TickForge

Binance L2 order book, bi-temporal Parquet lakehouse, and cross-feed reconciler

Connecting
GitHub Repo Arnab Mondal
Mid price
--
Bid -- / Ask --
Weighted microprice
--
Edge vs mid: -- bps
Bid-ask spread
-- bps
Invariant: pending
Order book imbalance (L5)
--
$500K buy slippage
-- bps
$100K buy: -- bps
Live feed rate
-- msg/s
Update ID: --

Live L2 Order Book Ladder

BTC/USDT (100ms)
Price (USDT) Size Cum. notional
Mid pending Spread pending

Cumulative Depth Curve and Liquidity Wall

Top 20 levels

Live Cross-Feed Reconciliation Auditor

WS vs REST /api/v3/depth
UTC time Pair WS mid REST mid Drift (bps) Crossed Status
Waiting for first live WS vs REST reconciliation pass…

50,000-Event Deterministic Replay Benchmark

Python engine (measured)
Throughput --
p50 latency --
p95 latency --
p99 latency --
Per-event L2 update latency distribution (microseconds)

Bi-Temporal Point-in-Time Replay Explorer

Zero future leaks

Every captured Binance snapshot stores dual timestamps (event_time_ms and knowledge_time_ms) in Zstd-compressed Apache Parquet. Scrub the knowledge horizon below to verify zero look-ahead bias during backtest replay.

--
Visible ticks: -- Excluded future ticks: -- Leakage count: 0

System Architecture and Quant Data Invariants

5/5 unit tests passing (buffer-first Binance sync)

1. Buffer-First WebSocket Sync

Opens combined Binance @depth@100ms streams first, buffers deltas, hydrates REST snapshots, and verifies U <= lastUpdateId + 1 <= u with zero missed updates.

2. In-Memory L2 Order Book

Applies incremental level updates and zero-qty deletes, guards against crossed books (best_bid >= best_ask), and computes microprice, OBI (L1/L5/L10), and VWAP slippage.

3. Bi-Temporal Zstd Parquet

Persists dual-timestamped records via PyArrow (event_time_ms vs knowledge_time_ms) so late historical backfills never leak into earlier point-in-time backtests.

4. Cross-Feed Drift Auditor

Compares live WS book state against independent REST depth snapshots, tracking mid, bid, and ask drift in bps, sequence gaps, and rolling 5-sigma return outliers.