What the bot does
The Polymarket 5-min bot is a market maker: it posts resting buy and sell orders on very short-dated binary markets and tries to earn a small, structural edge rather than betting on which way the price goes. It runs in two modes.
Primary mode — the risk-free pair. A 5-minute market resolves to exactly one of two outcomes, so the two sides ("up" and "down") must together be worth exactly $1.00 when it settles. When the bot can buy both sides for less than a dollar combined, the payout at resolution is fixed and known in advance. It holds both legs and collects the difference. This is the core, and it is structurally risk-free at resolution — provided both legs actually fill.
Secondary mode — late-window directional maker. In the closing seconds of a window the bot may place a single directional maker order, informed by a calibrated online model. This mode is a genuine directional bet, it is unproven, and it is held to a much higher bar than the pair strategy. Today it runs in shadow: measured, scored, and not used to size real paper risk until it earns the right to be.
The market it plays
Polymarket 5-minute markets are binary prediction markets: "will BTC be up or down at the close of this 5-minute window?" You buy shares of an outcome for somewhere between $0.00 and $1.00. If your outcome happens, each share pays $1.00; if not, it pays $0.00. The price is therefore a live probability — a share at $0.60 is the market saying "about 60% likely."
Two features of this venue make it a maker's playground:
- The two sides must sum to $1 at settlement. "Up" pays $1 exactly when "down" pays $0. That hard constraint is what makes a genuine, arithmetic arbitrage possible — no forecast required.
- Maker orders can be free. Posting liquidity (a resting limit order someone else trades against) avoids taker fees, so a maker keeps more of a thin edge than a taker ever could.
The flip side: windows are only five minutes long, order books are thin, and the clock is always running. Thin, fast markets punish sloppy execution — which is exactly why the honest accounting below matters so much.
The concepts, in plain English
Three ideas do all the work here: pair arbitrage, market-making, and inventory risk. Understand these and you understand the whole bot.
Pair arbitrage
Because the two outcomes must total $1 at resolution, buying both for less than $1 combined locks a fixed payout. You don't care who wins — one leg always pays $1. The gap between what you paid and $1 is the edge, banked risk-free at settlement.
Market-making
Instead of crossing the spread (paying up to buy now), a maker posts resting orders and waits to be filled. You provide liquidity, and in return you buy a touch cheaper and sell a touch richer than the mid — earning the spread instead of paying it.
Inventory risk
A maker's nightmare: you meant to hold a balanced book, but only one side of your pair fills. Now you're holding a naked directional position you never wanted — that's inventory risk. You must hedge or unwind it, usually at a loss.
Calibrated model
For the late-window mode, "calibrated" means when the model says 70% it's right about 70% of the time. An online model keeps updating as new windows resolve. Calibration is what lets you trust a probability enough to (eventually) size against it.
A worked example of the pair
Whichever way BTC moves, one leg settles at $1.00 and the other at $0.00. You paid $0.97 for a guaranteed $1.00 — a $0.03 gross edge, no forecast needed. If both legs fill. That last clause is where the whole story turns.
Where inventory risk bites: the orphaned leg
Suppose your "buy UP" order fills but the market moves before your "buy DOWN" fills. You're no longer arbitraged — you're holding a bare directional bet on UP that you never intended. This is the half-fill, or orphaned leg, and it is the single biggest threat to a pair strategy. To get back to safe, you either buy the missing side at a now-worse price or hedge it — and either way you pay. A pair engine lives or dies on how well it controls, and honestly prices, its orphans.
Why it's disciplined
Discipline here isn't a personality trait — it's accounting rules that make it impossible to lie to ourselves. Four of them:
- Every simulated fill is cost-loaded. Fees, modeled slippage, and the realistic cost of getting filled are subtracted inside the paper P&L. A fill that couldn't have happened at that price doesn't get booked at that price.
- The exit is modeled, not assumed. A position you can't unwind within a few percent isn't really a position. We never mark a thin book at its last-trade price and call it profit.
- Orphan drag is measured and booked. Every half-fill logs what it actually cost to hedge, and that loss is subtracted from the headline number. There is no "gross edge before the messy part" line item on the dashboard.
- Nothing graduates on vibes. The directional leg stays in shadow until it clears a pre-registered statistical bar — written down before the evaluation window, so the goalposts can't move.
Who actually wins (and why most don't)
The natural question: if these markets are so efficient, how do some people consistently make money on them? The honest answer is that the winners aren't predicting anything — they hold a structural edge most participants simply can't access. There are really only three ways anyone wins here, and all three are infrastructure-and-capital games, not clever-forecast games.
Market-making + rebates
Quote both sides of every window and earn the spread plus the venue's maker rebate. You get paid to provide liquidity — structural, not directional. It needs sub-100ms cancel/replace, a datacenter connection, and enough volume to reach rebate tiers. This is how the biggest consistent winners actually do it.
Latency arbitrage
These markets resolve on a crypto price. Whoever sees the underlying tick a few milliseconds before the book updates buys the mispriced side risk-free. A pure speed race — the vast majority of arb profit goes to the fastest bots. A laptop on home internet is hundreds of milliseconds behind: it's the slow money getting picked off.
Patient pair-arb
Wait for the rare window where both sides can be bought for under a dollar combined, and lock the risk-free profit. This one needs less hardware — but the opportunity appears in a tiny fraction of windows, requires capital to matter, and lives or dies on solving the half-fill problem.
So why does nearly everyone else lose? They try to predict which way the price moves in five minutes. In a hyper-efficient market that's a coin flip minus fees — and the people betting direction are the very liquidity the market-makers and speed bots feed on. It isn't that they lack a secret playbook; they're structurally sitting on the losing side of the table.
This is exactly why the honest accounting here matters — and why the next section is the most important one on this page. When we forced every simulated fill to pay real-world costs, the naive edge didn't shrink a little. It vanished. That isn't a flaw in our modeling; it's the mathematical fingerprint of an edge that lives entirely in infrastructure we don't have.
The honesty story
This is the part most trading pages leave out. Here's what actually happened when we stopped counting fills the flattering way and started counting them the real way.
On paper, a sub-$1 combined price looks like free money. Buy both sides for $0.97, collect $1.00, repeat forever. If you only count the pairs that completed cleanly, the strategy prints. That's the naive number — and it's the number a dishonest dashboard would show you.
Then we modeled real execution. In thin, five-minute books, a meaningful share of intended pairs half-fill: one leg lands, the other doesn't, and you're stuck hedging an orphan at a loss. When we booked every one of those orphan costs against the P&L, the picture changed completely.
The lesson generalizes: a naive edge that disappears under realistic execution costs was never an edge — it was an accounting error. Modeling those costs didn't make us less profitable; it revealed that the flattering version was fiction. We would rather ship a thin, true number than a fat, fake one.
What stays private (and why)
We'll happily explain the philosophy. We won't hand a competitor the parameters. The rule of thumb: a sentence that's also true of any competitor's bot is safe to say; a sentence that lets someone clone ours is not. So the following live only in config, never in public:
- The exact combined-price threshold that makes a pair worth taking
- All timing values — entry cutoffs, hedge deadlines, pair timeouts
- The queue / fill assumptions the maker model uses
- The late-window model's input features, weights, and calibration internals
- The directional entry conditions and price bands
These are the edge. Publishing them wouldn't make us more honest — it would just let the thin edge get arbitraged away by anyone who copied the numbers. Everything on this page is the honest category-level story; none of it is enough to rebuild the bot.
See it live
The public dashboard reads the exact same JSON the bot writes — no edited screenshots, no cherry-picked wins. Where the honest number is thin, negative, or zero, it shows it that way, in real time.