose best opportunities cluster at specific moments — the narrow, roughly day-ahead window when real uncertainty and crowd over-pricing overlap, before the market collapses to near-certainty at settlement. A system that only wakes on a schedule would miss the moment a fresh forecast or observation opens that gap. Ours is watching every market, continuously, everywhere it trades, so that when a clean edge appears it is already looking — and so that when nothing clears, it correctly does nothing.

The liquidity reality — read this carefully

It would be easy to imply that these markets are full of easy, abundant edge. They are not, and we want to be precise about why. These weather markets are not empty — every market carries tens to hundreds of thousands of dollars of resting orders. But that depth is overwhelmingly one-sided: it is market-maker sell walls (asks) with thin bids underneath. And in the price range where quotes are genuinely real — call it roughly 10 to 92 cents — the market is efficiently priced. There is no free money sitting in the liquid, real part of the book; sophisticated market-makers have already done that arithmetic.

So where do the model's apparent "edges" come from? Mostly from the tails — bands priced at a fraction of a cent, where a wall of 0.1¢ dust sits on outcomes we think are slightly less unlikely than that. On paper that looks like a huge percentage edge. In reality it is untradeable: you cannot buy size at 0.1¢, and the dust is there precisely because no one serious wants the other side. Our minimum-price gate correctly rejects these. The net effect is stark and worth stating plainly: of perhaps ~18 raw "edges" the model flags on a representative pass, only about 2 clear the (correctly calibrated) gates. The rest are efficient pricing or dust.

Any single forecast can be wrong, and some will be. The entire system is built so that no single trade, city, or day can meaningfully hurt the book. Four controls do most of that work, and each exists for a specific failure it prevents.

Fractional Kelly sizing. The Kelly criterion tells you the mathematically growth-optimal bet size for a given edge — but full Kelly is famously volatile and unforgiving of a mis-estimated edge. So we bet a fixed fraction of it. The intuition is worth internalizing: full Kelly maximizes long-run growth only if your edge estimate is exactly right, and it punishes over-estimates brutally — a book that thinks its edge is twice what it really is can, at full Kelly, grind itself down even while being "right on average." Betting a fraction (a quarter, say) gives up a sliver of theoretical growth in exchange for a large cut in volatility and a wide margin of safety against our own estimation error. Since our edge estimates are inputs we could be wrong about, that trade is obviously worth making. Positions scale up with conviction and edge and shrink when either is thin, but they are never all-in. This keeps every bet small and repeatable, which is exactly what a small-edge, high-repetition strategy needs.

Hard caps. On top of sizing sit blunt, non-negotiable limits: a per-market cap so no single contract can balloon, a total-exposure cap on how much of the book is at risk at once, and a daily-loss limit that auto-halts trading if a day goes badly. These aren't clever — they're guardrails that don't depend on any model being right.

Correlation cap. The subtle danger in weather is that trades which look independent aren't. A heat wave that pushes ten cities above their bands at once is one bet, not ten — and a naive book would happily over-concentrate into it. We budget same-region and weather-system exposure as a single position, so a correlated event can't quietly become the whole portfolio.

Kill switch. Finally, a single control stops everything, instantly. If a data feed goes stale, a forecast source misbehaves, or the system's own behavior drifts outside expected bounds, the kill switch halts all trading rather than letting a malfunction compound. When something is wrong, doing nothing is the correct move — and the system is built to reach for that.

This is a small-edge, high-repetition business: modest expected value per trade, compounded across many diversified, largely-uncorrelated city-days. It is not a get-rich-quick machine, and we won't present it as one.

The arithmetic is deliberately unglamorous. A few cents of net edge on a contract is not much on its own — but a real edge, applied across a broad, diversified book of independent city-days, is how the number grows. Because the bets are largely uncorrelated (New York's weather has little to do with Los Angeles's on the same day), the ups and downs of individual trades tend to wash out, leaving the small persistent edge visible over time. Diversification isn't a nicety here; it is the mechanism.

A realistic illustrative sketch

To make the shape concrete — and to be explicit that this is a sketch, not a forecast or a promise — imagine a book placing on the order of a few dozen small, edged bets a day, each with a few cents of net edge after costs, most of them independent. Even a genuine edge like that produces losing days and losing weeks; the signal only separates from the noise over hundreds of trades. The realistic outcome of such a strategy, if the edge is real, is a modest, grinding, positive drift — not a chart that goes up and to the right in a straight line. Any month can be red.

Capacity is the honest ceiling

The binding constraint on this strategy is liquidity, not ideas. As §5 lays out, these markets are not empty — but the tradeable, two-sided depth is thin, concentrated in a narrow window before resolution, and much of the visible size is one-sided market-maker walls we can't lean on. Past a certain size, our own orders start walking up that book and moving the price against us, eating the very edge we're trying to capture — so there is a real ceiling on how much capital this can deploy before returns degrade. Because genuinely clearable edges are scarce (a couple per representative scan, not dozens), that ceiling is lower and more honest than a glance at total resting volume would suggest. More cities and more venues raise it over time, but it is finite, and we'd rather state it plainly: this is a strategy that works at a disciplined size, not one that scales indefinitely. Returns come from discipline and coverage, never from leverage.