DEV Community

manja316
manja316

Posted on

Building a Polymarket Crash-Fade Bot with 89% Win Rate

Political prediction markets crash when news breaks. Most of these crashes recover. I built a bot that buys the crash and sells the recovery.

48 paper trades. 89.6 percent win rate. Running live on Polymarket.

The Edge

When a political market drops more than 15 percent in under 4 hours, it recovers to 90 percent of the pre-crash price about 70 percent of the time. The bot watches 200 markets every 15 minutes, detects crashes, and buys the dip.

How It Works

  1. Fetch 200 active political markets from the Gamma API
  2. Track 4-hour price highs for each market
  3. When current price drops more than 15 percent from the 4h high: buy
  4. Exit when price recovers to 90 percent of pre-crash high, or after 48 hours

Key Design Decisions

Position sizing: 5 dollars per trade. Small enough that losses do not matter, big enough to prove the edge.

Exit strategy: BUY NO to exit (not SELL YES). This is a quirk of Polymarket CLOB — selling YES tokens requires them as collateral, but buying the complementary NO token achieves the same economic result.

Kill switch: If win rate drops below 55 percent after 20 trades, the bot pauses automatically.

Results

After 48 closed paper trades:

  • Win rate: 89.6 percent (43 wins, 5 losses)
  • Average profit per trade: 0.70 dollars
  • Total paper PnL: +33.72 dollars
  • Most profitable: political qualification markets (FIFA World Cup, primaries)
  • Most losses: timeout exits on slow-moving markets

What I Learned

  1. Paper PnL is not real PnL. My live execution had bugs where exits could not fire due to a safety cap that was too low. Verify live round-trips before trusting paper numbers.

  2. The CLOB minimum order is 5 shares. At 2 dollars per trade with prices above 40 cents, you get fewer than 5 shares and the order gets rejected. Position size must account for this.

  3. Political markets have strong mean-reversion. Sports and crypto markets do not — they trend. Different market types need different strategies.

The bot runs autonomously via a Python process monitored by a Paperclip agent that restarts it if it crashes.


Building autonomous trading systems at github.com/LuciferForge

Top comments (0)