Short-horizon mean reversion in cryptocurrency markets: a matched cross-market measurement

Battle Count: 76

Highly relevant for quantitative trading: documents a pervasive but small (2-3 AUC points) 15-minute mean-reversion signal in crypto that is not exploitable under benchmark spot costs (gross edge ~1.3bp vs 5bp maker round-trip). Provides rigorous out-of-sample methodology (walk-forward, FDR, permutation nulls, holdout) that sets a standard for signal validation. The wrapper/ETF inheritance finding is practically important for fund managers and market structure researchers. The flow-conditioning evidence informs market-making and liquidity-provision strategies. However, the signal is explicitly too small to capture at retail costs, making it more relevant for understanding market microstructure and limits-to-arbitrage than for direct alpha generation.

The core measurement device is deliberately minimal: a three-parameter constrained logistic regression with a power-law kernel tie, fit by MLE. The walk-forward protocol, moving-block bootstrap, and FDR control are standard but require careful implementation. The main complexity lies in the extensive robustness battery (artifact channels, permutation nulls, refit-inclusive bootstraps, DFA, sign statistics), the wrapper panel construction with session-matched bars, and the one-minute timing analysis. The paper explicitly states the estimator is a low-variance device, not a model of markets, and results do not depend on it. Data engineering (multi-venue refetches, order-book snapshots, taker-flow signing) adds practical complexity.

Exceptional reproducibility: complete replication package on GitHub with pipeline code, frozen symbol lists, SHA-256 hashes, fetch scripts rebuilding every input from public sources, and a README mapping each exhibit to one command. All models use seeded procedures for deterministic reruns. The primary test design was fixed before the wide-universe run. Data sourced entirely from public APIs (Binance, Coinbase, OKX, Bybit, Dukascopy, Alpaca). Multiple independent estimators (constrained logit, free logit, parameter-free sign statistic, DFA) corroborate findings.