Prediction Market Arbitrage

Cross-venue prediction market arbitrage tooling for executable YES order flow between Predict.fun and Polymarket.

Prediction market arbitrage tooling focused on Predict.fun and Polymarket. The repository contains:

• Go binaries for fetching market data, building cross-market pairs, serving a dashboard, and scanning arb opportunities
• TypeScript utilities for local data fetching and a lightweight dev server
• Static dashboard assets under site/

Current scope

The live implementation in this repository currently supports:

Platform	Data source	Status
Predict.fun	GraphQL + WebSocket orderbook snapshot flow	Implemented
Polymarket	Gamma API + CLOB orderbook proxy	Implemented

Other platforms mentioned in older notes are not implemented in the current codebase.

Repo layout

cmd/
  arb_scan/           # CLI scanner for profitable YES cross opportunities
  fetch_open_markets/ # Fetch Predict.fun markets
  fetch_all_markets/  # Combine Predict.fun + Polymarket and build pairs
  server/             # Go HTTP dashboard server
  ws_debug/           # Debug utility for raw websocket streams

internal/
  config/             # CLI/config parsing helpers
  market/             # Shared market and payload types
  matcher/            # Pairing and text-matching logic
  polymarket/         # Polymarket client + normalization
  predict/            # Predict.fun client + normalization
  server/             # Dashboard HTTP/SSE server
  worker/             # Bounded concurrency helpers

site/
  data/               # Generated JSON payloads
  index.html          # Static dashboard

*.ts                  # TypeScript fetch/server utilities

Arbitrage model

The scanner looks for executable YES cross opportunities:

• buy YES on venue A at the best ask
• sell YES on venue B at the best bid
• subtract taker fees
• reject mid-price-only inputs that are not actually executable

This avoids double-counting complement-style math and keeps the output aligned with actionable orderbook prices.

Quick start

1. Build the Go binaries

make build

This produces:

• bin/fetch_open_markets
• bin/fetch_all_markets
• bin/server
• bin/arb

2. Generate market data

./bin/fetch_all_markets

Quick smoke run with bounded live data:

POLYMARKET_MAX_MARKETS=1 ./bin/fetch_all_markets \
  --max-markets 1 \
  --skip-orderbook \
  --skip-holders \
  --skip-comments \
  --skip-timeseries

Generated files land under site/data/:

• predict_markets_full.json
• predict_markets_view.json
• markets_full.json
• markets_view.json
• markets_pairs.json

3. Run the scanner

./bin/arb

Or pass a custom pairs file:

./bin/arb ./site/data/markets_pairs.json

4. Run the Go dashboard server

./bin/server

Default server URL:

http://localhost:8050

TypeScript utilities

The TypeScript files are useful for local development and debugging.

Install dependencies:

npm ci

Run the dev server:

npm run dev

Compile TypeScript output:

npm run build

Run type-only verification:

npm run typecheck

Run the TypeScript normalization tests:

npm run test:ts

npm run build emits compiled files into dist/, which is intentionally ignored by Git.

Verification

Go tests:

go test ./...

Go build:

go build ./...

TypeScript check:

npm ci
npm run check

Unified local check:

make check

Bounded live smoke:

make smoke-live

Useful environment variables

Scanner:

• ARB_MIN_NET_BPS
• ARB_MIN_FILL_RATIO

Dashboard / refresh loop:

• PORT
• REFRESH_INTERVAL_MS
• SSE_PING_MS
• AUTO_REFRESH
• FETCH_CONCURRENCY
• CATEGORY_CONCURRENCY

Polymarket orderbook proxy:

• POLYMARKET_CLOB_URL
• POLYMARKET_ORDERBOOK_TTL_MS
• POLYMARKET_ORDERBOOK_CONCURRENCY
• POLYMARKET_ORDERBOOK_LEVELS
• POLYMARKET_ORDERBOOK_MAX_TOKENS
• POLYMARKET_MAX_MARKETS

fetch_all_markets --max-markets N now limits Predict.fun input and, unless POLYMARKET_MAX_MARKETS is explicitly set, applies the same cap to the Polymarket side for smoke tests and small-batch runs.

Notes

• The Go server reads site/data/markets_pairs.json by default.
• site/data/*.json is runtime data, not a source of truth.
• If the cached JSON is malformed, the server now rejects it instead of silently serving invalid metadata.