Skew

Agora Agents Hackathon — RFB 2 submission.

An autonomous prediction-market mispricing agent that detects and paper-trades mispricings on Polymarket and Kalshi for BTC/ETH price markets, with onchain agent identity and USDC settlement anchored on Arc testnet.

---

What it does

The agent applies a log-normal options pricing model to BTC and ETH price prediction markets. Most market participants price these naively; the agent finds edges where the market probability diverges from what realized volatility and time-to-expiry imply.

1. Ingest — snapshots Polymarket + Kalshi every 5 minutes into SQLite
2. Price — runs a log-normal model (Black-Scholes style) calibrated via walk-forward backtest against 365 days of daily closes
3. Detect — identifies mispricings where |model_p − market_p| > 5pp
4. Trade — books paper trades on top edges, tracks P&L with mark-to-market unrealized values
5. Settle on Arc — each trade decision is anchored on-chain; when a trade resolves, USDC is transferred on Arc testnet as settlement proof

---

Architecture

packages/
  shared/     Types, env config, SQLite helpers
  ingester/   Polymarket + Kalshi market snapshotter → SQLite
  edge/       Log-normal pricing model, mispricing detector, calibration backtest
  agent/      Paper-trade ledger: book → resolve → P&L
  onchain/    Arc integration: ERC-8004 identity, trade anchoring, USDC settlement
  api/        Hono HTTP server (12 endpoints) — the agent's brain
  tui/        Textual Python terminal dashboard

data/         SQLite database (gitignored)
scripts/      One-shot utilities (rpc-ping, faucet check)

---

Arc integration

This agent uses Arc testnet as its settlement and accountability layer:

What	How
Agent identity	ERC-8004 register() on startup — agent gets a unique on-chain ID
Trade anchoring	Every booked trade fires a self-send tx with trade metadata in calldata
USDC settlement	On trade resolution: WIN → USDC transfer to agent address; LOSS → USDC burned to 0xdEaD

The agent's Arc address and registration tx are exposed via /agent and visible on testnet.arcscan.app.

Chain: Arc Testnet (Chain ID 5042002) USDC: 0x3600000000000000000000000000000000000000 (6 dec ERC-20) ERC-8004 IdentityRegistry: 0x8004A818BFB912233c491871b3d84c89A494BD9e

---

Pricing model

For a binary market "BTC ≥ $K at expiry T":

P(yes) = Φ( (ln(S/K) + (σ²/2)·τ) / (σ·√τ) )

Where:

• S = current BTC spot (CoinGecko)
• K = strike price
• σ = 30-day annualized realized volatility from daily closes
• τ = time to expiry in years
• Φ = standard normal CDF

Settlement type modifiers applied for Kalshi barrier markets (max, min, barrier).

Calibration: walk-forward backtest over 365d of daily closes produces per-bucket reliability scores (Brier score per probability decile).

---

API

The Hono server runs on port 3200 by default.

Endpoint	Description
GET /health	Server status + agent registration state
GET /agent	ERC-8004 agent identity + Arc explorer link
GET /dashboard	Spot, vol, top mispricings, open trades, P&L summary
GET /mispricings	Latest detected mispricings (sorted by |edge|)
GET /trades	Paper trades by status
GET /summary	Realized + unrealized P&L summary
GET /calibration	Reliability buckets from last backtest
GET /markets	Active markets in DB
GET /spot	Latest BTC/ETH spot prices
GET /vol	Realized volatility
GET /log	Agent log ring buffer
POST /tick	Run one full cycle manually
POST /backtest	Run 365d calibration backtest

Auto-tick runs every 5 minutes on server startup.

---

Setup

Prerequisites

• Bun ≥ 1.1
• Python ≥ 3.11 + uv
• arc-canteen CLI (for RPC key)

1. Install

arc-canteen login                  # one-time — issues your RPC key
git clone <this-repo>
cd rfb2-agent
cp .env.example .env               # fill in keys (see below)
bun install

2. Configure .env

# Arc RPC — get from arc-canteen
RPC=                               # arc-canteen rpc-url

# Agent wallet (testnet only — never use mainnet keys here)
AGENT_PRIVATE_KEY=                 # 0x-prefixed hex private key with testnet USDC
                                   # Faucet: https://faucet.circle.com (select Arc Testnet)

# Venues (public APIs, no key required for read)
POLYMARKET_API_BASE=https://gamma-api.polymarket.com
KALSHI_API_BASE=https://api.elections.kalshi.com/trade-api/v2

3. Seed the database

bun run ingest          # snapshot Polymarket + Kalshi markets → SQLite

4. Run

Terminal 1 — API server:

bun run api

Terminal 2 — Terminal UI:

bun run tui

The TUI auto-connects to http://localhost:3200. The API server auto-ticks every 5 minutes.

First run

On first start the API server will:

1. Register the agent on Arc (ERC-8004) — takes ~5 seconds
2. Run an initial tick to build mispricings
3. Begin the 5-minute auto-tick loop

---

TUI controls

Key	Action
1	Dashboard tab
2	Calibration tab
3	Settings tab
t	Run tick manually
b	Run full backtest (365d)
r	Refresh all
q	Quit

Click any row in the mispricings or trades table for full detail.

---

Dev scripts

bun run typecheck       # TypeScript type check (all packages)
bun run ingest          # Run ingester once
bun run edge            # Run edge detector once
bun run agent           # Run agent loop once
bun run rpc:ping        # Sanity-check Arc RPC

---

Chain facts (Arc Testnet)

Field	Value
Chain ID	5042002
RPC	Via arc-canteen rpc-url
Explorer	https://testnet.arcscan.app
USDC (ERC-20)	0x3600000000000000000000000000000000000000 (6 dec)
Faucet	https://faucet.circle.com
CCTP domain	26