This project asks whether BTC-USDT order-book and order-flow signals can forecast short-horizon future mid-price movement, and whether that predictive structure survives the operational frictions required to trade it. The goal was not to publish a profitability claim. The goal was to build a reproducible microstructure research system that keeps market-data causality, execution mechanics, transaction costs, and robustness checks in the same chain.
The repository implements a full research pipeline from immutable raw market
data through QA, order-book replay, causal research tables, leakage-controlled
features, forward labels, statistical research, predictive modeling,
walk-forward evaluation, signal construction, execution simulation, accounting,
cost and latency analysis, robustness analysis, final reporting, and performance
engineering. The canonical instrument is BTC-USDT, with explicit vendor symbol
mapping to Binance/Tardis BTCUSDT.
The central engineering challenge was preserving what the research process could have known at each timestamp. Exchange timestamps are retained, but local observation time and source order drive replay and feature eligibility. Features are backward-looking from each cutoff, labels are strictly forward-looking, and cross-day feature or label leakage is disabled. The data layer preserves raw source bytes, records checksums, and validates schema, duplicate, timestamp, price, quantity, sequence, book-crossing, update-gap, and stale-BBO issues.
Queue imbalance produced a strong and stable 1s predictive rank signal across the pre-registered development sample, with about 0.43 mean daily Spearman IC and 24/24 positive development dates. OFI and multivariate modeling added incremental information: Extended LightGBM produced +0.0107 incremental IC across 18 expanding walk-forward folds. The modeling result was useful but modest; QI remained the dominant simple signal.
Better predictive IC did not automatically produce better executable economics. Signals are desired states, not fills. Market orders filled reliably but paid spread, latency, and displayed-depth costs. Passive orders exposed queue uncertainty, low fill participation, adverse selection, and residual inventory. Cross-date economic robustness favored the simpler QI efficiency baseline more often than the larger multivariate signal, and generic fee stresses showed thin transaction-cost headroom.
The project emphasizes deterministic artifacts and reviewable controls. Public reports include source-traceable metrics, deterministic hashes, CI status, and explicit limitations. Phase 16 profiled the existing pipeline before optimization and found repeated trailing-window aggregation as a material Phase 5 bottleneck. A Python accumulator refactor reduced representative feature-engineering median runtime from 1.493s to 0.292s, a 5.11x speedup, while preserving an exact reference-vs-optimized feature output SHA-256.
This is a BTC-USDT historical research system using Binance/Tardis reconstruction. It uses displayed book data only, does not observe hidden liquidity, does not model endogenous self-impact, uses a simplified passive queue model, and applies generic fee stress scenarios rather than venue-specific live fee schedules. Execution robustness uses bounded development-date samples. The 2026 temporal holdout remains untouched and reserved for future confirmatory evaluation.