Not losing money: sizing and risk

In earlier chapters we built an edge — a reason to believe a market's price is slightly wrong. This chapter is about the other half of the job, the half that actually keeps you in the game: deciding how much to bet, and making sure a run of bad luck can never wipe you out.

Here is a hard truth that surprises newcomers. You can be right on average and still go broke. Being right about the odds is necessary, but it is not enough. If you bet too much on any single wager, one unlucky outcome can take away money you can never earn back — because you have less left to grow from. The math of losses is unforgiving: lose half your money and you now need to double what remains just to get back to where you started. So the first rule of the whole enterprise is not "find great bets." It is "survive long enough for the great bets to pay off."

Sit with that recovery math for a moment, because it is the whole reason this chapter exists. A 10% loss needs only an 11% gain to undo. A 25% loss needs a 33% gain. A 50% loss needs a 100% gain. A 90% loss needs a 900% gain — a tenfold return — just to break even. The damage from losses does not grow in a straight line; it accelerates. That asymmetry is why a strategy that "wins on average" can still march quietly to zero if it is sized carelessly: the big losses dig holes that the ordinary wins can never climb out of. Sizing is the discipline of never digging a hole that deep in the first place.

The two levers you control

When you place a bet you control two things. First, which bet — that is the edge work from Chapters 4 through 6. Second, how big — how much of your money, called your bankroll (the total pot you are trading with), you put on this one wager. People obsess over the first lever and ignore the second. That is backwards. Given a fixed edge, your bet size determines almost everything about whether you thrive or blow up.

Think of it like poker, or like farming. A farmer with a great crop plan still plants many fields, not one, because a single hailstorm shouldn't end the farm. The plan is the edge; the many small fields are the sizing.

To make this concrete, imagine two traders who share the exact same edge — both find bets that, on average, pay 5 cents of profit for every dollar risked. The first, call her Ada, risks 40% of her bankroll on each bet. The second, call him Boris, risks 3%. They see the identical stream of opportunities and the identical run of luck. Over hundreds of bets Boris's money climbs a gentle, bumpy staircase upward. Ada's money lurches: a couple of good runs make her briefly rich, but the first cluster of losses — and clusters always come — carves 60%, 70% off her pile, and now she needs enormous gains just to recover. Same edge, same luck, opposite fate. The only difference was the size of the bet. This is not a parable; it is arithmetic, and it is the single most important thing in this chapter.

Bet more when the edge is bigger — but never all-in

There is a famous formula, the Kelly criterion (named after a scientist, John Kelly), that answers one clean question: if you knew your exact edge, what single bet size grows your money the fastest over the long run? The intuition is beautifully simple and worth remembering even if you never touch the equation:

• Bigger edge → bigger bet. When the price is very wrong in your favor, lean in more.
• Smaller edge → smaller bet. When you barely have an advantage, risk very little.
• No edge → no bet. If the price looks fair, sit out.

So far so sensible. But "full Kelly" — betting the exact amount the formula recommends — is violently aggressive in practice. It assumes you know your edge perfectly, which you never do. Your edge is itself an estimate, and estimates are wrong sometimes. If you overestimate your edge and bet full Kelly, the swings up and down are stomach-churning and a bad patch can gut your bankroll.

The fix is fractional Kelly: compute the full-Kelly size, then bet only a fraction of it — say a quarter or a half. You give up a little long-run growth in exchange for far smaller, far safer swings. It is the difference between driving fast and driving fast with a seatbelt and inside the speed limit. polyAether uses a fraction of Kelly, deliberately, precisely because our edge is an estimate and humility is cheaper than ruin.

Let's actually turn the crank once, because the mechanism is less mysterious than the name suggests. Suppose polyAether's forecast says a market's "yes" outcome has a 62% chance of happening, and the market is selling that "yes" share for 50 cents. If we're right, a 50-cent share is really worth 62 cents — a genuine edge. The Kelly formula for this kind of bet asks a simple question: how much of the mispricing can I claim relative to how much I could lose? For a share that pays $1 if it wins and $0 if it loses, full Kelly works out to roughly the edge divided by the odds against you — here about 24% of the bankroll. That is a terrifyingly large slice for a single weather bet. Now apply a one-quarter fraction: 24% becomes about 6%. That 6% is a sane, survivable stake — and if our 62% estimate is even a little optimistic, betting a quarter of full Kelly means the mistake costs us a scrape, not a limb.

Notice what the fraction is really buying. Full Kelly is the theoretical growth-maximizing bet only if your probability is exactly right. Ours never is — it is the output of a forecast model, well-calibrated but not clairvoyant (that calibration story is the heart of Chapters 4–6). Shrinking to a quarter or a half of Kelly is the mathematical way of saying "I trust my edge, but not that much." The growth you sacrifice is small; the protection against a wrong estimate is enormous. That trade — a little less upside for a lot less ruin — is one polyAether takes on every single bet.

Many small, unrelated bets beat one big one

Even a well-sized single bet is a single bet: it wins or it loses. The way to smooth out luck is diversification — spreading your money across many bets whose outcomes don't move together. The magic word there is uncorrelated. Two bets are correlated when they tend to win or lose together; they are uncorrelated when the outcome of one tells you nothing about the other.

Here is why it matters. Betting on tomorrow's high temperature in Chicago and, separately, on tomorrow's high in Miami are two mostly-unrelated bets — different weather systems, different days. Ten such independent bets, each small, will rarely all lose at once; the good and bad cancel out and your bankroll drifts up smoothly. But betting on "will it be hot in Chicago" and "will it be hot in a suburb 20 miles from Chicago" is really one bet wearing two coats. If you're wrong, you're wrong on both. That hidden sameness is the danger.

So polyAether doesn't just cap how much goes into any one market. It also watches for correlation across positions — a correlation cap — so it never accidentally stacks a pile of bets that are secretly the same bet. Our roughly 80 curated weather stations across many regions exist partly for this reason: they give us genuinely different bets to spread across.

The way to feel the power of uncorrelated bets is to watch the swings shrink. One bet that risks $30 and either wins or loses is a coin-flip-sized lurch in your bankroll. But split that same $30 across ten independent bets of $3 each, and for all ten to lose at once is a wildly unlikely event — like flipping ten tails in a row. The typical day is now a handful of wins and a handful of losses that mostly cancel, leaving your small average edge to peek through. You didn't change your edge one bit; you changed how bumpy the ride is to collect it. Smoother rides are survivable rides, and survivable rides are the ones that get to compound.

There is an honest caveat worth stating plainly, because it matters more than the theory suggests. These weather markets are thin. In practice only a small handful of edges clear polyAether's gates on any given cycle — often zero. So the diversification we get is real but modest: we are usually spreading across a few genuinely different bets, not dozens. That is a feature of discipline, not a bug — we would rather hold a few well-separated positions than manufacture correlated ones just to look busy. Chapters 9 and 10 return to this "few trades is the correct outcome" reality in detail.

The guardrails that don't depend on being smart

Formulas can be misconfigured. Estimates can be wrong. So on top of the clever sizing math, polyAether enforces hard caps — dumb, absolute limits that hold no matter what the formula says. Think of them as guardrails on a mountain road: you hope never to touch them, but you are very glad they exist.

A per-market cap means no single market — however juicy it looks — can ever hold more than a set slice of the bankroll. A total cap means the machine keeps a reserve and is never fully committed. And the daily-loss halt is the most important of all: if losses in a single day cross a preset line, the system stops trading for the day, full stop. This is the circuit breaker that turns a bad day into merely a bad day instead of a catastrophe. Bad days happen even when everything is working; the halt makes sure they stay survivable.

Here is how the caps interact in practice, so you can see they are not redundant. Suppose fractional Kelly, looking at an unusually large edge, recommends putting 9% of the bankroll into one market. The per-market cap is set below that — say 5% — so the machine quietly trims the bet down to 5% and moves on. Nothing dramatic happens; the guardrail simply refuses to let a single confident-looking number dominate the book. Now suppose the machine already holds several positions and their total reaches the total cap. Even if a fresh, beautiful edge appears, the machine declines to add it, because being fully committed with no reserve is exactly the state that turns one bad settlement into a scramble. The clever Kelly math proposes; the dumb caps dispose. That ordering — smart sizing first, then hard ceilings that override it — is the belt-and-suspenders design at the core of the whole system.

One button that stops everything

Beyond the automatic halts, there is a kill switch: a single manual command that instantly stops all trading and prevents any new positions. Caps and halts are the automatic brakes; the kill switch is the human hand yanking the emergency cord. If anything looks wrong — a suspicious pattern, a settlement question from Chapter 7, an outage, or just a bad feeling — we can stop the whole machine in one motion and figure it out with nothing else at risk. Having an off switch you trust is what makes it psychologically possible to let an automated system run at all.

The distinction between the daily halt and the kill switch is worth making sharp. The halt is a rule: it fires on a number crossing a line, without judgment, and it resets tomorrow. The kill switch is a decision: it fires on human judgment, for reasons a rule could never anticipate — a news event that makes a whole category of weather markets untrustworthy, a suspicion that a data feed is stale, a hunch that something is off before any threshold has been breached. Rules are fast and tireless but literal; humans are slow and fallible but can smell trouble a formula would miss. You want both, and you want them independent, so that the failure of one does not disable the other.

Discipline is the whole game

Notice what this chapter did not promise: bigger wins, cleverer bets, a secret formula for riches. Everything here is about not losing — sizing modestly, spreading out, capping exposure, halting on bad days, and keeping a hand on the kill switch. That is not a footnote to the strategy. It is the strategy's foundation. A pile of edges with no risk discipline is a story about how someone went broke. The same edges with disciplined sizing is a business.

It is worth being blunt about where these controls sit today, because it changes how you should read them. In the paper simulation, the full loop now runs end to end: a position opens when an edge clears the gates, it is held across cycles, it settles against the real weather observation, and the realized profit or loss flows back into the shown balance — so the sizing and caps described here are not diagrams, they are code deciding real (simulated) stakes on every cycle. And because the edge itself is calibration — well-calibrated probabilities against a crowd that tends to overprice uncertainty — the risk controls are what let a modest, steady statistical advantage survive the noise long enough to show up as compounding. Speed keeps us from being picked off; calibration is the edge; sizing is what turns that edge into something that lasts. None of the three works without the others.

A necessary reminder from the whole course: polyAether is strictly on paper right now — simulated trades, no real money at stake — and although the paper profit-and-loss lifecycle now works end to end (positions open, hold, settle on the real observation, and feed realized P&L back into the balance), there is still no proven track record. A working simulation is not a validated strategy. These controls are designed and being exercised, not battle-tested with live capital, and the honest expectation is few trades — often zero on a given cycle — because the disciplined thing to do most of the time is nothing. The point of building all of this before going live is precisely so that survival is engineered in from the start, not bolted on after the first painful day.