Inside the five-minute Bitcoin bots

A review of more than fifty high-performing Polymarket bots found that most are not built around one perfect directional forecast. Their edge comes from a combination of pricing, position structure, execution, liquidity and risk control. Here is how the system works.

Suppose Bitcoin suddenly moves above the opening price of a live five-minute market. The move is already visible on an external exchange, but some sell orders for Up on Polymarket were placed before it happened. If those orders are still resting in the book, the contract can briefly stay available at a price that no longer matches the latest state of the market. That temporary mismatch is what the bot hunts for.

Instead of treating the Polymarket price as fair value, the bot makes its own estimate. Say Up trades at 42¢ and a new Bitcoin move lifts the model’s probability of Up to about 54%. The question stops being whether Bitcoin looks bullish. It becomes whether a 54¢ outcome can still be bought for well under 54¢. One simple way to update the probability is to scale the previous odds by the strength of the new signal. This is Bayes’ rule in odds form, and the multiplier is the signal’s likelihood ratio.

Suppose the bot values Up at 55¢ and the best sell order is 46¢. That looks like 9¢ of edge, but the bot cannot necessarily capture it. The first 80 shares may be available at 46¢, the next 200 at 48¢ and the next 400 already at 50¢. What matters is not the best ask but the average price of building the whole position, after fees, spread, slippage, partial fills, worse prices at larger size, and uncertainty in the model itself.

Tradable edge = fair value − expected average entry − trading costs − safety margin.

This is where strategies that look spectacular on a chart fall apart live. A mispricing can exist in historical data, at the top of the book, and even for 50 shares, then disappear completely for a bot that wants 5,000. Finding an edge is one thing; being able to execute enough size before it disappears is another.

The same Bitcoin move affects the current five-minute market, the next five-minute window, the fifteen-minute market and the short-term ETH and SOL markets at once, but they do not all update at the same speed. Suppose Bitcoin jumps and the current five-minute market reprices Up from 50¢ to 68¢, while the fifteen-minute market is still around 54¢. That alone does not make the fifteen-minute contract cheap: the two have different opening prices and different amounts of time left. So the bot estimates fair value for each separately.

Relative score = (current gap − typical gap) ÷ historical volatility of the gap. That is a z-score of the spread between connected markets. A large deviation doesn’t create a trade on its own, but it tells the system that one part of a connected structure has moved unusually far from the rest. A person watches one contract; a bot can watch a whole network of related probabilities at once.

Finding an underpriced outcome is only the first step; the next question is how to build the position. Across high-performing bots, five structures keep recurring.

The model sees an edge in Up and the bot starts buying Up. Then Bitcoin loses momentum, drifts back toward the opening level, and the estimated probability of Up falls. The bot doesn’t have to stay committed. It can stop adding Up, trim part of the position, start accumulating Down, and rotate again if the market turns once more. The first entry isn’t the final decision; the position keeps moving with the model. The danger is noise. If Bitcoin chops back and forth, the bot can switch sides repeatedly and bleed its edge away through spread, slippage and poor fills.

The simplest binary arbitrage is when Up + Down costs less than $1. At Up 45¢ and Down 48¢ the pair costs 93¢, and one of the two will settle at $1. Some bots build that pair without buying both sides at once. Bitcoin rises and Down falls to 27¢, so the bot buys Down. Later Bitcoin reverses toward the opening level, and Up can be picked up around 49¢. The finished pair costs 76¢, from two prices that never existed together in the book.

The catch is that until the second side is bought, it isn’t arbitrage. If Bitcoin keeps climbing after the Down purchase, cheap Up never appears and the bot is simply holding a directional bet. More conservative systems build the pair in small blocks (50 Down, then 50 Up, then another block) so less inventory is ever left exposed.

Taker fees change the arithmetic. As a taker, the 93¢ pair pays about 3.5¢ in fees and nets 3.5¢ rather than 7¢, and a 94¢ pair nets 2.5¢. The 76¢ legged pair still nets almost 21¢.

Other bots don’t try to hold equal quantities. With 260 Up and 235 Down, 235 on each side form the paired portion and the remaining 25 Up are the directional bias. Most of the position is protected, while some extra exposure rides on the outcome the model prefers. Directional exposure = Up quantity − Down quantity.

If the signal strengthens, the bot can widen the imbalance; if confidence falls, it adds Down and reduces directional risk. But holding both sides doesn’t make the position efficient. At average prices of Up 55¢ and Down 49¢, a complete pair costs $1.04 and can only ever return $1, so the “protected” part already carries a 4¢ deficit that the extra directional contracts must recover before the position is in profit.

Some bots don’t think in terms of one market, one trade. They run capital across the five- and fifteen-minute Bitcoin markets, the ETH and SOL markets and neighbouring windows, managing an inventory rather than a collection of separate bets. Suppose a Down position has climbed to 98¢. The bot can wait for settlement at $1, or sell part of it at 98¢ and move that capital to a market where a fresh edge has appeared. It may also buy a sliver of the opposite outcome at 1–2¢. That costs almost nothing if nothing unusual happens, and offsets part of the loss if Bitcoin crosses the line at the last moment. The question for this kind of bot is how efficiently the entire inventory is using edge, not whether one entry was good.

Another group trades almost only at the very end, when the result looks close to certain but the likely winner still trades below $1: buy at 98.7¢, settle at $1, keep 1.3¢. The return per trade is tiny, so the strategy lives on turnover and volume. The risk is extremely lopsided. A bot can collect 1¢ again and again, then lose almost the whole contract on one wrong trade. That is why a very high win rate doesn’t make the strategy safe; the size of the rare losses matters far more.

Two details now favour this style. The taker fee at 98.7¢ is only about 0.09¢ a share, because the fee shrinks toward the extremes. And since settlement moved to a window average, a last-second spike can no longer flip a nearly settled market the way it could when the final price decided it.

Suppose the bot finds what looks like perfect arbitrage, with Up + Down at 94¢. It sends two orders. Up fills, but Down fills only partly, then the market moves and the rest of Down becomes too expensive. The arbitrage is gone and the bot is holding an open Up position. The pricing model wasn’t necessarily wrong; the failure happened in execution.

One answer is to make the working quote depend on inventory. Working price = fair value − inventory penalty, and inventory penalty = q × λ × σ² × τ, where q is the current position imbalance, λ is risk sensitivity, σ is volatility and τ is time remaining. It is the reservation price from the Avellaneda–Stoikov market-making model, with the mid price swapped for the bot’s fair value. If the bot already holds too much Up, it becomes less willing to buy more Up, and buying Down becomes more valuable because it reduces the imbalance. It sounds simple, and it is one of the biggest differences between a real execution system and a bot that just keeps buying while its signal stays positive.

Even a strong edge doesn’t justify putting all the capital into one market. The model can be right and the trade can still fail because liquidity disappears, one leg never fills, the API lags, an external data feed updates late, the real average entry is worse than expected, or several correlated positions open at once. So sizing is a separate decision, and one approach is to use only a fraction of the full Kelly allocation.

Correlation matters most. A five-minute and a fifteen-minute Bitcoin position are technically two markets, but if both carry large Up exposure they are effectively the same bet. The same happens when BTC, ETH and SOL all move together in a broad crypto swing.