VERTEX
Automated Prediction Market Trading System
AIPythonTradingPrediction Markets
I got curious about prediction markets the way most people get curious — I saw someone make a wild bet on Polymarket and thought 'I could do better than that.' Turns out, building a system that actually does it is a lot harder than clicking a button. VERTEX is a suite of trading bots I've been building, breaking, and rebuilding since early 2025. Each bot targets a different edge in the prediction market ecosystem.
The Bots
- VERTEX LITE — LLM-powered event scanner using Claude to detect mispriced markets every 30 minutes
- Temporal Arb — high-frequency crypto directional trading exploiting Polymarket repricing lag vs. Binance
- Whale Tracker — copy trading top-performing wallets with dynamic sizing and stop losses
- Calibrated probability model built from 187K historical signal-outcome pairs
- Kelly-inspired position sizing with per-window exposure caps
- Full risk management: position limits, consecutive loss pauses, bankroll halt, total loss kill switch
Technical Stack
PythonasyncioBinance WebSocketPolymarket CLOB APIGamma APIClaude APITelegram Bot API
The Process
A suite of trading bots targeting different edges in the prediction market ecosystem — from LLM-powered event analysis to high-frequency crypto directional trades.
VERTEX LITE — LLM-Powered Event Scanner
- The first bot. It scans Polymarket every 30 minutes for mispriced markets using an LLM (Claude) as the analysis engine. The model reads the market question, checks current news and data sources, estimates a fair probability, and flags trades where the market price diverges enough from the model's estimate.
- Trades non-crypto event markets (politics, sports, culture). Quality gates include minimum entry price, liquidity floors, and slippage estimation. Kelly-inspired position sizing capped at $1,000 per trade.
Temporal Arb — High-Frequency Crypto Directional
- Polymarket runs 5-minute, 15-minute, and 4-hour 'up or down' markets on BTC, ETH, SOL, and XRP. The strike price is set at market open — if the coin is above/below the strike at expiry, you win. The catch: market makers on Polymarket reprice these contracts slower than Binance moves. Temporal arb exploits that lag.
- Streams real-time prices from Binance via WebSocket. Calibrated probability model built from 187K historical signal-outcome pairs. Enters positions in the final 2 minutes before expiry when edge is highest. Per-window exposure caps to prevent correlated blowups.
Whale Tracker — Copy Trading
- Monitors the top-performing wallets on Polymarket and mirrors their positions with dynamic sizing, stop losses, and drawdown limits.
What I Learned (The Hard Way)
- Your model is wrong until proven otherwise. The first version of the temporal arb probability model used a theoretical normal CDF. Backtesting against 187,000 actual outcomes showed it was underestimating win rates by 15-30 percentage points.
- Resolution infrastructure is as important as signal generation. We had trades that were correctly predicted wins but got voided because the resolution watcher checked too early — catching mid-resolution API prices that looked like cancellations.
- Correlated risk kills you. Five trades in the same 5-minute window, all betting 'Up' across different coins, all expiring at the same time. When they all lost, it wiped out a week of gains in 60 seconds. Now there's a per-window exposure cap.
- Paper trading lies. Our paper trading showed amazing results because of a bug in orderbook sorting — it was reading the wrong side of the book and seeing artificially good prices. The fix was one line of code, but weeks of 'results' were meaningless.
Technical Architecture
- Signal Engine: Real-time price streaming → probability model → edge calculation → entry decision.
- Risk Manager: Position limits, per-asset caps, per-window exposure caps, consecutive loss pauses, bankroll halt thresholds, total loss kill switch.
- Execution: Order book walking for slippage estimation, fill verification, simulated and live execution modes.
- Resolution Watcher: Polls Gamma API for market outcomes, handles edge cases (mid-resolution prices, API lag, force-void after TTL).
- Alerting: Telegram bot for trade entries, exits, P&L updates, and risk warnings.
Challenges & Iterations
- Migrating from CLOB v1 to v2 when Polymarket deprecated the old API mid-trade
- Building a calibration dataset from 187K historical outcomes when Gamma's text search API was broken
- Debugging silent orderbook sorting bugs that made paper trading results look 10x better than reality
- Handling Gamma API lag where resolved markets take minutes to update outcome prices
- Managing bot lifecycle on a laptop — sleep/wake killing WebSocket connections