Everything polyAether does rests on three plain ideas: what a probability means, what it means to be honest about your probabilities, and how you turn an honest probability into money. This chapter builds all three from zero.
01 — What a probability really means
A probability is just a number between 0 and 1 (or, if you prefer, between 0% and 100%) that measures how likely something is. Zero means "impossible." One means "certain." Everything interesting lives in between.
The cleanest way to feel what a probability is comes from a die. Roll a fair six-sided die. The probability of rolling a 4 is 1/6, or about 17%. What does that number actually promise? Not that your next roll will be a 4 — a single roll is a single roll. It promises something about the long run: if you roll the die ten thousand times, roughly one-sixth of those rolls will come up 4. The probability is the fraction you'd see if you could repeat the situation over and over.
Weather works the same way, even though you can't literally re-roll a day. When a forecast says "70% chance of rain tomorrow," it is making a long-run claim about a whole class of days: across all the days that looked like this one — same season, same pressure patterns, same everything the forecast can see — it rained on about 70% of them. Tomorrow is one draw from that pile. It will either rain or it won't; the 70% describes the pile, not the single day.
It helps to hold two questions apart, because people constantly mix them up. The first is: what will happen tomorrow? — a single, sharp answer we don't get to know in advance. The second is: how likely is rain, given what we can see today? — which has a number for an answer, and is the only one a forecast can honestly speak to. A probability isn't a hedge or a shrug; it's a precise statement about a pile of similar situations, of which tomorrow is one anonymous member.
The markets polyAether trades sharpen this even further. A typical contract isn't "rain, yes or no" but something like "Will today's high at Chicago's official station land in the 78–79 °F bucket?" Nature will settle on exactly one bucket. Before it does, our job is to spread 100% of belief across all the buckets — maybe 6% on 76–77, 22% on 78–79, 31% on 80–81, and so on — so the numbers add to one and each is an honest long-run frequency. The forecast isn't a single guess; it's a full distribution of belief over every outcome that could occur.
A probability is a long-run frequency. "70%" doesn't predict tomorrow — it says that among all the days like this one, about 7 in 10 turn out that way. A single outcome can never prove a probability right or wrong; only many outcomes can.
Why one outcome tells you almost nothing
This is the trap that fools almost everyone. Suppose someone says "80% chance of sun" and it rains. Were they wrong? No — not from that one day. An 80% forecast is supposed to be wrong about 1 day in 5. If it were never wrong, it wasn't really 80%, it was more like 100%. Rare things happening at their promised rate is exactly what a good forecast looks like. To judge a probability, you have to zoom out and count many days.
Flip it around to feel how sharp the trap is. Imagine a forecaster who says "90% chance of sun" on 100 sunny-looking mornings across a year, and it turns out sunny on 91 of them. That's a superb forecaster — reality (91%) landed almost exactly on the promise (90%) — even though on 9 of those days it wasn't sunny. Each of those 9 misses wasn't a failure; it was the forecast working as designed. The single most common mistake people make about probability is grading the forecast by one outcome instead of by the long-run rate. The outcome is one noisy sample. The rate is the truth. Which brings us to the actual test.
02 — Calibration: the test of an honest probability
So if a single day can't grade a forecaster, what can? The answer is calibration. A forecaster is well-calibrated when their stated probabilities match reality in the long run: of all the times they say "70%," the thing actually happens about 70% of the time. Of all the times they say "30%," it happens about 30% of the time. And so on, up and down the whole range.
Here's the honest way to picture it. Collect every day a forecaster said "70% chance of rain." Maybe that's 200 days. Now count: on how many did it actually rain? If it's around 140 (which is 70% of 200), they're calibrated at that level. If it rained on only 100 of them, they were overconfident — they said 70% but reality was 50%. If it rained on 180, they were underconfident. Calibration is simply: do your numbers mean what they say?
A worked calibration check. Say we run the model for a year and, whenever it lands on "70%," we drop that day into a bucket. Over the year the "70%" bucket collects 200 days. We wait for each day to resolve, then tally the yeses:
- 139 of 200 came true → observed rate 69.5%, promised 70%. Essentially perfect — the 0.5-point gap is well within the noise you'd expect from only 200 samples.
- 108 of 200 came true → observed rate 54%. We said 70% but reality was 54%: badly overconfident. Every bet we placed on the strength of that "70%" was buying a 54-cent thing while believing it was worth 70.
- 151 of 200 came true → observed rate 75.5%. Mildly underconfident — the truth is a touch better than we claimed, so we were leaving a little edge on the table but never in danger.
Repeat that tally for the "10%," "20%," … "90%" buckets and you have a full calibration audit. The whole point: a probability isn't checked by whether one call came true — it's checked by whether a bucketful of identical calls came true at the promised rate.
Notice what calibration is not. It is not about being right on any single day, and it is not about being bold. A forecaster who simply says "50%" about everything can look vaguely reasonable but is useless — they're not telling you anything specific. Good forecasting needs two things together: probabilities that are sharp (close to 0 or 100 when the situation warrants) and calibrated (meaning what they say). polyAether cares about calibration above all, because a probability you can't trust is worse than no probability at all.
How we actually score it: Brier and PIT
Counting buckets by hand is the intuition; in practice polyAether keeps two running scores that measure the same honesty more precisely, and it does so continuously as real outcomes arrive.
The first is the Brier score. Each time a market resolves, we take the probability we assigned to what actually happened, subtract it from 1, and square the miss. If we said 70% and it came true, the penalty is (1 − 0.70)² = 0.09. If we said 70% and it didn't, the penalty is (0 − 0.70)² = 0.49 — much larger, because we were confident and wrong. Average that penalty over thousands of resolved markets and you get one number where lower is better. The squaring is the important part: it punishes confident mistakes far harder than timid ones, so the only way to earn a good Brier score over the long run is to be both sharp and honest. You cannot game it by hedging to 50%, and you cannot game it by being reckless.
The second is the PIT check — the probability integral transform, which is the distribution-shaped version of the bucket audit above. For every resolved market we ask: where did the real outcome fall inside the full distribution we predicted? If our distributions are honest, those positions should spread out perfectly evenly — the truth should land in our "bottom 10%" region 10% of the time, our "top 10%" region 10% of the time, and so on. If instead the truth keeps landing in our tails more often than it should, that's a fingerprint of overconfidence: our distributions are too narrow, too sure of themselves. PIT is how we catch a distribution that's been squeezed too tight, long before it costs us money.
Both scores accumulate. A single market tells us almost nothing — the same reason one rainy day can't grade a "70%" forecast — but over hundreds and then thousands of resolutions they converge on the truth about whether our numbers deserve trust. We don't get to declare ourselves calibrated; we earn the claim one resolved market at a time.
Well-calibrated means your numbers are honest: when you say 70%, it happens about 70% of the time — checked across hundreds of cases, not one. Calibration is the only thing that makes a probability safe to bet on.
03 — Edge: our probability minus the market price
Now we connect the forecast to the money. Recall from Chapter 2 that in a prediction market, a contract that pays $1 if some event happens trades at a price that behaves like a probability. If "Rain in Chicago today?" is priced at 55 cents, the market is effectively saying "55% chance." Buy it for 55 cents; if it rains you collect $1, if it doesn't you get nothing.
So now we have two probabilities sitting side by side for the very same event:
- The market price — the crowd's probability, in cents.
- Our model probability — what our weather ensemble (the ~122-member GFS + ICON + ECMWF forecast from Chapter 3) says.
The gap between them is the whole game. We call it the edge:
Edge = our probability − the market price
Suppose the market is charging 55 cents for "Rain today," but our calibrated ensemble puts the true chance at 70%. We think a $1 payout is worth about 70 cents, and we can buy it for 55. That 15-cent gap is our edge — the amount we believe the market has mispriced the contract in our favor. If our probability is honest, buying at 55 is buying a 70-cent thing at a discount, over and over.
Following the money on one edge. We pay 55 cents for a contract our model values at 70. What do we expect to earn per contract, on average, if the 70% is honest?
70% of the time it rains. The contract pays $1.00; we paid $0.55, so we net +$0.45.
30% of the time it doesn't. The contract pays nothing; we lose the $0.55 we staked, netting −$0.55.
Blend them by their odds. Expected profit = 0.70 × (+$0.45) + 0.30 × (−$0.55) = $0.315 − $0.165 = +$0.15 per contract.
That +15 cents of expected value per dollar-contract is the edge made concrete — and notice it's exactly our probability (70) minus the price (55). We will still lose outright on 3 of every 10 of these bets. Losing individual trades isn't the model being wrong; it's the 30% doing exactly what it promised. The edge only shows up as money after the wins and losses pile up in their honest proportions.
Edge can point the other way too. If the market charges 80 cents but we think the real chance is only 60%, there's no edge in buying — the contract is overpriced by our lights, and we pass (or, where possible, take the other side). And when our number and the market's number roughly agree, edge is near zero and there's simply nothing to do. No edge, no trade. Most days, most markets, that's the correct answer.
An edge on paper is not an edge you can trade. A gap between our number and a quoted price is only the start. Before it becomes a real opportunity it has to survive several honest filters, and most gaps don't:
- Is there a real price on the other side? These weather markets are not empty — every one carries tens to hundreds of thousands of dollars in resting asks. But that depth is overwhelmingly one-sided: thick market-maker sell walls sitting above thin bids. A lot of the model's apparent edges are against a fraction of a cent of "dust" resting on unlikely tail buckets. A minimum-price gate correctly throws those out — you cannot meaningfully profit from buying a 0.1-cent lottery ticket, however favorable the ratio looks.
- Is the price even contestable? Where prices are genuinely alive — roughly the 10-to-92-cent band — the market is usually priced efficiently, and our number and theirs agree. Efficient agreement is not a bug; it's most of the market, most of the time.
- Is the window open? The genuinely two-sided, tradeable moment is only about a day before resolution. As resolution approaches the outcome collapses toward certainty and the market prices in step, leaving nothing to capture.
Put those together and the honest expectation is few trades. Out of a couple dozen raw model edges on a given day, it's common for only a small handful — sometimes zero — to clear the gates. Zero trades is frequently the correct, disciplined outcome, not a failure. The gates exist precisely so a fantasy edge against dust never becomes a real loss.
Edge = our probability minus the market price. It's the discount (or premium) between what we think a contract is worth and what it costs. Positive edge is the only reason to place a bet; when the two numbers agree, we sit still.
Why edge only works if we're calibrated
Here is the thread that ties the chapter together. Edge is measured against our probability — so if our probability is dishonest, our edge is a fantasy. Imagine we're chronically overconfident: we say 70% when reality is only 55%. We'll "see" a fat 15-cent edge against a 55-cent market that is, in fact, perfectly priced. We'd bet confidently into markets where we have no advantage at all, and slowly bleed money while feeling clever.
That's why calibration isn't an academic nicety — it's the safety catch on the whole machine. An edge is only real if the probability behind it is trustworthy. This is exactly why polyAether validates its numbers against real outcomes before risking a cent, and why it's still strictly paper-trading with no proven track record. A believable edge starts with a believable probability.
What "paper" now means in practice is worth stating plainly, because the full loop does run end to end — just with imaginary money. When an edge clears the gates, a position opens at the real quoted price, is held across cycles as the day develops, and settles on the actual official observation when the market resolves; the resulting win or loss flows straight back into a running balance and equity figure. So the mechanism that would move real money is exercised in full — what's absent is the real money and, therefore, any track record. A working machine is not the same as a proven one.
One more honest framing, because people reach for it as an explanation and it's the wrong one. Our edge is not forecasting genius, and it is not raw speed. The system is fast — it reacts to order-book changes in milliseconds and sits a few milliseconds from the exchange — but for weather the useful information changes on the scale of seconds to minutes (a fresh forecast run lands; a new hourly observation prints), not nanoseconds. Speed's job here is defensive: to keep us from being picked off on a stale quote and to be first to react when genuinely new information arrives. The edge itself comes from calibration — from carrying probabilities that are honest against a crowd that reliably overprices uncertainty.
In the next chapter we'll turn to the specific, repeatable reason these gaps exist at all: the crowd systematically overpays for surprises — pricing uncertain, dramatic outcomes at roughly 1.3× their true odds. That predictable mistake is where our edge comes from.