Polymarket’s “Bitcoin Up or Down” markets open a new round every five minutes and ask a single question: will Bitcoin be higher or lower five minutes from now than it was at the open? The markets launched on February 12 this year. They run 288 windows a day, and daily volume has peaked at close to $60 million.
Most people trade them on instinct. They chase headlines and jump in whenever a candle twitches. Machines don’t trade that way. According to some on-chain tallies, trading bots on Polymarket have taken more than $60 million in profit since 2025, about 77% of it from crypto Up/Down markets. Those numbers are hard to verify one by one, but the direction is clear: software is taking over this corner of the market.
What follows looks first at a few bot strategies that already work, then breaks down one an individual can build: the open-source AI agent framework Hermes, wired to a large language model, Polymarket’s CLOB V2 order API and Telegram. The idea behind it is simple. Once Bitcoin settles into a directional state, the odds that the state persists can be measured, and the market price often hasn’t caught up.
Why the 5-Minute Markets
The shorter the window, the more emotion drives the price and the easier it is for pricing to drift. By this strategy’s own estimates, a single asset offers 288 windows a day, a potential opportunity roughly every 81 seconds, and a typical gap of 5 to 15 percentage points between model probability and market price.
Settlement is clean, too. Every window resolves against Chainlink’s BTC/USD feed, the winning side pays $1 per share, and there is no room for human judgment calls. It is the kind of job you can hand to a program and let run around the clock.
The framework here is Hermes, an agent project Nous Research open-sourced last year. It ships with memory, scheduled tasks and a multi-platform messaging gateway, and it has climbed fast on GitHub this year. Nous Research closed a $50 million Series A led by Paradigm in 2025 and has raised about $70 million in total. Trading bots are shifting from scripts with hard-coded rules to agents that can call a model, read their own logs and review their own trades.
Three Bot Strategies That Already Work
The most profitable accounts in the 5-minute Up/Down markets are almost all bots. The three types below all opened their accounts in March this year. They share the same underlying logic but execute it in very different ways. Profile P&L figures are each account’s own numbers, so we also sampled about 4,000 recent public trades from each one to reconstruct how they enter.
High-Confidence Spread Capture: Buy Both Sides, Then Close Out in the Final Minute
The first type trades almost nothing but Bitcoin, mostly the 5-minute markets, with some 15-minute and hourly ones. From just before the open through the first two minutes, it buys both “Up” and “Down” at around 50 cents. In the final minute, when the direction is mostly settled, it puts the bulk of its money on the side that is already a heavy favorite, usually paying 95 to 99 cents. In our sample, more than half of its final-minute fills were above 85 cents, and that minute carried the most dollar volume of the whole window.
It earns the last 1 to 5 cents. The median spend per window is under $100, and the profit comes from doing that across hundreds of windows a day. According to its public profile, it has made about $746,000 in total across more than 37,000 markets, with a biggest single win of $11,900.
The weakness is obvious. Buying at 99 cents earns a cent. One reversal in the final seconds wipes out the profit from dozens of trades.
Dual-Mode Expected Value: Hedge Early, Buy Cheap Tickets Late
The second type trades the 5- and 15-minute markets for Bitcoin, Ether, SOL and XRP at the same time. For the first minute and a half, it also buys both sides near 50 cents. From minute two to minute four, it changes gears and concentrates on the side the market has written off; close to half of its fills in that stretch were below 15 cents.
Put the two modes together and the average cost of both sides in a window adds up to a median of about $0.91, below the $1 payout at resolution. That part is a fairly steady spread. The late cheap tickets are long-odds bets: buy at 10 cents and a hit pays 10 times. It has reportedly made about $797,000 across more than 33,000 markets. Its biggest single win, $42,200, is the largest of the three and most likely came from one of those cheap tickets.
Most cheap tickets expire worthless, and the account depends on a few hits to make up the difference. In a one-way market the drawdowns get deep. Public records also show the account has made no new trades since early May.
Multi-Asset Variance Reduction: The Busiest and the Most Balanced
The third type covers Bitcoin, Ether and SOL across 5-minute, 15-minute and 4-hour markets. It keeps buying both sides throughout the window, its volume barely changes from minute to minute, and its positions are the most balanced of the three. In our sample, the imbalance between Up and Down shares was only about 20%, against 40% to 60% for the other two.
It has the thinnest margin per trade and makes it up on volume. It filled 4,000 trades in the roughly two hours we sampled, and its cumulative volume in crypto markets is close to $300 million. Running several assets and timeframes at once evens out the wins and losses in any single market. According to its profile, it has made about $569,000 across more than 22,000 markets, with a biggest single win of $18,500.
By a rough count, its profit is only about 0.2% of volume. A small fee increase, or a rival that gets its orders matched faster, could be enough to erase the edge.
The Easier One to Copy: A Markov Persistence Strategy
All three of the above lean toward market making and spread capture, which reward speed and fast turnover of capital. Better suited to an individual is the directional strategy covered in the rest of this article: use a Markov chain to judge whether a state will persist, trade only when the probability gap is wide enough, and size positions with Kelly. The open-source community has two other common variants. One estimates probabilities with Brownian motion and calibrated volatility and pairs that with quarter-Kelly sizing, which makes it more conservative. The other combines arbitrage and momentum logic and lets the program tune its own parameters.
| Strategy | Entry | Sizing | Strength | Risk |
|---|---|---|---|---|
| High-confidence spread capture | Buys both sides at the open, then the favorite at 95–99¢ near the close | About $100 per window | High confidence per trade | Late reversals |
| Dual-mode expected value | Hedges both sides early, then buys the cheap side below 15¢ | Mostly hedged, small cheap tickets | Occasional big wins | Deep drawdowns |
| Multi-asset variance reduction | Buys both sides all window, across assets and timeframes | Small trades, huge volume | Lowest volatility | Razor-thin margin |
| Markov persistence | p(j*,j*) ≥ 0.87 and gap ≥ 5% | Kelly, often fractional | Low barrier, self-learning | Depends on win-rate estimate |
Where the Edge Comes From
At the core is a Markov chain. Bitcoin’s short-term price action is divided into a set of “states,” such as strong uptrend, weak uptrend, sideways and downtrend, and the model counts how often the current state moves to each of the others. Prices are not always a random walk. Once a persistent move is under way, the odds that it continues in the same direction can be well above 50%.
Written as a formula, the entry rule is Δ = p̂ − q ≥ ε. Here p̂ is the model’s probability, q is the market price and ε is the minimum gap, set at 5%. If the model gives “Up” a 70% chance and the market is selling it at 62 cents, the 8-point gap clears the 5% bar and the trade becomes a candidate.
The return per trade is r = (1 − q) / q. Buy at 64.7 cents and a win returns about 55%; buy at 44.1 cents and it returns about 127%. The cheaper the price, the bigger the payout, and the bigger the chance of losing.
A stricter gate sits on top: the state-persistence probability p(j*,j*) must be at least 0.87, or there is no trade. Most of the time the system just waits, and it only acts in highly persistent states. By design, the strategy expects a win rate of 63% to 72% after that filter.
Fees have to be part of the math. Polymarket charges takers on short-dated crypto markets according to the formula shares × 0.07 × p × (1 − p), which comes to about 1.6 cents per share at a price near 65 cents. That takes a real bite out of a 5-point threshold, so the agent should be told to evaluate every trade net of each market’s fee rate.
Sizing the Bets
Position size comes from the Kelly formula: f* = p − (1 − p) / b. Here p is the win probability and b is the net payout on a win, the same (1 − q) / q as above.
With p = 0.87 and an entry at 64.7 cents, b is about 0.546 and f* comes out around 0.63, which would mean staking 60% of the bankroll. That number only shows how the formula behaves. In live trading, p should be the actual win rate measured from the trade log, not the persistence threshold, and the resulting stake will be much smaller. Many open-source projects simply use quarter-Kelly and add a cap on any single bet.
So the system doesn’t count on winning every trade. It stacks four filters: model probability, state persistence, a gap of at least 5%, and Kelly sizing to set the stake.
Tools and Costs
The whole setup is assembled from off-the-shelf parts: the Hermes framework and its desktop distribution Atomic, Claude Opus as the brain, an open-source trading bot as the base, CLOB V2 for execution and Telegram for reporting.
By the original plan’s estimates, running costs stay under $10 a month, mostly model API fees, depending on usage. The minimum starting balance is $10, $50 is the recommended starting point, and you need a little POL for gas. Someone who knows the tools can set it up in half an hour.
Getting the Agent Running in Three Steps
Step one: install Atomic. On the home screen, choose Hermes agent. You can run it locally or click Run in Cloud in the top-right corner; the interface is the same either way.
Step two: connect a model. Go to Atomic settings → AI Models → Anthropic, paste in your API key and pick the current Claude Opus model, which handles real-time analysis, reads the trade log and adjusts the parameters. Set temperature to 0.2 and max tokens to 4096. Trading decisions need consistency, and there is no reason for the model to get creative each time. Alternatives include pay-as-you-go OpenRouter or Codex through ChatGPT Pro.
Step three: connect Telegram. Under Atomic → Skills → Messengers → Telegram, click Connect, create a bot with @BotFather and paste the token back in. From then on, trade reports and daily reviews arrive on your phone.
Handing the Strategy to the Agent
You don’t have to write a trading system from scratch. A good base is polymarket-BTC-15-Minute-Trading-Bot on GitHub. It was built for 15-minute markets and comes with a seven-phase architecture, Grafana monitoring, Redis, and stop-loss and take-profit logic, which makes it a good fit for a version with Markov entries and Kelly sizing.
The instructions to Hermes boil down to this: keep the existing architecture; move the execution layer to py_clob_client_v2 in Python; support Polymarket’s Safe proxy wallets; refer to the collateral balance instead of the legacy USDC wording; evaluate trades net of fees using CLOB V2 market metadata; add the p(j*,j*) ≥ 0.87 persistence filter and Kelly sizing; default to DRY_RUN=true; and never let private keys appear in chats or logs.
Next, have the agent create a trading wallet, return a deposit address and approve three contracts: CTF Exchange, Neg Risk CTF Exchange and Neg Risk Adapter. The environment variables look roughly like this:
PRIVATE_KEY=your_wallet_key
SAFE_ADDRESS=your_safe_address
CLOB_HOST=https://clob.polymarket.com
DRY_RUN=true # simulate first
MIN_EDGE=0.05 # minimum 5% gap
MIN_PROB=0.87 # Markov persistence threshold
MIN_BET=1.00 # $1 minimum per trade
MAX_BET=50.00 # per-trade cap
BANKROLL=100.00 # starting capitalThe “$100” in “starting with $100” is simply BANKROLL=100. It is the stake for testing the framework, not a promise of what it will grow into.
Before going live, run the bot in simulation for a full 24 hours. Log the number of signals, entry prices, the state at entry, simulated P&L per trade and the win rate at the threshold, and send a Telegram summary every six hours. If the signals are erratic or the simulated P&L doesn’t hold up, don’t rush to go live. Once the results are steady, keep practicing with small real trades of $1 to $2 each.
The Self-Learning Loop
What sets Hermes apart from an ordinary bot is that its parameters aren’t fixed. A state that works today may stop working tomorrow.
During the day, the agent trades whenever the thresholds are met and writes every entry, exit and P&L into a trade journal. Without the journal, there is nothing to review. At midnight, Claude Opus reads through the day’s log: which states had the highest win rate, which price ranges had the best expected value, which hours lost the most, and whether the current MIN_PROB and Kelly fraction still make sense. Then it rewrites the thresholds, adjusts position sizing and updates the parameters in .env. If a loose threshold led to losses, it tightens it. If a price range performed well, it gets more weight. If sizing was too aggressive, it comes down. The next day runs on the new parameters.
Every morning, a Telegram report arrives with the previous day’s P&L, win rate and trade count, which rules changed and why, and today’s new thresholds. Read it, confirm, and then decide whether to keep running.
After 50 to 100 trades, the agent has a history of its own and starts to tell which states carry a real edge and which hours to avoid. Its value isn’t in predicting the market. It turns every trade into material for the next round.
Before You Start
This strategy lives on thin margins. Fees are highest at mid-range prices, and only the numbers after fees count. Win rates over a few dozen trades swing widely, so keep early bets small and don’t mistake a lucky streak for skill.
The private key is the biggest risk. Read open-source code before you run it, use a dedicated wallet that holds only a small amount, and keep the key out of chats and logs. Polymarket is restricted in many countries and regions, so check the rules where you live before you start.
This article is for informational purposes only and does not constitute investment advice.
