Polymarket 机器人回测：86% ROI 背后的参数陷阱与现实局限

⚡️ TL;DR (Snippet Optimized)

研究员构建 Polymarket 机器人，在 BTC 15 分钟涨跌市场中，使用 movePct=15%、sumTarget=0.95 参数实现 86% ROI（$1,000 → $1,869）。
但激进参数（movePct=1%、sumTarget=0.6）导致 -50% ROI 仅在 2 天内，凸显参数敏感性。
回测基于自录 6GB 实时数据（因 Polymarket CLOB API 无历史数据），但忽略订单簿冲击、网络延迟和部分成交等现实因素。

🎯 Why it Matters

This isn’t just another “I made a bot” flex. It exposes a critical truth about prediction markets like Polymarket: micro-inefficiencies exist, but exploiting them requires infrastructure, precision, and ruthless parameter tuning. While retail traders manually scalp these gaps, automated strategies could dominate—if they survive real-world friction. However, the moment such strategies scale, they self-destruct by eliminating the very inefficiency they rely on. This case study is a masterclass in the gap between backtested alpha and live-market reality.

🧠 Deep Dive: The Alpha

The bot targets Polymarket’s BTC 15-minute UP/DOWN binary markets—where each round lasts 15 minutes, and tokens pay $1 if correct, $0 otherwise. In efficient markets, priceUP + priceDOWN = $1. But during volatility, this sum can dip below $1 due to asymmetric liquidity, panic selling, or slow order propagation.

The strategy is a two-leg arbitrage loop:

Leg 1: Detect a ≥15% price drop on one side within ~3 seconds during the first 2 minutes of a round. Buy that side.
Leg 2: Only hedge by buying the opposite side if leg1EntryPrice + currentOppositeAsk ≤ sumTarget (e.g., 0.95). This ensures the total cost is below $1, locking in theoretical profit.

But why does this work? Because Polymarket’s order book is thin and reactive. A sudden BTC move triggers cascading sells on one side, temporarily decoupling prices from fair value. The bot exploits this lag.

However, the parameter sensitivity is extreme:

High movePct (15%) + high sumTarget (0.95): Fewer trades, but higher win rate and margin.
Low movePct (1%) + low sumTarget (0.6): Frequent trades, but most are false signals—market hasn’t stabilized, so hedging locks in losses.

Critically, the researcher could not use official historical data—Polymarket’s CLOB API returns empty for this market. So he built a custom recorder logging best bid/ask every second. This enabled deterministic replay, but introduced limitations:

No sub-second data: Misses micro-flashes that trigger real arb.
No order book depth: Assumes full fill at best ask, ignoring slippage.
No market impact: Ignores how the bot’s own orders move prices.
Conservative failure mode: If Leg 2 fails before round end, Leg 1 is marked as 100% loss—even though it might still win.

For infrastructure, running on a Raspberry Pi with JavaScript is fine for testing, but production demands Rust for speed, dedicated Polygon RPC to reduce blockchain latency, and co-located VPS to minimize network delay. Without these, the bot is just a spectator.

💬 Q&A: Key Insights

Q: Can retail traders profit from Polymarket arbitrage bots?

A: Highly unlikely. Profit requires sub-second execution, deep liquidity awareness, and infrastructure most lack. Manual trading misses the window; basic bots get front-run.

Q: How does this impact my portfolio?

A: Don’t chase this alpha. Instead, understand that prediction markets are inefficient short-term but efficient long-term. Use them for hedging or speculation—not as passive income.

Q: Why did the aggressive parameter set lose 50%?

A: Low movePct (1%) triggered on noise, not real moves. With sumTarget=0.6, the bot hedged too early—before prices stabilized—locking in losses on both legs.

Q: Is Polymarket’s lack of historical data intentional?

A: Possibly. Limiting historical tick data raises the barrier to entry, protecting early arbitrageurs and encouraging real-time participation over backtesting.

📊 Data Points & Citations

Source: @the_smart_ape via Bitpush News
Key Stat: 86% ROI ($1,000 → $1,869) with conservative parameters over ~4 days
Data Volume: 6 GB of custom-recorded best bid/ask snapshots
Fee Model: 0.5% trading fee + 2% spread applied conservatively

🚦 Market Verdict

Outlook: Bearish for retail replication
Risk Level: High

Disclaimer: Not financial advice. DYOR.

Polymarket 机器人回测：86% ROI 背后的参数陷阱与现实局限

传输日志