How an Ex-Citibank Quant Turned $1,708 into $83,864 on Polymarket in 60 Days: The Dual-Engine Architecture

A former Citibank quant executed 30,280 trades and delivered ~49x return by running two parallel, highly specialized engines on Polymarket. I reverse-engineered the full strategy using Claude and backtested it against a 78-million-trade dataset.

The secret isn’t a single magic model — it’s architectural separation of alpha sources.

The Two-Engine System

Engine 1 — Deep Discount Hunter (Low-Conviction, High-Edge Outliers)

• Scans contracts priced <10¢ on BTC, ETH, SOL, XRP windows
• Market-implied probability <10%, but internal model outputs ~45–55%
• Position size: $8–$100 (tiny to survive variance)
• Goal: capture extreme mispricings where the crowd is irrationally pessimistic

Technical Implementation:

• Real-time scan across thousands of active markets via CLOB V2 + GraphQL
• Calibrated probabilistic classifier (XGBoost + Platt scaling + Bayesian update)
• Features: historical resolution baseline, on-chain activity, sentiment delta, time-to-resolution decay
• Strict entry filter: model_prob - market_prob > 0.40 after expected slippage

Engine 2 — High-Conviction Fade (Crowd Overpricing Grinder)

• Targets markets where crowd is slightly wrong-sided at 50–57¢
• Larger position size: $1,000–$1,200
• Goal: consistent edge on moderate mispricings with higher liquidity

Technical Implementation:

• Different feature set optimized for mean-reversion signals
• Ensemble model focused on order-book pressure, aggressor flow, and short-term momentum exhaustion
• Entry filter: model_prob - market_prob > 0.08 (tighter because size is larger)

Shared Production Infrastructure

Both engines share:

• Unified Polymarket CLOB V2 SDK for sub-second order placement and full order-book reconstruction
• Live oracle-gap monitoring (Coinbase vs Chainlink proxy)
• Dynamic fractional Kelly sizing per engine (aggressive on Engine 1, conservative on Engine 2)
• Global risk engine: daily drawdown circuit breaker, per-market exposure caps, one-click kill-switch
• Tick-by-tick logging → perfect replay and continuous model recalibration

Why This Architecture Crushes

1. Risk Isolation — Small size on Engine 1 survives massive variance; larger size on Engine 2 benefits from higher Sharpe
2. Complementary Alpha — One hunts fat tails, the other grinds steady returns
3. No Single-Point Failure — Different models, different feature sets, different regimes
4. Scale Without Blowup — 30k+ trades with tiny average position size shows extreme discipline

Backtest Insight: The combined system maintained positive expectancy even after realistic fees and slippage. Most single-model bots fail exactly because they try to do both jobs with one engine.

This is quant-grade thinking applied to prediction markets: separate the outlier-hunting from the mean-reversion grinding, give each engine its own calibrated model and risk parameters, and let them run in parallel.

The result? One of the cleanest real-world examples of institutional-grade edge on a retail prediction market.

---

If you have more questions, please feel free to contact me at any time: https://t.me/FatherSon97

---

Tags: #Polymarket #TradingBots #DualEngine #PredictionMarkets #DeFi #Web3 #QuantitativeTrading #AlgorithmicTrading #CLOB #Fintech