Polymarket · educated copy-trading
A bot that follows the wallets who know — copying informed consensus into prediction markets, sized by conviction, held to resolution.
Four rounds of tuning on one window invites overfitting. So the config was frozen — no thresholds, filters, or sizing touched — and replayed on a separate quarter it was never shown. The profit engine reproduced almost exactly.
Sorted by entry price, the profit concentrates in one band: 0.70–0.93. Below it, consensus is just retail herding on longshots; above it, near-certainties pay pennies. The middle is where informed agreement actually moves the outcome.
Across both windows, the 0.70–0.93 mid-band carried 70–85% of Book A's profit at 90–98% win rate — the single most consistent structure in the results.
Several independent, qualified wallets converge on the same side within a window. The strongest, most consistent engine.
A single proven wallet places a bet far above its own median size — a fresh, high-conviction entry worth shadowing.
Wallets with a "knows-the-outcome" signature, entering early and holding to resolution. Gated hardest; opens with a news-aware governor.
Every copy is deliberately late — a realistic 180-second follower delay — sized by fractional Kelly under a stack of per-trade, per-market, and per-book caps, with kill-switches and circuit-breakers underneath.
Running on live Polymarket data, constrained to the same liquid event universe — geopolitics, elections, macro. Early days, but the wins are landing exactly where the thesis says they should.
Event markets resolve over days-to-weeks, so a real live sample takes patience — but for the first time the bot is winning on geopolitics and politics, not sports noise.