Where the Money Comes From on a Prediction Market — Five Families of Strategies and the Price of Entry to Each

Where the Money Comes From on a Prediction Market — Five Families of Strategies and the Price of Entry to Each

A map of strategies grouped by source of income, not by tactic — and an honest look at what’s hard in each

A new article in the cycle. The last one ended with a conclusion: a directional edge — guessing whether the price goes up or down — is rare and hard, and without it a bot just pays the costs. Hence the natural question: “so you can’t make money?” You can — but directional betting is only one of several families, and far from the most popular among bots. This article is a map of five families, grouped by where the money in them comes from, and an honest look at the difficulty of each. There still won’t be a ready strategy: each family has its own price of entry, and you have to pick the one within your reach.

---

Where this topic even comes from

Leaf through what bots actually run on these markets and you’ll find a dozen names: momentum, paired arbitrage, mean reversion, market-making, latency, delta-neutral, model-based, and so on. The list is intimidating, but there’s less substance in it than it seems: most of these tactics are a few sources of income in different wrappers. So taking them apart by name is pointless; you have to take them apart by one question — where does the money in the strategy come from.

There aren’t many answers to that question. You can earn by: guessing the direction; estimating the fair probability more accurately than the market with a model; catching a price inconsistency; serving those who trade, for a fee; or extracting a price skew while hedging away the directional risk. Those are five different sources, five different skills, and five different difficulties. None is “easy” — each is just hard in its own way. Let’s take all five apart: where the money is and what exactly is hard.

One caveat that runs through all five, keep it in mind as you read. The entry price decides everything. Buying the “obvious” winner at 85–90 cents is a loss even at high directional accuracy: to come out ahead at a price of 0.85 you’d have to guess right more than 85% of the time — which almost never happens. There was a separate article about this at the start of the cycle (the breakeven win rate equals the entry price), and it runs through every family below.

---

The five families at a glance

Now each one, with numbers.

---

Family 1. Directional / momentum — the most intuitive and the hardest

This is what everyone pictures as “trading”: you think BTC will rise, you buy “up,” you guess right, you earn. The money here comes from prediction: you estimated the probability of the outcome more accurately than the market. A typical tactic — skip the first minutes of the window (noise and stop-hunting there), enter only if the direction has confirmed and the contract price is still lagging the real move, exit on a take-profit or hold to resolution.

The previous article showed why this is hard: to earn, your estimate has to differ from the price by more than the costs, and such divergences are rare. Without a real edge, every trade brings a loss of the costs — about three cents round trip, steadily, just for the right to play. To come out ahead, the edge has to cover those three cents.

The takeaway for the family: directional betting is a bet that you predict better than the market. Sometimes true, but rarely, and it has to be checked hard. It’s the hardest way, because it demands the scarcest thing — a real predictive advantage. On short windows (5–15 minutes) it’s especially hard: there speed and microstructure rule, not a “pretty” prediction; on longer ones (an hour and up) there’s a bit more room for a directional edge.

---

Family 2. Quant / model-based — the same prediction, but by discipline

This is the “grown-up” version of the first family. The money is the same — from your probability estimate being more accurate than the price — but you get it not from intuition but from a model of the fair probability. The key difference: edge = model probability − contract price, and you enter only when that edge exceeds a threshold that accounts for costs (usually 5–10 points). This is exactly our “frequency-price gap beyond costs” from the article on finding an edge, only the probability is computed by a formula.

Where the model’s probability comes from. For short windows there’s a simple, honest formula from stochastic calculus: the probability of being above the strike by the window’s close is

P(up) = Φ( deviation / remaining volatility )

Let’s take it apart so it isn’t a black box. “Deviation” is how far the price has moved from the strike (the price at the window’s open) by now. “Remaining volatility” is how far the price can still travel in the time left to close. Φ is the normal-distribution function, which turns their ratio into a probability. The intuition is plain: the further the price has already moved up and the less time is left to reverse it, the higher the probability of closing at the top.

Here’s how it works over a 15-minute window in numbers (volatility ~0.4% over the window):

It reads logically: at the very start, at the strike, it’s all 50/50. The price moved 0.2% up by the middle of the window — the model already gives 76%. It moved 0.4% and a quarter of the time is left — nearly decided, 98%. Now compare with the contract price: if the model says 76% and the contract trades at 55 cents, your edge is 21 points, and (if it survives the costs and the check) that’s an entry.

A more powerful source of the probability is the options market: from the implied-volatility surface on Deribit (via SSVI calibration and the Breeden-Litzenberger formula) one extracts the distribution of the future price and computes the probability more accurately than a simple GBM. But two honest caveats at once. The first, practical: Deribit data is an external source, outside what gogobots gives you (Binance spot and the Polymarket market); it’s something you connect yourself, not a built-in feature. The second, in principle: the probability from options is a risk-neutral estimate, just like the contract price. We discussed this in the article on finding an edge: a gap between two risk-neutral estimates may be not an error but a difference in risk premiums between two markets. So an “edge against Deribit” has to be checked just as hard as any other, not treated as a holy grail.

The takeaway for the family: quant is the most systematic and scalable way, and it works better on longer windows (an hour and up), where the model has time to say something meaningful. But its difficulty is in two things. First: the model has to be calibrated to Polymarket’s resolution (the oracle, often Chainlink), not just to the spot, or your probability is about one market while you’re paid by another. Second: a model, especially a machine-learning one, is easy to overfit — and here the whole arsenal of the cycle (the out-of-sample fan, the combinatorial check, counting degrees of freedom) is mandatory, or you’ll get a pretty probability out of noise.

---

Family 3. Arbitrage / paired — a thin edge, but the most popular among bots

The third family earns on a price inconsistency, and contrary to intuition it’s the most common bot strategy — not because the edge is big, but because it’s nearly riskless and scales well.

The logic is ironclad. The “up” contract and the “down” contract together cover all outcomes: one is sure to settle at a dollar. So buying both together guarantees you a dollar. They should cost 100 cents together. But in moments of imbalance the sum sags — say, “up” at 48 and “down” at 49, together 97. Then you buy both for 97, get 100 at resolution, and three cents are yours with no risk at all: the direction is irrelevant to you, one of the sides wins anyway. Often it’s done more actively: they wait for one side to be dumped sharply below 0.30–0.40, accumulate it cheap, then hedge with the opposite when the sum gives an edge again.

The edge here is small — two to four cents on the dollar. The whole math is in volume: a small edge over many trades and many markets. Let’s count, and notice where the risk hides:

Say the edge is 3 cents and the win rate is 97% (sometimes only one side fills, and you’re left with an unhedged position). Then a trade averages about +2.3 cents accounting for rare misses — but those misses are expensive: a position caught on one side without a hedge carries the very directional risk the whole strategy is meant to avoid. Over 500 trades this adds up to a solid plus, but it rests on two things: on volume and on executing both legs faster than the competitors.

The takeaway for the family: arbitrage is nearly riskless and therefore popular, but its difficulty isn’t in the search (everyone sees the windows) — it’s in execution and volume. You have to catch both contracts atomically, auto-hedge partial fills, and do it on a large flow, because each trade drips little. Lose on speed or on managing partial fills, and the thin edge turns into a directional loss.

---

Family 4. Market-making — earning without prediction

The fourth family: “no directional edge — come here.” The key shift: a market maker doesn’t guess direction at all.

Plainly. Until now you were a buyer — you took at the price offered. A market maker is a shopkeeper: he posts two prices himself, say, willing to buy at 49 and sell at 51. The two-cent difference — the spread — goes to him when one person sells to him at 49 and another buys from him at 51. He matched them and earned, guessing nothing. On top of the spread, the venue often pays the maker a rebate — for keeping prices on the market for others to trade against.

Sounds like easy money — and here’s the difficulty. It’s called adverse selection. Not everyone trading against the shopkeeper is random: some know more than he does — they saw the Binance spot move a fraction of a second before it reached the contract. Such a trader buys from the shopkeeper at 51 exactly when the price is about to go up — and the shopkeeper sold cheap what’s worth more a moment later.

Let’s count the balance. The spread gives half a cent on each side, the rebate another third of a cent; the gross income is 0.8 cents per fill. On an informed trade the shopkeeper loses, say, two cents. The result depends on what fraction of the flow is informed:

While the informed are few, the shopkeeper is comfortably in the black. At 40% comes breakeven: the losses to the “smart” ones eat the whole spread and rebate. Above that — negative, without even trying to guess direction.

And here’s the link that puts everything in place: the informed flow that fleeces the market maker is exactly the fast directional traders from the previous article, the ones who saw Binance first. The shopkeeper’s loss to adverse selection is precisely the fast directional trader’s gain. Two sides of one coin.

The takeaway for the family: market-making doesn’t require prediction — the edge is structural, in the spread and rebate. But it isn’t “easy”: the difficulty is keeping selection below breakeven — pulling quotes in time when informed flow surges, managing your inventory of positions. There’s also a directional variant of MM: at the start of the window take both sides, then actively accumulate the one the tape is really moving toward — then income comes both from the spread and from the excess of shares on the winning side. But that mixes directional risk back into the structural.

---

Family 5. Delta-neutral / hedged — extracting the skew without direction

The fifth family: you hold a position on Polymarket and simultaneously hedge it on a perpetual future on Binance or Bybit. The point is to extract a price skew of the shares while removing the directional risk: if Polymarket priced the probability of a rise wrong, you take that difference, while you’re protected from the BTC move itself by the perp hedge.

The money here is from share mispricing, as in the quant family, but the risk is arranged differently: there’s almost no directional risk, but a new one appears — basis risk. A perp hedge is imperfect: there’s a difference in behavior between the Polymarket contract and the perp, there’s the funding cost of the perpetual, and there’s the risk of executing two legs instead of one. That is, you trade directional risk for basis risk and doubled operational complexity.

The takeaway for the family: delta-neutral gives a systematic edge without betting on direction — attractive for those who don’t want to guess. But the difficulty is real: two simultaneous executions, managing basis and funding, and a mispricing model that still has to be honest. It’s not “removed the risk and print money,” but “shifted the risk from direction to basis and operations.”

---

Not families, but entry tactics

So the map doesn’t blur: latency, mean reversion, and last-second aren’t separate sources of income but tactics for entry within the families above. Latency (exploiting the contract’s lag relative to the spot) is a mechanism feeding directional, quant, and MM strategies. Mean reversion (buying the sagged side after a sharp move, expecting a bounce) is an entry tactic within directional or arbitrage logic. Last-second (trading in the final 10–30 seconds of the window, when the direction is nearly decided, taking the rebate as a maker) is an MM tactic. Worth knowing, but these are tactics on top of a source of income, not the source itself.

---

What would refute this — and how to choose your family

The condition on which the map would collapse (and didn’t). If market-making had no structural income — if the spread and rebate were always eaten clean by selection — the family would be a fiction. But the balance is positive while informed flow is below breakeven. The same with arbitrage: if the sum of “up + down” never sagged beyond the costs, the family wouldn’t exist — but the windows open, thin as they are.

How to choose. The five families are five different hard problems, and you have to pick not “where it’s easier” (it’s easy nowhere) but which difficulty is within your reach:

• Strong at analysis and hypotheses with a reason — directional or quant, but be ready to check the edge hard and calibrate the model to Polymarket’s resolution.
• Able to build careful logic for pulling quotes and managing inventory — market-making gives a structural income without prediction.
• Have fast infrastructure and love the race for execution — arbitrage/paired: nearly riskless, but crowded, and speed decides it.
• Ready to run two legs and manage basis — delta-neutral removes directional risk at the price of operational complexity.

No single skill covers all five. And remember the main practical rule from the practice of these markets: usually the profitable ones are either pure/near-arbitrage or strict timing and selectivity — don’t trade every window, wait for yours. Whoever trades everything feeds those who wait.

---

What I don’t know, and where the limits are

• All the numbers are illustrative, from synthetics or formulas. Spread, rebate, share of informed flow, arbitrage edge, model parameters were set by me as plausible; on your market they’re your own, and they have to be measured.
• The rebate is a volatile parameter. The maker reward and the venue’s fee schedule change; verify the specific figure in the current documentation as of the date. The families’ logic doesn’t depend on it, but MM’s viability does: without the rebate, breakeven on selection comes sooner.
• Deribit is an external source. The IV model is more powerful than GBM, but Deribit data lies outside the spot and the contract; it’s something you connect yourself. And the probability from options is risk-neutral — its edge has to be checked like any other.
• The model has to be calibrated to Polymarket’s resolution (the oracle), not to the spot. Otherwise your probability is about one market while you’re paid by another.
• Short windows (5–15 min) are heavily botized; there latency and microstructure often matter more than the model. There’s more room for a model/directional edge on longer windows.
• Edges burn out. What worked six months ago may have been arbitraged away by bots; how to tell that an edge has died will be covered separately.

---

What to do tomorrow

1. Decide which source you take money from before you build a bot. Prediction, a model, arbitrage, spread, or a hedge — those are five different projects with different success metrics. Mixing them is a common beginner’s mistake.
2. Remember the entry price in every family. Buying the obvious winner expensive is a structural loss even at high accuracy. You need an edge, not “obviousness.”
3. For quant, calibrate to Polymarket’s resolution and run it through the whole validation of the cycle. A model overfits easily; the out-of-sample fan and counting degrees of freedom are mandatory.
4. For arbitrage and MM, measure execution, not the edge. In arbitrage — the share of partial fills without a hedge; in MM — the share of informed flow and the net after selection, not the gross spread.
5. Run the chosen family on real execution without risking money. Paper mode shows the spread, selection, misses, and thin arbitrage windows on the data of both venues — free, before any subscription.

---

FAQ

Which family is the most profitable? The wrong question. Profitable is the one whose difficulty is within your reach: prediction, model calibration, execution speed, managing selection, or basis. A skill you don’t have will make any family unprofitable.

If arbitrage is nearly riskless, why isn’t everyone rich on it? Because the edge is thin (two to four cents), and all the earning is in volume and the speed of executing both legs. Lose on latency or on partial fills, and a nearly riskless strategy turns into a directional loss. It’s an infrastructure race.

Is quant with an options model the grail? No. First, Deribit data is external, you have to connect it yourself. Second, the probability from options is risk-neutral, like the contract price — an “edge against it” may be a difference in risk premiums, not an error. Third, a model overfits easily. It’s powerful, but not free and not automatic.

Who is the “informed flow” that fleeces the market maker? Traders who know more than you at the moment of the trade — usually the ones who saw the Binance spot move before it reached the contract price. That very speed edge from the previous article. Their earning is your loss as a market maker.

Will a model help on short 5-minute markets? Probably not: latency and microstructure rule there, and a “pretty” probability lags the speed. A model and directional edge make more sense on windows of an hour and up.

---

Disclaimer

This material is educational and is not financial advice. Past results do not predict future results. Trading on prediction markets carries real risk, up to and including the total loss of capital. Access to Polymarket is restricted or prohibited in some jurisdictions — verify legality where you live before trading.