drift

Inspiration

Polymarket surfaces crowd-sourced probabilities, but crowds aren't always efficient. When a geopolitical event breaks or sentiment shifts, there's a window where market odds lag reality. We asked: what if you could systematically detect those windows by combining two independent signals — AI-estimated probabilities and on-chain smart-money positioning?

What it does

Drift is a signal engine for Polymarket that fuses two alpha sources:

AI Probability Modeling — For each active market, an LLM ingests recent news and event context to produce an independent probability estimate. The delta between AI estimate and current market odds produces a mispricing score.
On-Chain Flow Intelligence — Drift tracks large-position wallets on Polygon's CTF (Conditional Token Framework) contracts, scoring wallet conviction and historical accuracy. When high-accuracy wallets cluster on one side, that's a flow signal.
Signal Fusion — When AI mispricing and smart-money flow converge directionally, Drift generates a high-conviction trading signal with Kelly-criterion position sizing.

The output is a ranked opportunity table with edge estimates, conviction scores, and suggested allocations — backtested against resolved Polymarket contracts.

How we built it

Polymarket CLOB API and Gamma Markets API for real-time and historical market data
Polygon RPC for on-chain wallet tracking and CTF contract event parsing
Claude API (Sonnet) as the probability estimation engine with web search for live context
Python/FastAPI backend orchestrating all three layers
SQLite for historical data storage and backtest state
Lightweight React dashboard for signal visualization

Challenges we ran into

Polymarket's historical data for resolved markets required stitching multiple API endpoints together with on-chain event logs
Calibrating the AI probability engine to avoid overconfidence required prompt engineering around base rates and explicit uncertainty
Defining "smart money" heuristically — we settled on wallets with >65% hit rate across 20+ resolved positions, weighted by recency

What we learned

The convergence signal (AI + flow agreement) showed meaningfully higher accuracy than either signal alone. The biggest edge wasn't in either model individually — it was in the disagreement filtering: knowing when to stay out is more valuable than knowing when to get in.