Icarus

Cryptocurrency market-making research platform. Ingests L2 order book and trade data from four venues over websockets, estimates a cross-venue fair value with a state-space filter, and uses the output for alpha research and strategy backtesting.

System

websockets ──▶ capture ──▶ SQLite ──▶ Kalman filter ──▶ drift predictor ──▶ maker simulator
  (4 venues)                          (fair value +                          (backtest + paper trade)
                                       per-venue basis)

Venues: Coinbase, Kraken, OKX, Hyperliquid (spot; Hyperliquid perp optional).

Key components

Component	Location
Fair-value filter (multi-venue Kalman)	src/icarus/strategy/fair_value/filters/venue_basis_kalman_filter.py
Drift predictor (Ridge, walk-forward)	scripts/walk_forward_lagged.py
Maker simulator (queue + latency)	scripts/simulate_basis_maker.py
Live capture pipeline	scripts/capture_filter_eval.py
Filter accuracy eval	scripts/filter_innovation_mae.py
Stat-arb backtest	scripts/basis_mean_reversion_backtest.py

Findings

1. Filter tracks each venue at sub-spread accuracy One-step-ahead MAE on fair-value prediction across 1M+ observations (12h BTC capture, 4 venues):

venue	MAE	bps
kraken	$4.04	0.53
coinbase	$10.40	1.37
okx	$10.69	1.41
hyperliquid	$11.46	1.51

Typical BTC bid-ask spread on these venues is 1–2 bps, so the filter's prediction error is at or below the spread.

2. Identified alpha, fee-gated at retail Walk-forward Ridge regression on 40 cross-venue and lagged features identifies Coinbase as the price leader (1–2s ahead of Kraken/OKX/Hyperliquid). Maker simulation produces ~$7/fill of simulated edge on Coinbase. At 40 bps retail maker fee this is deeply unprofitable; the strategy breaks even only under a negotiated market-maker rebate (≤1 bps).

3. Cross-venue basis mean-reversion has no edge Tested entry on z-score deviations of per-venue basis vs cross-venue common price; gross P&L (pre-fee) was <0.5 bps per trade on $10K notional across 2000+ trades, indistinguishable from noise. Ruled out quantitatively.

Getting started

poetry install
./scripts/pycheck

# Live fair-value plot (coinbase anchor, spot venues)
make basis ASSET=BTC

# Replay a capture to measure filter accuracy
PYTHONPATH=src python scripts/filter_innovation_mae.py \
    --db-path data/capture/YYYY-MM-DD.sqlite3

# Backtests
PYTHONPATH=src python scripts/simulate_basis_maker.py
PYTHONPATH=src python scripts/basis_mean_reversion_backtest.py

Tooling

• poetry for dependency and environment management
• ruff for formatting and linting
• pytest for unit tests
• scripts/pycheck for the full local quality check
• scripts/socket_sanity.py for live websocket health checks